A Study Compendium
Compiled from a personal knowledge graph. Each entry is a node; the Requires / Builds toward / See also links are the graph's typed edges — read it front-to-back as a curriculum or jump between concepts.
Part 1 — Foundations · 11
Backpropagation · Bayes' Theorem · Calculus · Discrete Mathematics · Eigendecomposition · Gradient Descent · Gradient, Jacobian, Hessian · Linear Algebra · Ordinary Differential Equations · Probability · Singular Value Decomposition
Part 2 — Programming & Data · 5
Data Structures & Algorithms · NumPy · pandas · Python · SQL
Part 3 — Statistics & Experimentation · 14
A/B Testing · Central Limit Theorem · Central Tendency · Correlation Analysis · Delta Method · Distribution Shape · Econometrics · Hypothesis Testing · Improving Sensitivity (CUPED) · Minimum Detectable Effect · Ratio Metrics · Statistical Power · Statistics · Time Series
Part 4 — Machine Learning · 33
Autoencoders · Confusion Matrix · Cross-Validation · Data Cleaning · Decision Trees · Dimensionality Reduction · Ensemble Methods · Feature Engineering · Feature Scaling · Feature Selection · Gradient Boosting · Hyperparameter Tuning · K-Means Clustering · K-Nearest Neighbors · Linear Regression · Logistic Regression · Model Evaluation · Model Selection · Polynomial Regression · Precision, Recall & F1 · Predictive Analytics · Prescriptive Analytics · Principal Component Analysis · Random Forest · Recommendation Systems · Regularization · Reinforcement Learning · ROC-AUC · scikit-learn · Self-Supervised Learning · Supervised Learning · Support Vector Machine · Unsupervised Learning
Part 5 — Deep Learning · 26
Activation Functions · Attention · Convolution · Convolutional Neural Network · Deep Q-Network · Fine-tuning · Forward Propagation · Generative Adversarial Network · Image Classification · Image Segmentation · ImageNet · Keras · Loss Functions · LSTM · Multi-Head Attention · Multilayer Perceptron · Neural Network · Perceptron · Pooling · PyTorch · Recurrent Neural Network · Self-Attention · TensorFlow · Transfer Learning · Transformer · Video Recognition
Part 6 — NLP & Transformers · 10
Attention Is All You Need · Embeddings · Large Language Models · Lemmatization · Natural Language Processing · Retrieval-Augmented Generation · Stemming · Tokenization · Vector Database · Word Embeddings
Part 7 — Reinforcement Learning · 3
Actor-Critic Methods · Policy Gradient · Q-Learning
Part 8 — Data Engineering · 13
Apache Airflow · Apache Kafka · Apache Spark · Batch Processing · Data Modeling · Data Warehouse · dbt · Descriptive Analytics · Diagnostic Analytics · dplyr · ETL / ELT · Stream Processing · Web Scraping
Part 9 — Systems & MLOps · 7
CAP Theorem · CI/CD · Distributed Systems · Docker · Kubernetes · MLOps · Model Deployment
Part 10 — AI Systems · 7
Agentic AI · Agentic Commerce · AI Agents · Embodied AI · GEO / AEO · Prompt Engineering · RAG Reranker Recall Experiment
Part 11 — AI & Cybersecurity · 17
Adversarial Attacks · Agentic AI Security · AI Red Teaming · AI Red Teaming · AI Security · AI Security Governance · Anomaly Detection · Data Poisoning · DefenseClaw · Detection Engineering · Identity & Access Management · OpenAnt · Prompt Injection · Promptfoo · PyRIT · Security Operations Center · Threat Intelligence
Part 12 — Responsible AI · 10
AI Fairness 360 · Algorithmic Auditing · Barocas et al. — Fairness and ML · Bias Mitigation · Datasheets for Datasets · Fairlearn · Fairness in ML · Gebru et al. (2021) — Datasheets · Mitchell et al. (2019) — Model Cards · Model Cards
Part 13 — Strategy & Time Management · 11
Commons Governance · Digital Dispossession · Evidence Log · First-Party Data · Late Capitalism · Marketing Mix (7P / 4C) · Technological Sovereignty · Three Clocks · Time Boxing · Unipolar to Multiplex · Work Slots
Part 14 — Graph & Network Science · 2
Graph Theory · Network Science
Part 15 — Computational Neuroscience · 22
Action Potential (Spike) · AdEx Neuron · Bellec et al. (2020) — e-prop · Brain-Computer Interface · Computational Neuroscience · e-prop · Energy Efficiency of AI · Event-Based Vision · Hodgkin & Huxley (1952) · Hodgkin-Huxley Model · Intel Loihi 2 · Izhikevich Neuron · Leaky Integrate-and-Fire (LIF) · Membrane Potential · Neftci et al. (2019) — Surrogate Gradients · Neuromorphic Computing · snnTorch · Sparse Computation · Spiking Neural Network (SNN) · STDP · Surrogate Gradient Learning · Synaptic Conductance
Part 16 — Connectomics · 8
Connectome-Constrained Network · Connectomics · Dorkenwald et al. (2024) — FlyWire · Electron Microscopy · FlyWire Connectome · Hybrid Neuro Architecture · Neuron Segmentation · Synapse Detection
Part 17 — Concurrent Systems (BEAM) · 6
Actor Model · Backpressure · BEAM VM · Elixir · ETS · GenServer
Part 18 — References · 6
Ashish Vaswani · ggplot2 · Matplotlib · Seaborn · Tableau · Wulfram Gerstner
Appendices: Projects · Central & adjacent · Open questions · Gaps
The mathematics every model rests on — linear algebra, calculus, probability, optimisation.
Reverse-mode automatic differentiation over the network graph, applying the chain rule once per layer. It makes gradient descent tractable for billions of parameters. In machine learning, backpropagation is a gradient computation method commonly used for training a [[Neural Network]] in computing parameter updates.
Updates a prior belief with evidence to yield a posterior. The engine behind Naive Bayes classifiers, Bayesian inference, and principled reasoning under uncertainty. Bayes' theorem, named after Thomas Bayes, gives a mathematical rule for inverting conditional probabilities, allowing the [[Probability]] of a cause to be found given its effect. For example, with Bayes' theorem, the [[Probability]] that a patient has a disease given that they tested positive for that disease can be found using the [[Probability]] that the test yields a positive result when the disease is present. The theorem was developed in the 18th century by Bayes and independently by Pierre-Simon Laplace.
Derivatives, the chain rule, partial derivatives, and integrals. Training any differentiable model is calculus applied millions of times: gradient descent follows the derivative of a loss surface downhill. In mathematical logic, the lambda calculus is a formal system for expressing computation based on function abstraction and application using variable binding and substitution. Untyped lambda calculus, the topic of this article, is a universal machine, i.e. a model of computation that can be used to simulate any Turing machine. It was introduced by the mathematician Alonzo Church in the 1930s as part of his research into the foundations of mathematics. In 1936, Church found a formulation which was logically consistent, and documented it in 1940.
Logic, sets, combinatorics, and graph theory. Captured but deliberately left unconnected — it sits in the orphan queue as a reminder that a node without edges is invisible to the graph.
Factors a square matrix into eigenvectors and eigenvalues — the directions it merely scales. Foundation for spectral methods and a stepping stone to the SVD.
Iteratively step parameters against the loss gradient. SGD, momentum, and Adam are the variants that actually train every model here. Gradient descent is a method for unconstrained mathematical optimization. It is a first-order iterative algorithm for minimizing a differentiable multivariate function.
The gradient, Jacobian, and Hessian generalise the derivative to many dimensions. The gradient points uphill on the loss surface; optimisation walks the opposite way.
Vectors, matrices, and the operations between them — matrix multiplication, rank, eigenvalues, and factorisations. It is the substrate under everything else here: a neural network's weights, PCA's projection, and a transformer's attention are all linear algebra at scale. Linear algebra is the branch of mathematics concerning linear equations such as
Equations relating a function to its derivatives. Every biophysical neuron model (Hodgkin-Huxley, LIF, AdEx) is an ODE integrated over time. In mathematics, an ordinary differential equation (ODE) is a differential equation (DE) dependent on only a single independent variable. As with any other DE, its unknown(s) consists of one function(s) and involves the derivatives of those functions. The term "ordinary" is used in contrast with partial differential equations (PDEs) which may be with respect to more than one independent variable, and, less commonly, in contrast with stochastic differential equations (SDEs) where the modeled process is random.
Random variables, distributions, expectation, and conditional probability. The language for uncertainty that underlies statistics, evaluation metrics, Bayesian methods, and reinforcement learning. Probability and [[Statistics]] are two closely related fields in mathematics that are sometimes combined for academic purposes. They are covered in multiple articles and lists:Probability [[Statistics]] Glossary of probability and [[Statistics]] Notation in probability and [[Statistics]] Timeline of probability and [[Statistics]]
Factors any matrix into rotation · scale · rotation. The workhorse behind low-rank approximation, PCA, and latent-factor models.
The Python data stack and the query language under every warehouse.
