Skip to main content

> ML_ALGORITHMS_ATLAS_v1.0

Algorithms & Method Families

100 qualified algorithmic method families across 9 disciplines: Classical Supervised, Unsupervised, Time Series, Anomaly Detection, Deep Learning, Recommenders, Causal Inference, Reinforcement Learning, and Specialized Methods.

Accelerator Matrix14D Decision Engine
Showing 18 of 100 Method Families (Page 2 of 6)Mathematically Formulated
Algorithm
O(k * p^2)classical-unsupervised

Gaussian Mixture Models (GMM / Expectation-Maximization)

Outputs probabilistic soft cluster memberships (posteriors) rather than hard assignments.

Implementing Tools:
Scikit learnCaret rstatsmodels
clusteringdensity estimationanomaly detection
#gaussian-mixture-modelsDetails
Algorithm
O(trees * log(subsample))classical-unsupervised

Isolation Forest (iForest)

Anomaly score is a direct monotonic function of average tree path depth.

Implementing Tools:
Scikit learnPyodLinfa
anomaly detectionoutlier rejection
Paper (2008)Details
Algorithm
O(n) for novelty modeclassical-unsupervised

Local Outlier Factor (LOF)

LOF ratio directly indicates the factor by which the point is less dense than its local neighborhood.

Implementing Tools:
Scikit learnPyod
anomaly detection
#local-outlier-factorDetails
Algorithm
O(support_vectors * p)classical-unsupervised

One-Class Support Vector Machine (OC-SVM)

Decisions are driven by high-dimensional kernel distances from an origin hyperplane.

Implementing Tools:
Scikit learnPyodLinfa
anomaly detectionnovelty detection
#one-class-svmDetails
Algorithm
O(p * components)classical-unsupervised

Independent Component Analysis (FastICA)

Yields an unmixing matrix that recovers physically meaningful underlying independent source signals.

Implementing Tools:
Scikit learnSciPy
feature extractionsignal processing
#independent-component-analysisDetails
Algorithm
O(p * k)classical-unsupervised

Non-Negative Matrix Factorization (NMF)

Components are non-negative and directly interpretable as additive building blocks of the data.

Implementing Tools:
Scikit learnCaret rSciPy
feature extractiontopic modelingdimensionality reduction
#non-negative-matrix-factorizationDetails
Algorithm
O(length^2 * d)self-supervised

Masked Language Modeling (MLM / BERT)

Attention maps can be probed post-hoc; internal multi-head representations are highly distributed.

Implementing Tools:
TransformersPyTorchKerashub
text classificationtoken classificationfeature extraction
Paper (2018)Details
Algorithm
O(forward_pass)self-supervised

Simple Framework for Contrastive Learning (SimCLR)

Learns a metric embedding space where cosine distance reflects semantic invariant similarity.

Implementing Tools:
PyTorchTorchvisionKerashub
feature extractionimage classification
Paper (2020)Details
Algorithm
O(forward_pass)self-supervised

Momentum Contrast (MoCo v1 / v2 / v3)

Learns a contrastive metric space uncoupled from mini-batch GPU memory constraints.

Implementing Tools:
PyTorchTorchvision
feature extractionimage classificationobject detection
Paper (2020)Details
Algorithm
O(forward_pass)self-supervised

Bootstrap Your Own Latent (BYOL)

Generates robust visual representations without relying on negative samples or large contrastive batches.

Implementing Tools:
PyTorchTorchvision
feature extractionimage classification
#bootstrap-your-own-latent-byolDetails
Algorithm
O(ViT_forward)self-supervised

Self-Distillation with No Labels (DINO & DINOv2)

ViT self-attention heads naturally segment objects and scene semantics without explicit pixel-level supervision.

Implementing Tools:
PyTorchTorchvisionTransformers
feature extractionimage segmentationdepth estimation
Paper (2021)Details
Algorithm
O(ViT_encoder)self-supervised

Masked Autoencoders (MAE)

Reconstruction outputs can be inspected directly to verify semantic visual understanding.

Implementing Tools:
PyTorchTorchvisionTransformers
feature extractionimage classificationobject detection
Paper (2022)Details
Algorithm
O(text_enc + vision_enc)self-supervised

Contrastive Language-Image Pre-training (CLIP)

Enables zero-shot classification via natural language prompt cosine similarity without specialized training heads.

Implementing Tools:
TransformersOpen clipPyTorch
zero shot classificationimage text retrievalfeature extraction
Paper (2021)Details
Algorithm
O(1) dictionary vector lookupself-supervised

Word2Vec (CBOW & Skip-Gram)

Embedding vector geometry satisfies linear semantic analogies (e.g., King - Man + Woman = Queen).

Implementing Tools:
GensimfastText
feature extractiontoken embedding
#continuous-bag-of-words-skipgram-word2vecDetails
Algorithm
O(1) vector lookupself-supervised

Global Vectors for Word Representation (GloVe)

Vector dot products directly approximate logarithms of empirical word co-occurrence frequencies.

Implementing Tools:
GensimspaCy
feature extractiontoken embedding
#global-vectors-gloveDetails
Algorithm
O(kv_cache * layers) memory, O(params * 2 FLOPs) per tokenself-supervised

Causal Autoregressive Next-Token Prediction (GPT)

High capacity enables complex emergent chain-of-thought, but internal mechanistic interpretability remains an active research frontier.

Implementing Tools:
vLLMTransformersSGLangTensorRT-LLM
text generationcode generationin context learning
Paper (2018