Arrays, hash maps, trees, graphs, and the complexity analysis to reason about them. The difference between code that works on a sample and code that works at scale. In mathematics and computer science, an algorithm is a finite sequence of mathematically rigorous logical instructions, typically used to solve a class of specific problems or to perform a computation. Algorithms are used as specifications for performing calculations and data processing. More advanced algorithms can use conditionals to divert the code execution through various routes and deduce valid inferences.
Vectorised n-dimensional arrays with C-speed operations. The numerical foundation the rest of the Python data stack is built on. NumPy is a library for the [[Python]] programming language, adding support for large, multi-dimensional arrays and matrices, along with a large collection of high-level mathematical functions to operate on these arrays. The predecessor of NumPy, Numeric, was originally created by Jim Hugunin with contributions from several other developers. In 2005, Travis Oliphant created NumPy by incorporating features of the competing Numarray into Numeric, with extensive modifications. NumPy is open-source software and has many contributors. NumPy is fiscally sponsored by NumFOCUS.
Labelled tabular data with fast group-by, join, and reshape. Where most exploratory data analysis and feature engineering actually happens.
The lingua franca of data and ML work: readable, batteries-included, and the home of NumPy, pandas, scikit-learn, and every deep-learning framework.
Declarative querying of relational data: joins, aggregation, window functions, indexing. Still the most durable data skill — every warehouse and transform tool speaks it.
Inference, hypothesis testing, and running A/B experiments that actually detect effects.
Randomised controlled experiments on product metrics. Sensitivity is the bottleneck — see CUPED — and ratio metrics need care with the delta method. A/B testing is a user-experience research method. A/B tests consist of a randomized experiment that usually involves two variants, although the concept can be also extended to multiple variants of the same variable. It includes application of statistical [[Hypothesis Testing]] or "two-sample [[Hypothesis Testing]]" as used in the field of [[Statistics]]. A/B testing is employed to compare multiple versions of a single variable, for example by testing a subject's response to variant A against variant B, and to determine which of the variants is more effective.
Sample means of i.i.d. draws converge to a normal distribution as n grows — the bridge that lets you do inference on a single sample. In [[Probability]] theory, the central limit theorem (CLT) states that, under appropriate conditions, the distribution of a normalized version of the sample mean converges to a standard normal distribution. This holds even if the original variables themselves are not normally distributed. There are several versions of the CLT, each applying in the context of different conditions.
Mean, median, and mode — the different ways to summarise where a distribution's mass sits, and when each one lies.
Measuring and testing association between variables — and why correlation is not causation.
A first-order Taylor approximation for the variance of a function of estimators. The standard fix for correct confidence intervals on ratio metrics. In [[Statistics]], the delta method is a method of deriving the asymptotic distribution of a random variable. It is applicable when the random variable being considered can be defined as a differentiable function of a random variable which is asymptotically Gaussian. More generally, the delta method applies to Hadamard directionally differentiable functionals of stochastic processes that converge to a limiting process.
Skewness and kurtosis — asymmetry and tail weight — and what they break in the Gaussian assumptions behind most tests. In [[Statistics]], the concept of the shape of a [[Probability]] distribution arises in questions of finding an appropriate distribution to use to model the statistical properties of a population, given a sample from that population. The shape of a distribution may be considered either descriptively, using terms such as "J-shaped", or numerically, using quantitative measures such as skewness and kurtosis.
Applying statistical models to economic data — regression, time series, and causal identification. Econometrics is an application of statistical methods to economic data in order to give empirical content to economic relationships. More precisely, it is "the quantitative analysis of actual economic phenomena based on the concurrent development of theory and observation, related by appropriate methods of inference." An introductory economics textbook describes econometrics as allowing economists "to sift through mountains of data to extract simple relationships." Jan Tinbergen is one of the two founding fathers of econometrics.
Null vs alternative, p-values, and error rates. The formal machinery for deciding whether an observed effect is signal or noise. A statistical hypothesis test is a method of statistical inference used to decide whether the data provide sufficient evidence to reject a particular hypothesis. A statistical hypothesis test typically involves a calculation of a test statistic. Then a decision is made, either by comparing the test statistic to a critical value or equivalently by evaluating a p-value computed from the test statistic. Roughly 100 specialized statistical tests are in use.
Deng et al., 'Improving the Sensitivity of Online Controlled Experiments' — CUPED regresses out pre-experiment covariates to cut variance, buying large sensitivity gains at Booking and Netflix without extra traffic. [[A/B Testing]] is a user-experience research method. A/B tests consist of a randomized experiment that usually involves two variants, although the concept can be also extended to multiple variants of the same variable. It includes application of statistical [[Hypothesis Testing]] or "two-sample [[Hypothesis Testing]]" as used in the field of [[Statistics]]. [[A/B Testing]] is employed to compare multiple versions of a single variable, for example by testing a subject's response to variant A against variant B, and to determine which of the variants is more effective.
The smallest effect an experiment can reliably catch given its sample size. Fixing the MDE up front is how you size a test honestly. In frequentist [[Statistics]], power is the [[Probability]] of detecting an effect given that some prespecified effect actually exists using a given test in a given context. In typical use, it is a function of the specific test that is used, the sample size, and the effect size.
Metrics that are ratios of two random quantities (e.g. clicks per session). Their variance can't be computed naively — the randomisation unit and the metric unit differ. Web analytics is the measurement, collection, analysis, and reporting of web data to understand and optimize web usage. Web analytics is not just a process for measuring web traffic but can be used as a tool for business and market research and assess and improve website effectiveness. Web analytics applications can also help companies measure the results of traditional print or broadcast advertising campaigns. It can be used to estimate how traffic to a website changes after launching a new advertising campaign.
The probability a test detects a real effect. Low power is why underpowered A/B tests 'find nothing' and quietly waste weeks.
Estimation, inference, and sampling — turning data into defensible claims. Underlies experiment design, model evaluation, and every 'is this difference real?' question.
Data indexed by time, with autocorrelation, trend, and seasonality that break the i.i.d. assumption. Demands its own models and validation splits.
Classic supervised and unsupervised algorithms, features, and honest evaluation.
An encoder–decoder trained to reconstruct its input through a bottleneck, learning a compressed representation. A non-linear cousin of PCA and a route to generative models. In machine learning, deep learning (DL) focuses on utilizing multilayered neural networks to perform tasks such as classification, regression, and representation learning. The field takes inspiration from biological neuroscience and revolves around stacking artificial neurons into layers and "training" them to process data. The adjective "deep" refers to the use of multiple layers in the network. Methods used can be supervised, semi-supervised or unsupervised.
The 2×2 (or k×k) table of predicted vs actual classes. Every classification metric is just a ratio read off it. A confusion matrix, also known as error matrix, is a specific table layout that allows visualization of the performance of a person or an algorithm on a specific task. In audiology these matrices measure how a person can hear specific words or sounds. In machine learning these matrices show the success of the learning system both in [[Supervised Learning]] and [[Unsupervised Learning]], where they are usually called matching matrix.
K-fold resampling to estimate out-of-sample error honestly. The main guard against fooling yourself with the training set. Machine learning (ML) is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without being explicitly programmed. Advances in the field of deep learning have allowed neural networks, a class of statistical algorithms, to surpass many previous machine learning approaches in performance.
Fixing missing, duplicate, and inconsistent data before it poisons downstream models. Machine learning (ML) is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without being explicitly programmed. Advances in the field of deep learning have allowed neural networks, a class of statistical algorithms, to surpass many previous machine learning approaches in performance.
Recursive axis-aligned splits that carve the feature space into pure regions. Interpretable alone and the building block of the strongest tabular ensembles.
Projecting high-dimensional data to fewer dimensions while keeping structure. Fights the curse of dimensionality and makes data visualisable. Dimensionality reduction, or dimension reduction, is the transformation of data from a high-dimensional space into a low-dimensional space so that the low-dimensional representation retains some meaningful properties of the original data, ideally close to its intrinsic dimension. Working in high-dimensional spaces can be undesirable for many reasons; raw data are often sparse as a consequence of the curse of dimensionality, and analyzing the data is usually computationally intractable.
Combine many weak learners — bagging, boosting, stacking — into one strong predictor; random forests and gradient boosting are the flagship cases. In [[Statistics]] and machine learning, ensemble methods use multiple learning algorithms to obtain better predictive performance than could be obtained from any of the constituent learning algorithms alone. Unlike a statistical ensemble in statistical mechanics, which is usually infinite, a machine learning ensemble consists of only a concrete finite set of alternative models, but typically allows for much more flexible structure to exist among those alternatives.
Encoding, interactions, aggregations, and domain transforms — turning raw data into signal a model can use. Still the highest-leverage lever on tabular problems. In supervised machine learning and statistical modeling, feature engineering is a preprocessing step which transforms raw data into a more effective set of inputs. Each input comprises several attributes, known as features. By providing models with relevant information, feature engineering significantly enhances their predictive accuracy and decision-making capability.
Standardisation and normalisation so features share a comparable range. Essential for distance- and gradient-based methods; irrelevant to trees. Feature scaling is a method used to normalize the range of independent variables or features of data. In data processing, it is also known as data normalization and is generally performed during the data preprocessing step.
Keeping the features that carry signal and dropping the rest — filter, wrapper, and embedded methods. Fewer, better features means less overfitting and faster models. In machine learning, feature selection is the process of selecting a subset of relevant features for use in model construction. Feature selection techniques are used for several reasons:simplification of models to make them easier to interpret, shorter training times, to avoid the curse of dimensionality, improve the compatibility of the data with a certain learning model class, to encode inherent symmetries present in the input space.
Fits trees sequentially, each correcting the residual errors of the last. XGBoost/LightGBM dominate tabular competitions. Gradient boosting is a machine learning technique based on boosting in a functional space, where the target is pseudo-residuals instead of residuals as in traditional boosting. It gives a prediction model in the form of an ensemble of weak prediction models, i.e., models that make very few assumptions about the data, which are typically simple [[Decision Trees]]. When a decision tree is the weak learner, the resulting algorithm is called gradient-boosted trees; it usually outperforms [[Random Forest]].
Searching the hyperparameter space (grid, random, Bayesian) without leaking test information. In machine learning, a hyperparameter is a parameter that can be set in order to define any configurable part of a model's learning process. Hyperparameters can be classified as either model hyperparameters or algorithm hyperparameters. These are named hyperparameters in contrast to parameters, which are characteristics that the model learns from the data.
Partitions points into k clusters by iteratively minimising within-cluster variance. Fast, simple, and sensitive to initialisation and the choice of k. k-means clustering is a method of vector quantization, originally from signal processing, that aims to partition n observations into k clusters in which each observation belongs to the cluster with the nearest mean. This results in a partitioning of the data space into Voronoi cells. k-means clustering minimizes within-cluster variances, but not regular Euclidean distances, which would be the more difficult Weber problem: the mean optimizes squared errors, whereas only the geometric median minimizes Euclidean distances. For instance, better Euclidean solutions can be found using k-medians and k-medoids.
Classifies by majority vote of the closest training points — no training, all inference. A useful baseline and a lens on the curse of dimensionality. In [[Statistics]] and machine learning, the k-nearest neighbors algorithm (k-NN) is a non-parametric [[Supervised Learning]] method that assigns weightage only to the k nearest neighbors of an entity in making a decision about the entity. It is used both in classification -- where a new example is assigned a label based on the labels of its k nearest training examples; and in regression -- where the prediction is computed from the values of those neighbors. Its more frequent use is for classification, as the k-NN classifier, the output of which is a class membership decided by a plurality vote of its neighbors.
Fits a linear map by least squares — the simplest supervised baseline. Everything more complex is measured against how much it beats it. In machine learning and pattern recognition, a feature is an individual measurable property or characteristic of a data set. Choosing informative, discriminating, and independent features is crucial to producing effective algorithms for pattern recognition, classification, and regression tasks. Features are usually numeric, but other types such as strings and graphs are used in syntactic pattern recognition, after some pre-processing step such as one-hot encoding. The concept of "features" is related to that of explanatory variables used in statistical techniques such as linear regression.
A linear model with a sigmoid link for classification, trained by gradient descent. Strong, interpretable, and still a smart first model for many problems. In [[Statistics]], a logistic model is a statistical model that models the log-odds of an event as a linear combination of one or more independent variables. In regression analysis, logistic regression estimates the parameters of a logistic model. In binary logistic regression there is a single binary dependent variable, coded by an indicator variable, where the two values are labeled "0" and "1", while the independent variables can each be a binary variable or a continuous variable.
Choosing and reading the right metric — accuracy, precision/recall, ROC-AUC, log loss. Picking the metric is really picking the problem you care about.
Choosing among candidate models honestly — nested CV, information criteria, and the bias–variance framing. Model selection is the task of selecting a model from among various candidates on the basis of performance criterion to choose the best one. In the context of machine learning and more generally statistical analysis, this may be the selection of a statistical model from a set of candidate models, given data. In the simplest cases, a pre-existing set of data is considered. However, the task can also involve the design of experiments such that the data collected is well-suited to the problem of model selection.
Linear regression on polynomial features — captures curvature while staying linear in the parameters. A quick lesson in the bias–variance trade-off. In [[Statistics]], polynomial regression is a form of regression analysis in which the relationship between the independent variable x and the dependent variable y is modeled as a polynomial in x. Polynomial regression fits a nonlinear relationship between the value of x and the corresponding conditional mean of y, denoted E(y |x). Although polynomial regression fits a nonlinear model to the data, as a statistical estimation problem it is linear, in the sense that the regression function E(y | x) is linear in the unknown parameters that are estimated from the data.
Precision = of predicted positives, how many are right; recall = of actual positives, how many were caught. The F1 score balances them; imbalance decides which matters. In statistical analysis of binary classification and information retrieval systems, the F-score or F-measure is a measure of predictive performance. It is calculated from the precision and recall of the test, where the precision is the number of true positive results divided by the number of all samples predicted to be positive, including those not identified correctly, and the recall is the number of true positive results divided by the number of all samples that should have been identified as positive.
Forecasting what will happen by fitting models to historical patterns. Predictive analytics encompasses a variety of statistical techniques from data mining, predictive modeling, and machine learning that analyze current and historical facts to make predictions about future or otherwise unknown events.
Recommending what to do about it — optimisation and decision rules on top of prediction.
Linear projection onto the directions of maximum variance, computed via the SVD. The default first move for compression and de-noising. Principal component analysis (PCA) is a linear [[Dimensionality Reduction]] technique with applications in exploratory data analysis, visualization and data preprocessing.
Bagged decision trees with feature subsampling — averages away variance. A robust default that rarely overfits catastrophically. Random forests or random decision forests is an ensemble learning method for classification, regression and other tasks that works by creating a multitude of [[Decision Trees]] during training. For classification tasks, the output of the random forest is the class selected by most trees. For regression tasks, the output is the average of the predictions of the trees. Random forests correct for [[Decision Trees]]' habit of overfitting to their training set.
Ranking items for a user — collaborative filtering, matrix factorisation, and two-tower models. Machine learning (ML) is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without being explicitly programmed. Advances in the field of deep learning have allowed neural networks, a class of statistical algorithms, to surpass many previous machine learning approaches in performance.
Penalising model complexity to curb overfitting: L2 (ridge) shrinks weights, L1 (lasso) drives them to zero for selection, ElasticNet blends both. In mathematics, [[Statistics]], finance, and computer science, particularly in machine learning and inverse problems, regularization is a process that converts the answer to a problem to a simpler one. It is often used in solving ill-posed problems or to prevent overfitting. There is a strong connection between regularization methods and Bayesian approaches for solving such ill-posed problems.
An agent learns a policy by acting and receiving rewards, trading off exploration and exploitation. The framing behind game-playing agents and RLHF. In machine learning and optimal control, reinforcement learning (RL) is concerned with how an intelligent agent should take actions in a dynamic environment in order to maximize a reward signal. Reinforcement learning is one of the three basic machine learning paradigms, alongside [[Supervised Learning]] and [[Unsupervised Learning]].
Plots true- vs false-positive rate across thresholds; the area under it summarises ranking quality independent of the cutoff. Careful under heavy class imbalance. A receiver operating characteristic curve, or ROC curve, is a graphical plot that illustrates the performance of a binary classifier model at varying threshold values. ROC analysis is commonly applied in the assessment of diagnostic test performance in clinical epidemiology.
The classic-ML toolbox — a consistent fit/predict API across every algorithm in this graph. scikit-learn is a free and open-source machine learning library for the [[Python]] programming language. It features various classification, regression and clustering algorithms including support-vector machines, random forests, [[Gradient Boosting]], k-means and DBSCAN, and is designed to interoperate with the [[Python]] numerical and scientific libraries [[NumPy]] and SciPy. Scikit-learn is a NumFOCUS fiscally sponsored project.
Generates its own labels from the data (predict the masked token, the next frame). The pretraining paradigm behind modern foundation models.
Learning a mapping from inputs to labelled outputs — classification and regression. The most common and best-understood setting.
Finds the maximum-margin separating hyperplane, and via the kernel trick, non-linear boundaries. Strong on small, high-dimensional data.
Finding structure without labels — clustering, dimensionality reduction, density estimation. Useful when labels are scarce or the goal is discovery.
Neural networks from the perceptron to CNNs and recurrent models.
The non-linearities — ReLU, sigmoid, tanh, GELU — that let stacked layers represent more than a single linear map. Their shape governs gradient flow. In artificial neural networks, the activation function of a node is a function that calculates the output of the node based on its individual inputs and their weights. Nontrivial problems can be solved using only a few nodes if the activation function is nonlinear.
A content-based weighted lookup over a set of representations — every position can attend to every other. Removing the sequential bottleneck of RNNs is the pivot of modern NLP. In machine learning, attention is a method that determines the importance of each component in a sequence relative to the other components in that sequence. In [[Natural Language Processing]], importance is represented by "soft" weights assigned to each word in a sentence. More generally, attention encodes vectors called token [[Embeddings]] across a fixed-width sequence that can range from tens to millions of tokens in size.
A sliding weighted window that exploits translation invariance and shares parameters. The prior that makes vision models sample-efficient. In mathematics, convolution is a mathematical operation on two functions and that produces a third function , as the integral of the product of the two functions after one is reflected about the y-axis and shifted. The term convolution refers to both the resulting function and to the process of computing it. The integral is evaluated for all values of shift, producing the convolution function. The choice of which function is reflected and shifted before the integral does not change the integral result. Graphically, it expresses how the 'shape' of one function is modified by the other.
Stacks of convolution + pooling that learn a hierarchy of visual features. AlexNet on ImageNet in 2012 kicked off the deep-learning era. A feedforward [[Neural Network]] is an artificial [[Neural Network]] in which information flows in a single direction – inputs are multiplied by weights to obtain outputs (inputs-to-output). It contrasts with a [[Recurrent Neural Network]], in which loops allow information from later processing stages to feed back to earlier stages. Feedforward multiplication is essential for [[Backpropagation]], because feedback, where the outputs feed back to the very same inputs and modify them, forms an infinite loop which is not possible to differentiate through [[Backpropagation]].
Replaces the Q-table with a neural network, plus experience replay and target networks for stability. DeepMind's Atari result that reignited deep RL. In machine learning, deep learning (DL) focuses on utilizing multilayered neural networks to perform tasks such as classification, regression, and representation learning. The field takes inspiration from biological neuroscience and revolves around stacking artificial neurons into layers and "training" them to process data. The adjective "deep" refers to the use of multiple layers in the network. Methods used can be supervised, semi-supervised or unsupervised.
Continuing training on your own data (SFT, LoRA) to bend a base model toward a task or style. A large language model (LLM) is an AI model trained on a vast amount of text for [[Natural Language Processing]] tasks, especially language generation. LLMs can typically generate, summarize, translate, and analyze text in many contexts. They are the basis for many modern chatbots, such as ChatGPT, Claude, Gemini, Grok, and DeepSeek.
Running inputs through the network layer by layer to produce an output and a loss. The pass whose intermediate values backprop later reuses. In machine learning, [[Backpropagation]] is a gradient computation method commonly used for training a [[Neural Network]] in computing parameter updates.
A generator and discriminator trained against each other until the fakes are indistinguishable. Produced the first photorealistic image synthesis. A generative adversarial network (GAN) is a class of machine learning frameworks and a prominent framework for approaching generative AI. The concept was initially developed by Ian Goodfellow and his colleagues in June 2014. In a GAN, two neural networks compete with each other in the form of a zero-sum game, where one agent's gain is another agent's loss.
Assigning a whole image to one of k categories — the canonical benchmark task for vision models and the one that proved deep learning at scale. Computer vision tasks include methods for acquiring, processing, analyzing, and understanding digital images, and extraction of high-dimensional data from the real world in order to produce numerical or symbolic information, e.g. in the form of decisions. "Understanding" in this context signifies the transformation of visual images into descriptions of the world that make sense to thought processes and can elicit appropriate action.
Classifying every pixel — U-Net and friends — the bridge from 'what is here' to 'exactly where'. In digital image processing and computer vision, image segmentation is the process of partitioning a digital image into multiple image segments, also known as image regions or image objects. The goal of segmentation is to simplify and/or change the representation of an image into something that is more meaningful and easier to analyze. Image segmentation is typically used to locate objects and boundaries in images. More precisely, image segmentation is the process of assigning a label to every pixel in an image such that pixels with the same label share certain characteristics.
1.2M labelled images across 1000 classes. The benchmark that made CNN progress legible and directly comparable year over year. The ImageNet project is a large visual database designed for use in visual object recognition software research. More than 14 million images have been hand-annotated by the project to indicate what objects are pictured and in at least one million of the images, bounding boxes are also provided. ImageNet contains more than 20,000 categories, with a typical category, such as "balloon" or "strawberry", consisting of several hundred images. The database of annotations of third-party image URLs is freely available directly from ImageNet, though the actual images are not owned by ImageNet.
The high-level deep-learning API — now the front door to TensorFlow and JAX. Keras is an open-source library that provides a [[Python]] interface for artificial neural networks. Keras was first independent software, then integrated into the [[TensorFlow]] library, and later added support for more. "Keras 3 is a full rewrite of Keras [and can be used] as a low-level cross-framework language to develop custom components such as layers, models, or metrics that can be used in native workflows in JAX, [[TensorFlow]], or [[PyTorch]] — with one codebase." Keras 3 will be the default Keras version for [[TensorFlow]] 2.16 onwards, but Keras 2 can still be used.
The objective the optimiser minimises — MSE for regression, cross-entropy for classification. Choosing it encodes what 'wrong' means. In computability theory, a primitive recursive function is, roughly speaking, a function that can be computed by a computer program whose loops are all "for" loops. Primitive recursive functions form a strict subset of those general recursive functions that are also total functions.
An RNN with gated memory cells that let gradients survive long sequences. The workhorse of sequence modelling before transformers. In machine learning, deep learning (DL) focuses on utilizing multilayered neural networks to perform tasks such as classification, regression, and representation learning. The field takes inspiration from biological neuroscience and revolves around stacking artificial neurons into layers and "training" them to process data. The adjective "deep" refers to the use of multiple layers in the network. Methods used can be supervised, semi-supervised or unsupervised.
Runs several attention heads in parallel, each learning a different relation, then concatenates them. More representational capacity at the same asymptotic cost. In machine learning, [[Attention]] is a method that determines the importance of each component in a sequence relative to the other components in that sequence. In [[Natural Language Processing]], importance is represented by "soft" weights assigned to each word in a sentence. More generally, [[Attention]] encodes vectors called token [[Embeddings]] across a fixed-width sequence that can range from tens to millions of tokens in size.
Stacked fully-connected layers with non-linear activations — a universal function approximator. The baseline every specialised architecture refines. In deep learning, a multilayer [[Perceptron]] (MLP) is a kind of modern feedforward [[Neural Network]] consisting of fully connected neurons with nonlinear [[Activation Functions]], organized in layers, notable for being able to distinguish data that is not linearly separable.
Composed layers of differentiable linear maps and non-linearities, trained end-to-end by gradient descent. The common ancestor of the MLP, CNN, RNN, and transformer.
A single linear threshold unit — the 1958 ancestor of deep learning. Its inability to solve XOR motivated stacking layers.
Downsamples feature maps (max or average) to add spatial invariance and cut compute. The other half of the classic conv block. Graph neural networks (GNNs) are artificial neural networks designed for tasks whose inputs are graphs.
Meta's eager, Pythonic deep-learning framework — the research default. PyTorch is an open-source deep learning library, originally developed by Meta Platforms and currently developed with support from the Linux Foundation. The successor to Torch, PyTorch provides a high-level API that builds upon optimised, low-level implementations of deep learning algorithms and architectures, such as the [[Transformer]], or SGD. Notably, this API simplifies model training and inference to a few lines of code. PyTorch allows for automatic parallelization of training and, internally, implements CUDA bindings that speed training further by leveraging GPU resources.
Carries a hidden state across time steps to model sequences. Elegant but hard to train over long ranges — vanishing gradients — and now largely superseded by attention.
Attention where queries, keys, and values all come from the same sequence, letting each token mix in information from the rest. The core operation of the transformer block.
Google's end-to-end deep-learning framework — static graphs, the Keras API, and production serving.
Reusing a model trained on one task as the starting point for another — why pretrained backbones dominate vision and NLP.
Self-attention + feed-forward blocks with residual connections and no recurrence, so the whole sequence trains in parallel. The backbone of every modern LLM and of the Open Intelligence Stack work. In deep learning, the transformer is a family of artificial [[Neural Network]] architectures based on the [[Multi-Head Attention]] mechanism, in which input data such as text, images, or audio, is converted to a sequence of numerical representations called tokens, and each token is converted into a vector via lookup from a word embedding table. At each layer, each token is then contextualized within the scope of the context window with other (unmasked) tokens via a parallel [[Multi-Head Attention]] mechanism, allowing the signal for key tokens to be amplified and less important tokens to be diminished.
Understanding frames in time — spatial plus temporal modelling, where CNNs meet sequence models.
Tokenisation, embeddings, attention, and the transformer architecture.
Vaswani et al., 2017. Introduces the transformer — self-attention and feed-forward blocks, no recurrence or convolution — the architecture every current LLM descends from. The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an [[Attention]] mechanism. We propose a new simple network architecture, the [[Transformer]], based solely on [[Attention]] mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train.
Dense vectors where geometric proximity encodes semantic similarity. The bridge from discrete text to everything numeric — retrieval, clustering, and RAG. An embedded system is a specialized computer system—a combination of a computer processor, computer memory, and input/output peripheral devices—that has a dedicated function within a larger mechanical or electronic system. It is embedded as part of a complete device, often including electrical or electronic hardware and mechanical parts.
Transformers scaled to web corpora — emergent in-context learning and the substrate of every agent and RAG stack. A large language model (LLM) is an AI model trained on a vast amount of text for [[Natural Language Processing]] tasks, especially language generation. LLMs can typically generate, summarize, translate, and analyze text in many contexts. They are the basis for many modern chatbots, such as ChatGPT, Claude, Gemini, Grok, and DeepSeek.
Reducing words to their dictionary lemma using morphology and part-of-speech. More accurate than stemming, and heavier. [[Natural Language Processing]] (NLP) is the processing of natural language information by a computer. NLP is a subfield of computer science and is closely associated with artificial intelligence. NLP is also related to information retrieval, knowledge representation, computational linguistics, and linguistics more broadly.
Getting computers to work with human language — the pipeline from raw text through tokens, embeddings, and models to output. Transformed wholesale by attention. Natural language processing (NLP) is the processing of natural language information by a computer. NLP is a subfield of computer science and is closely associated with artificial intelligence. NLP is also related to information retrieval, knowledge representation, computational linguistics, and linguistics more broadly.
Retrieval-Augmented Generation conditions the model on retrieved context instead of parameters alone — grounding answers and side-stepping stale weights. The central pattern of the Open Intelligence Stack. A [[Vector Database]], vector store or vector search engine is a database that stores and retrieves [[Embeddings]] of data in vector space. Vector databases typically implement approximate nearest neighbor algorithms so users can search for records semantically similar to a given input, unlike traditional databases which primarily look up records by exact match. Use-cases for vector databases include similarity search, semantic search, multi-modal search, recommendations engines, object detection, and retrieval-augmented generation (RAG).
Crudely chopping words to a root ('running' → 'run') by rule. Fast and lossy — a classical-NLP normalisation step largely retired by subword tokenisation.
Splitting text into subword units (BPE, WordPiece, Unigram). The unglamorous layer that quietly bounds a model's vocabulary and its behaviour on rare words.
An approximate-nearest-neighbour index over embeddings (HNSW, IVF). The retrieval substrate that makes RAG fast at corpus scale.
Static word vectors (word2vec, GloVe) learned from co-occurrence. The precursor to contextual embeddings from transformers.
Learning policies from reward — value-based and policy-gradient methods.
Learn a policy and a value function together — the family (A2C, PPO) behind most modern deep RL. The actor-critic algorithm (AC) is a family of [[Reinforcement Learning]] (RL) algorithms that combine policy-based RL algorithms such as [[Policy Gradient]] methods, and value-based RL algorithms such as value iteration, [[Q-Learning]], SARSA, and TD learning.
Optimises the policy directly by ascending the gradient of expected reward. The family (REINFORCE, PPO) behind continuous control and RLHF. Stochastic [[Gradient Descent]] is an iterative method for optimizing an objective function with suitable smoothness properties. It can be regarded as a stochastic approximation of [[Gradient Descent]] optimization, since it replaces the actual gradient by an estimate thereof. Especially in high-dimensional optimization problems this reduces the very high computational burden, achieving faster iterations in exchange for a lower convergence rate.
Learns an action-value table by bootstrapping from the Bellman equation, off-policy. The classic entry point to value-based RL. Q-learning is a [[Reinforcement Learning]] algorithm that trains an agent to assign values to its possible actions based on its current state, without requiring a model of the environment (model-free). It can handle problems with stochastic transitions and rewards without requiring adaptations.
Distributed batch and streaming pipelines, warehouses, and orchestration.
DAG-based scheduling and orchestration with retries, backfills, and lineage. Runs the batch and dbt jobs of the platform on a dependable schedule. Apache Airflow is an open-source workflow management platform for data engineering pipelines. It started at Airbnb in October 2014 as a solution to manage the company's increasingly complex workflows. Creating Airflow allowed Airbnb to programmatically author and schedule their workflows and monitor them via the built-in Airflow user interface. From the beginning, the project was made open source, becoming an Apache Incubator project in March 2016 and a top-level Apache Software Foundation project in January 2019.
A durable, partitioned commit log for event streams. The spine between producers and stream processors, decoupling them in time. Apache Kafka is a distributed event store and stream-processing platform. It is an open-source system developed by the Apache Software Foundation written in Java and Scala. The project aims to provide a unified, high-throughput, low-latency platform for handling real-time data feeds. Kafka can connect to external systems via Kafka Connect, and provides the Kafka Streams libraries for [[Stream Processing]] applications. Kafka uses a binary TCP-based protocol that is optimized for efficiency and relies on a "message set" abstraction that naturally groups messages together to reduce the overhead of the network roundtrip.
Distributed in-memory compute over DataFrames and RDDs. The workhorse for large-scale batch transformation in the data platform.
Processing bounded datasets on a schedule. Simpler and cheaper than streaming when freshness of minutes-to-hours is fine.
Structuring data for its use — normalisation for transactions, star/snowflake schemas and slowly-changing dimensions for analytics. The Courant Institute School of Mathematics, Computing, and Data Science, previously known as the Courant Institute of Mathematical Sciences (CIMS), is a constituent school of New York University.
A columnar analytics store (BigQuery, Snowflake, Redshift) optimised for OLAP scans over history. The serving layer of the batch platform. In computing, a data warehouse, also known as an enterprise data warehouse (EDW), is a system used for reporting and data analysis and is a core component of business intelligence. Data warehouses are central repositories of data integrated from disparate sources. They store current and historical data organized in a way that is optimized for data analysis, generation of reports, and developing insights across the integrated data. They are intended to be used by analysts and managers to help make organizational decisions.
SQL-first transformation with tests, docs, and lineage inside the warehouse. Declared level 2 here — the shipped platform's grant lifts it higher. DBT may refer to:Danish Board of Technology, a technology assessment institution in Denmark data build tool (dbt), a data analytics tool .dbt, the extension of a DBase file format DBT (gene) DBT Online Inc.
Summarising what happened — aggregation, visualisation, and descriptive statistics. Mathematics is a broad subject that is commonly divided in many areas or branches that may be defined by their objects of study, by the used methods, or by both. For example, analytic number theory is a subarea of number theory devoted to the use of methods of analysis for the study of natural numbers.
Why did it happen — drill-down, correlation, and root-cause analysis. This is a list of women who have made noteworthy contributions to or achievements in mathematics. These include mathematical research, mathematics education, the history and philosophy of mathematics, public outreach, and mathematics contests.
R's data-manipulation grammar — filter, mutate, group, summarise. dplyr is an R package whose set of functions are designed to enable dataframe manipulation in an intuitive, user-friendly way. It is one of the core packages of the popular tidyverse set of packages in the R programming language. Data analysts typically use dplyr in order to transform existing datasets into a format better suited for some particular type of analysis, or data visualization.
Extract, transform, load — moving data from sources into a queryable store. ELT flips the order, transforming inside the warehouse with tools like dbt. In computing, a [[Data Warehouse]], also known as an enterprise [[Data Warehouse]] (EDW), is a system used for reporting and data analysis and is a core component of business intelligence. Data warehouses are central repositories of data integrated from disparate sources. They store current and historical data organized in a way that is optimized for data analysis, generation of reports, and developing insights across the integrated data. They are intended to be used by analysts and managers to help make organizational decisions.
Processing unbounded event streams as they arrive, with windowing and exactly-once concerns. The price of freshness is complexity.
Pulling structured data from the web — requests, parsing, and the robots/ethics boundary.
Distributed-systems fundamentals and shipping models to production.
Under a network partition you must trade consistency against availability — you cannot have both. The lens for choosing a datastore. [[Bayes' Theorem]], named after Thomas Bayes, gives a mathematical rule for inverting conditional probabilities, allowing the [[Probability]] of a cause to be found given its effect. For example, with [[Bayes' Theorem]], the [[Probability]] that a patient has a disease given that they tested positive for that disease can be found using the [[Probability]] that the test yields a positive result when the disease is present. The theorem was developed in the 18th century by Bayes and independently by Pierre-Simon Laplace.
Automated build, test, and deploy on every change. The habit that makes shipping small and reversible instead of scary.
Consensus, partitioning, replication, and failure across many machines. The ground truth every data tool has to negotiate with.
Packages an app and its dependencies into a portable image that runs the same everywhere. The unit of deployment for most modern services. Docker most often refers to:A dockworker, a manual laborer who is involved in loading and unloading ships, also called a longshoreman or stevedore Docker (software), an open-source software project automating the deployment of applications inside software containers Docker, Inc., the company promoting Docker software
Orchestrates containers across a cluster — scheduling, scaling, self-healing. Powerful and heavy; the default for running services at scale. Kubernetes, also known as K8s, is an open-source container orchestration system for automating software deployment, scaling, and management. Originally designed by Google, the project is now maintained by a worldwide community of contributors, and the trademark is held by the Cloud Native Computing Foundation.
Bringing software discipline to ML: versioning data and models, reproducible pipelines, CI/CD, monitoring for drift. The gap between a notebook and a system. MLOps or ML Ops is a paradigm that aims to deploy and maintain machine learning models in production reliably and efficiently. It bridges the gap between machine learning development and production operations, ensuring that models are robust, scalable, and aligned with business goals.
Serving a trained model behind an API or batch job, with versioning and rollback. Where model value is actually realised — or lost to latency and drift. Automated machine learning (AutoML) is the process of automating the tasks of applying machine learning to real-world problems. It is the combination of automation and ML.
Retrieval-augmented generation, vector search, and evaluation of LLM systems.
Autonomous LLM-driven systems that plan, call tools, and act across multiple steps with limited supervision. Raises both capability and new security surfaces.
Comercio donde un agente de IA compra en nombre del humano: Google lanzó el Universal Commerce Protocol (enero 2026, Shopify como codesarrollador) y OpenAI el Agentic Commerce Protocol con Instant Checkout (febrero 2026). Reescribe el funnel: alcance y adquisición ya no describen a un humano navegando, sino a un agente comprando por él.
An LLM in a loop that plans, calls tools, and acts over time — the difference between answering and doing. Artificial intelligence (AI) is the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making. It is a field of research in engineering, mathematics, and computer science that develops and studies methods and software that enable machines to perceive their environment and use learning and intelligence to take actions that maximise their chances of achieving defined goals.
AI embedded in physical agents — robots, drones — that must sense and act in the world under tight latency and energy budgets. A driver of neuromorphic hardware.
Generative / Answer Engine Optimization: el comportamiento zero-click ya es el default y el SEO por palabra clave se desplaza. La métrica nueva no es la posición en el listado, es la citación dentro de una respuesta generada — ser citado por el modelo, no rankeado por el buscador.
Designing inputs that steer a frozen model's behaviour — instructions, few-shot exemplars, and structured output. Prompt engineering is the process of structuring natural language inputs to produce specified outputs from a generative AI model. Context engineering is the related area of software engineering that focuses on the management of non-prompt and prompt contexts supplied to the GenAI model, such as system instructions, metadata, API tools and tokens.
Measuring recall@k with and without a cross-encoder reranker on the local corpus, to decide whether the reranker earns its latency. Part of the Open Intelligence Stack.
Defending at machine speed — prompt injection, red teaming, agent security, and the SOC rebuilt around AI.
Inputs perturbed imperceptibly to flip a model's output. The canonical demonstration that high accuracy does not imply robustness.
Securing agents as systems — context windows, non-determinism, agentic identity, shifting trust boundaries, and the tool/memory surfaces an attacker can reach through them. Artificial intelligence (AI) is the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making. It is a field of research in engineering, mathematics, and computer science that develops and studies methods and software that enable machines to perceive their environment and use learning and intelligence to take actions that maximise their chances of achieving defined goals.
Adversarially probing an AI system before attackers do — automated jailbreaks, fuzzing, and prompt attacks to surface verified weaknesses; the defender's mirror of the attack chain. In the field of artificial intelligence (AI), alignment aims to steer AI systems toward a person's or group's intended goals, preferences, or ethical principles. An AI system is considered aligned if it advances the intended objectives. A misaligned AI system pursues unintended objectives.
Systematically probing an AI system for failures and attacks before adversaries do. Ethical penetration testing adapted to models.
Studying and defending the vulnerabilities specific to AI systems — adversarial inputs, prompt injection, data poisoning, model theft. A fast-growing discipline as models enter critical infrastructure.
Policies and guardrails for AI in an organisation — who may deploy what, red-team gates, runtime memory protection (e.g. OWASP Agent Memory Guard), and human-in-the-loop workflows. Artificial intelligence (AI) is the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making. It is a field of research in engineering, mathematics, and computer science that develops and studies methods and software that enable machines to perceive their environment and use learning and intelligence to take actions that maximise their chances of achieving defined goals.
Flagging events that deviate from learned normal behaviour — the backbone of AI-driven cyberdefense over network, power, and transaction streams.
Corrupting training data to implant backdoors or degrade a model. An attack on the pipeline rather than the deployed model.
Cisco's open-source agent-security framework — scans agent skills, sandboxes them, verifies MCP servers, and inventories AI assets. Claude is a series of [[Large Language Models]] (LLMs) developed by American software company Anthropic. Claude was released as an AI-based chatbot in March 2023. It is also used in AI-assisted software development.
Turning threat knowledge into continuously tuned detections — the discipline where security meets data science: behavioural analytics, false-positive budgets, drift monitoring. Computer science is the study of computation, information, and automation. Included broadly in the sciences, computer science spans theoretical disciplines to applied disciplines. An expert in the field is known as a computer scientist.
Identity as the new perimeter — multi-factor authentication, least privilege, and out-of-band verification for high-value requests; the baseline defence that stops most AI-assisted attacks. The trusted computing base (TCB) of a computer system is the set of all hardware, firmware, and/or software components that are critical to its security, in the sense that bugs or vulnerabilities occurring inside the TCB might jeopardize the security properties of the entire system. By contrast, parts of a computer system that lie outside the TCB must not be able to misbehave in a way that would leak any more privileges than are granted to them in accordance to the system's security policy.
Knostic's open-source LLM-based vulnerability discovery — stage 1 detects, stage 2 attacks; what survives is real, with low false positives. This is a list of free and open-source software (FOSS) packages, computer software licensed under free software licenses and open-source licenses. Software that fits the Free Software Definition may be more appropriately called free software; the GNU project in particular objects to their works being referred to as open-source. For more information about the philosophical background for open-source software, see free software movement and Open Source Initiative. However, nearly all software meeting the Free Software Definition also meets the Open Source Definition and vice versa.
The attacker's main lever on LLM systems: crafting input that overrides system instructions — the OWASP #1 for AI, and the reason agents need guardrails. [[Prompt Engineering]] is the process of structuring natural language inputs to produce specified outputs from a generative AI model. Context engineering is the related area of software engineering that focuses on the management of non-prompt and prompt contexts supplied to the GenAI model, such as system instructions, metadata, API tools and tokens.
Open-source evaluation and red-teaming harness for LLM prompts and agents — regression-test model behaviour and security properties in CI.
Microsoft's open-source red-teaming framework for AI systems — automated attack generators, scorers, and orchestrators for safety-and-security evaluation. The following list contains a list of computer programs that are built to take advantage of the OpenCL or WebCL heterogeneous compute framework.
The SOC — SIEM correlation, MDR/XDR telemetry, alert triage — now being rebuilt around AI co-pilots and autonomous response at machine speed.
Curated structured knowledge about attackers — TTPs, indicators, campaigns — that keeps detection and response ahead of the adversary instead of behind.
Fairness, bias mitigation, auditing, and the documentation that makes models accountable.
IBM's AI Fairness 360: a comprehensive library of fairness metrics and bias-mitigation algorithms across the ML lifecycle.
Independently evaluating a deployed model for bias, safety, and compliance — sometimes adversarially and publicly. Where responsible AI meets security.
Barocas, Hardt & Narayanan, 'Fairness and Machine Learning' (fairmlbook.org). The open-access textbook of the field.
Pre-, in-, and post-processing techniques that reduce disparate outcomes across groups. Every choice trades some accuracy or another fairness metric.
Documentation accompanying a dataset — its motivation, composition, collection, and biases — analogous to a hardware component's datasheet.
An open-source Python toolkit for assessing and improving the fairness of ML models, with metrics and mitigation algorithms.
Detecting and mitigating bias so a model's errors don't fall disproportionately on protected groups. Requires choosing an explicit, often contested, definition of 'fair'.
Gebru et al., 2021, 'Datasheets for Datasets'. Proposed structured documentation to surface a dataset's provenance and biases.
Mitchell et al., 2019, 'Model Cards for Model Reporting' (FAccT). Introduced the model-card standard for transparent reporting.
Standardised documentation reporting a model's intended use, performance across groups, and limitations. Transparency as an artefact shipped with the model.
The operating system of the study plan — three clocks, time-boxed work slots, evidence logging, and the strategic anchors (sovereignty, commerce, geopolitics).
Elinor Ostrom demostró empíricamente que existe gobernanza colectiva de recursos compartidos que no es ni propiedad privada ni tragedia de los comunes. neni.lat — cooperativa con liquidación directa CoDi/SPEI, atestaciones Ed25519 y documentación ante Banxico — es el experimento real de esa tercera vía en la capa de pagos.
La acumulación por desposesión de Harvey aplicada al siglo XXI: el dato personal como común no mercantilizado convertido en propiedad extraíble (Zuboff), y la capa más reciente — entrenar LLMs sobre texto humano sin consentimiento es un cercamiento del commons cultural. El bloqueo anti-scraping es cercar de vuelta lo que el capital cercaba de ti.
Todo slot importante deja evidencia: commit, screenshot, tabla, notebook, propuesta o lección completada. "Estudié dos horas" es evidencia débil; "reproduje el método y documenté el resultado" es fuerte. Cierre en tres líneas: Produced / Blocked by / Next physical action.
El dato propio post-cookie como único activo medible — email marketing como canal que sigues poseyendo en un mundo zero-click, dark social y messy middle matizando el funnel clásico. El bloqueo a crawlers de IA (GPTBot, ClaudeBot) plantea la tensión: proteger el contenido vs. volverse invisible en las respuestas generadas.
Genealogía del concepto: Sombart (1928) lo acuña, Mandel (1972) lo vuelve teoría — fusión de ciencia y producción, capital multinacional —, Jameson le da la lectura cultural, Harvey la acumulación por desposesión, Streeck la descomposición lenta ("mil pequeños cortes") y Fisher el realismo capitalista. "Tardío" no significa terminal.
Las 7 P's (Booms & Bitner: + personas, procesos, evidencia física) y las 4 C's en sus dos versiones (Lauterborn 1990 vs. Kotler Marketing 4.0). En la era agéntica cada P se rompe distinto: el producto debe ser un feed estructurado, el precio verdadero en checkout por API, la promoción se vuelve GEO/AEO, la evidencia física schema markup verificable — y a las 4 C de Lauterborn les falta el quinto vértice: comprensibilidad para máquinas.
Capacidad de desarrollar y controlar infraestructura tecnológica propia sin dependencia externa. Del stack soberano chino (XV Plan Quinquenal) al Cloudflare Worker anti-scraping de finitud.com — el portafolio de dominios como infraestructura de proyecto, no solo narrativa.
Tres relojes que organizan todo el estudio: finalización de la carrera, práctica profesional (finitud.com) e investigación de frontera (tesis ML interpretable). Cada material nuevo se cuelga de uno de los tres en vez de flotar como documento aislado.
Time slot = duración + modo de trabajo + entregable + condición de cierre. "Para los próximos 90 minutos, produce un artefacto revisable y detente cuando acabe el slot." Limita el alcance, no la ambición — convierte la curiosidad en activo de investigación programado.
El momento unipolar tiene autor y fecha: Krauthammer, Foreign Affairs, 1990/91 — con la advertencia de que sería un momento, no un estado. Zakaria habla del "ascenso del resto"; Acharya prefiere "mundo multiplex": no son pocos polos reemplazando a uno, es el orden fragmentándose en sistemas regionales sin arquitecto único.
Tres modos de trabajo: transaccional (el resultado importante que debe avanzar hoy), lateral (hacer ese resultado más rápido, repetible o escalable — la palanca del developer multiplicador) y reactivo (contener imprevistos sin destruir el plan). Distribución base: 4.5h transaccional · 2h lateral · 1h reserva reactiva · 30min cierre.
Graph theory and the network measures used to make sense of a connectome's wiring.
Nodes, edges, paths, and the structural properties between them. The formal language a connectome is expressed in. In mathematics and computer science, graph theory is the study of graphs, which are mathematical structures used to model pairwise relations between objects. A graph in this context is made up of vertices which are connected by edges. A distinction is made between undirected graphs, where edges link two vertices symmetrically, and directed graphs, where edges link two vertices asymmetrically. Graphs are one of the principal objects of study in [[Discrete Mathematics]].
Degree distributions, hubs, motifs, communities, and small-world structure. The toolkit for asking what a connectome's wiring means. The Courant Institute School of Mathematics, Computing, and Data Science, previously known as the Courant Institute of Mathematical Sciences (CIMS), is a constituent school of New York University.
Biophysical neuron models, spiking networks, and biologically plausible learning — from membrane potential to neuromorphic hardware.
The all-or-nothing spike a neuron fires when its membrane crosses threshold. The discrete event that spiking networks propagate. An action potential is a series of quick changes in voltage across a cell membrane. An action potential occurs when the [[Membrane Potential]] of a specific cell rapidly rises and falls. This "depolarization" then causes adjacent locations to similarly depolarize. Action potentials occur in several types of excitable cells, which include animal cells like neurons and muscle cells, as well as some plant cells. Certain endocrine cells such as pancreatic beta cells, and certain cells of the anterior pituitary gland are also excitable cells.
Adaptive Exponential Integrate-and-Fire: adds exponential spike initiation and an adaptation current on top of LIF. The sweet spot between biological realism and scalability — the recommended starting prototype.
Bellec et al., 2020, 'A solution to the learning dilemma for recurrent networks of spiking neurons'. Introduces e-prop as a biologically plausible, online learning rule.
Decoding neural activity into control signals (and back). A downstream application of connectomics and computational neuroscience in health and rehabilitation.
Modelling nervous systems mathematically — from single-neuron membrane dynamics to network activity. The bridge from biology to code for this whole domain. Computational neuroscience is a branch of neuroscience which employs mathematics, computer science, theoretical analysis and abstractions of the brain to understand the principles that govern the development, structure, physiology and cognitive abilities of the nervous system.
Bellec et al.'s eligibility-propagation: an online, local learning rule for recurrent SNNs that approximates backprop-through-time without a global backward pass. A candidate for the open problem of biologically plausible learning. In machine learning, a [[Neural Network]] (NN) or neural net, is a computational model inspired by the structure and functions of biological neural networks.
The widening gap between model capability and its power draw — a GPU pulls ~1,500 W, a brain ~20 W. The economic and physical ceiling that makes neuromorphic and sparse approaches strategic.
Cameras (DVS) that emit asynchronous per-pixel brightness-change events instead of frames — sparse, low-latency, and a natural input for spiking networks.
Hodgkin & Huxley, 1952. The Nobel-winning biophysical model of the action potential via voltage-gated Na⁺/K⁺ channels. The foundation of quantitative neuroscience. Sir Alan Lloyd Hodgkin was an English physiologist and biophysicist who shared the 1963 Nobel Prize in Physiology or Medicine with Andrew Huxley and John Eccles.
The 1952 biophysical model: sodium and potassium channels with voltage-gated conductances generate the spike from first principles. More realistic than AdEx, and far more costly. [[Computational Neuroscience]] is a branch of neuroscience which employs mathematics, computer science, theoretical analysis and abstractions of the brain to understand the principles that govern the development, structure, physiology and cognitive abilities of the nervous system.
Intel's neuromorphic research chip: ~1M spiking neurons per chip with on-chip learning, orders of magnitude more energy-efficient than GPUs for spiking workloads. [[Neuromorphic Computing]] is a computing approach inspired by the human brain's structure and function. It uses artificial neurons to perform computations, mimicking neural systems for tasks such as perception, motor control, and multisensory integration. These systems, implemented in analog, digital, or mixed-mode VLSI, prioritize robustness, adaptability, and learning by emulating the brain’s distributed processing across small computing elements.
A two-variable spiking model that reproduces a wide range of firing patterns at a fraction of Hodgkin-Huxley's cost. A common middle ground for large-scale simulation.
The simplest spiking neuron: a leaky membrane that integrates current until it crosses threshold, then resets. Cheap enough to scale to large networks. Biological neuron models, also known as spiking neuron models, are mathematical descriptions of the conduction of electrical signals in neurons. Neurons are electrically excitable cells within the nervous system, able to fire electric signals, called action potentials, across a [[Neural Network]]. These mathematical models describe the role of the biophysical and geometrical characteristics of neurons on the conduction of electrical activity.
The voltage across a neuron's membrane, evolving in time as ionic currents charge and discharge the membrane capacitance. What makes a model 'electrical' rather than abstract. Natural computing, also called natural computation, is a terminology introduced to encompass three classes of methods: 1) those that take inspiration from nature for the development of novel problem-solving techniques; 2) those that are based on the use of computers to synthesize natural phenomena; and 3) those that employ natural materials to compute. The main fields of research that compose these three branches are artificial neural networks, evolutionary algorithms, swarm intelligence, artificial immune systems, fractal geometry, artificial life, DNA computing, and quantum computing, among others.
Neftci, Mostafa & Zenke, 2019, 'Surrogate Gradient Learning in Spiking Neural Networks'. The reference that made gradient-based SNN training practical.
Hardware that computes with spikes and co-located memory, mimicking the brain's event-driven efficiency. Where SNNs' energy advantage becomes real. Neuromorphic computing is a computing approach inspired by the human brain's structure and function. It uses artificial neurons to perform computations, mimicking neural systems for tasks such as perception, motor control, and multisensory integration. These systems, implemented in analog, digital, or mixed-mode VLSI, prioritize robustness, adaptability, and learning by emulating the brain’s distributed processing across small computing elements.
A PyTorch library for building and training spiking neural networks with surrogate gradients. The most practical on-ramp to SNNs for someone who knows PyTorch.
Activating only the neurons a task needs instead of the whole network. The principle behind the brain's ~20 W efficiency and the pitch of sparse AI chips.
A network of spiking neurons communicating with discrete events in continuous time — event-driven, sparse, and energy-efficient. The biologically plausible alternative to dense artificial nets.
Spike-Timing-Dependent Plasticity: synapses strengthen or weaken based on the relative timing of pre- and post-synaptic spikes. A local, biologically observed learning rule.
Trains SNNs by replacing the non-differentiable spike with a smooth surrogate during the backward pass. Effective, but essentially backprop in disguise — not biologically plausible.
Voltage-dependent excitatory and inhibitory conductances (g_e, g_i) with reversal potentials — a biologically faithful synapse instead of a scalar weight added to a sum.
Mapping and analysing the brain's wiring, and using real connectomes as network architectures.
Using a real connectome's connectivity as the fixed architecture (inductive bias) of a trainable network, learning only neuron/synapse parameters. The Lappalainen 2024 line — the core thesis direction. [[Connectomics]] is the production and study of connectomes, which are comprehensive maps of connections within an organism's nervous system. Study of neuronal wiring diagrams looks at how they contribute to the health and behavior of an organism. There are two very different types of connectomes; microscale and macroscale. Microscale [[Connectomics]] maps every neuron and synapse in an organism or chunk of tissue, typically using [[Electron Microscopy]] and histology. This level of detail is only possible for small animals or tiny portions of large animal brains.
Mapping the complete wiring of a nervous system — which neurons connect to which, and what those patterns mean — then analysing the resulting graph. Connectomics is the production and study of connectomes, which are comprehensive maps of connections within an organism's nervous system. Study of neuronal wiring diagrams looks at how they contribute to the health and behavior of an organism. There are two very different types of connectomes; microscale and macroscale. Microscale connectomics maps every neuron and synapse in an organism or chunk of tissue, typically using [[Electron Microscopy]] and histology. This level of detail is only possible for small animals or tiny portions of large animal brains.
Dorkenwald et al., 2024. The complete connectome of the adult Drosophila brain — ~130k neurons, ~50M synapses — released as a publicly queryable graph. A Drosophila connectome is a list of neurons in the Drosophila melanogaster nervous system, and the chemical synapses between them. The fly's central nervous system consists of the brain plus the ventral nerve cord, and both are known to differ considerably between male and female. Complete or partial connectomes exist for the female adult brain, the male and female adult nerve cords, the male and female adult central nervous system, and the female larval central nervous system.
Nanometre-resolution imaging that resolves individual synapses. The slow, data-heavy front end of a connectomics pipeline. Dark-field microscopy, also called dark-ground microscopy, describes microscopy methods, in both light and electron microscopy, which exclude the unscattered beam from the image. Consequently, the field around the specimen is generally dark.
The adult Drosophila connectome: ~130,000 neurons and ~50 million synapses at synaptic resolution, public via FlyWire/Codex with an API. Concrete data for a connectome-constrained network. A Drosophila connectome is a list of neurons in the Drosophila melanogaster nervous system, and the chemical synapses between them. The fly's central nervous system consists of the brain plus the ventral nerve cord, and both are known to differ considerably between male and female. Complete or partial connectomes exist for the female adult brain, the male and female adult nerve cords, the male and female adult central nervous system, and the female larval central nervous system.
A split design: Elixir orchestrates the connectome (routing, backpressure, ETS graph) while Python (NumPy) vectorises the AdEx neuron updates, bridged over Erlang Ports. Separation with union at the data boundary. Neuro-symbolic AI is a subfield of artificial intelligence that combines neural networks and symbolic AI approaches, such as knowledge representation and automated reasoning, to create more robust, more reliable, and more trustworthy AI. This combination allows statistical patterns to be combined with explicitly defined rules and knowledge to give AI systems the ability to better represent, reason and generalize. Thus, neuro-symbolic AI provides a reasoning infrastructure to state-of-the-art machine learning for solving a wider range of problems more effectively.
Tracing individual neurons through millions of EM images — the biggest bottleneck in reconstruction, where merge and split errors create fake or missing connections. Medical image computing (MIC) is the use of computational and mathematical methods for solving problems pertaining to medical images and their use for biomedical research and clinical care. It is an interdisciplinary field at the intersection of computer science, information engineering, electrical engineering, physics, mathematics and medicine.
Finding pre/post-synaptic contacts to build the directed edges of the connectome graph. Errors here propagate straight into the wiring diagram.
The actor model on the Erlang VM — the orchestration half of the hybrid neuro stack.
Concurrency as isolated processes that communicate only by messages — no shared memory. Maps naturally onto a connectome: nodes as actors, synapses as message channels. The actor model in computer science is a mathematical model of concurrent computation that treats an actor as the basic building block of concurrent computation. In response to a message it receives, an actor can: make local decisions, create more actors, send more messages, and determine how to respond to the next message received. Actors may modify their own private state, but can only affect each other indirectly through messaging.
Letting slow consumers signal fast producers to slow down, so queues don't grow unbounded. Essential when streaming spikes through a network in real time (GenStage/Broadway). In queueing theory, a discipline within the mathematical theory of [[Probability]], the backpressure routing algorithm is a method for directing traffic around a queueing network that achieves maximum network throughput, which is established using concepts of Lyapunov drift. Backpressure routing considers the situation where each job can visit multiple service nodes in the network. It is an extension of max-weight scheduling where each job visits only a single service node.
The Erlang virtual machine: millions of lightweight preemptively-scheduled processes with per-process heaps and supervision. Real concurrency, unlike Python's GIL-bound threads. Calcpad is a free and open source mathematical software, created by PhD Eng. Nedelcho Ganchovski in 2014. It is primarily designed to perform engineering calculations and document the results into calculation notes/reports in HTML, PDF or Microsoft Word format. Calcpad employs its own programming language with simplified syntax, suitable for engineers with minimal programming experience. The open-source GitHub project was discontinued in March 2026 and the development was continued as closed-source.
A modern, Ruby-flavoured language on the BEAM. The orchestration layer of the hybrid stack: routing, backpressure, and concurrent state while Python does the numerics. ELIXIR is an initiative that allows life science researchers across Europe to share and store their research data as part of an organised network. Its goal is to bring together data and software experts from Europe's research organisations and data centres to help coordinate the collection, quality control and storage of large amounts of biological data produced by life science experiments. ELIXIR aims to ensure that biological data is integrated into a federated system easily accessible by the scientific community.
Erlang Term Storage: in-memory tables with read_concurrency for thousands of simultaneous lookups without blocking. Where the connectome graph lives for fast concurrent reads.
Elixir's stateful server process abstraction. Each connectome region or event router is a supervised GenServer. Confidential computing is a security and privacy-enhancing computational technique focused on protecting data in use. Confidential computing can be used in conjunction with storage and network encryption, which protect data at rest and data in transit respectively. It is designed to address software, protocol, cryptographic, and basic physical and supply-chain attacks, although some critics have demonstrated architectural and side-channel attacks effective against the technology.
Papers, people, and datasets cited across the graph.
Ashish Vaswani, lead author of the transformer paper. A provenance anchor for citations rather than a topic to study. Ashish Vaswani is an Indian computer scientist and entrepreneur. He conducted research at Google Brain, co-founded Adept AI, and co-founder and chief executive officer of Essential AI. As of June 2026, Vaswani, along with Essential AI, was acqui-hired by Nvidia.
R's grammar-of-graphics — layered, composable statistical plots. ggplot2 is an open-source data visualization package for the statistical programming language R. Created by Hadley Wickham in 2005, ggplot2 is an implementation of Leland Wilkinson's Grammar of Graphics—a general scheme for data visualization which breaks up graphs into semantic components such as scales and layers. ggplot2 can serve as a replacement for the base graphics in R and contains a number of defaults for web and print display of common scales. Since 2005, ggplot2 has grown in use to become one of the most popular R packages.
The foundational Python plotting library — everything else builds on its figure/axes model. Matplotlib is a plotting library for the [[Python]] programming language and its numerical mathematics extension [[NumPy]]. It provides an object-oriented API for embedding plots into applications using general-purpose GUI toolkits like Tkinter, wxPython, Qt, or GTK. There is also a procedural "pylab" interface based on a state machine, designed to closely resemble that of MATLAB, though its use is discouraged. SciPy makes use of Matplotlib.
Statistical visualisation on top of matplotlib with a tidy, high-level API.
Drag-and-drop BI dashboards — the industry default for interactive reporting.
Wulfram Gerstner (EPFL) — author of 'Neuronal Dynamics' and senior author on e-prop. A provenance anchor for the computational-neuroscience references.
Experiment platform with CUPED variance reduction and guarded ratio metrics.
A portfolio project: red-team a public model for prompt injection and adversarial robustness, then audit it for bias — producing a model card and an audit report.
Ingestion (Kafka), batch compute (Spark), warehouse transforms (dbt), all orchestrated by Airflow.
A hybrid prototype: Elixir orchestrates a sparse connectome (ETS, routing, backpressure) while Python vectorises AdEx neuron updates, bridged over Erlang Ports.
A CNN image classifier trained on ImageNet-style data, containerised and served behind an API.
A two-year self-study plan toward a UNAM master's on connectome-constrained spiking networks: computational neuroscience, SNNs, graph/network science, and biologically plausible learning.
A local-first retrieval + generation stack: embeddings, vector search, transformer generation, and an evaluation harness.
A planned reinforcement-learning agent (Q-learning → DQN) on a gridworld, to build RL intuition before applying it to RLHF.
Simulación que contrasta la tesis del capitalismo tardío con el fin del mundo unipolar, usando el ascenso tecnológico chino como caso de estudio — síntesis que se resiste a ser una sola tesis.
Centrality = incoming edges.
A/B Testing · Activation Functions · Actor-Critic Methods · Adversarial Attacks · Agentic AI · AI Fairness 360 · AI Red Teaming · AI Security Governance · Algorithmic Auditing · Anomaly Detection · Apache Airflow · Autoencoders · Batch Processing · Bayes' Theorem · Brain-Computer Interface · CAP Theorem · Central Limit Theorem · Central Tendency · Commons Governance · Connectome-Constrained Network
Topics adjacent or central to this research but not yet nodes — the backlog for a complete curriculum.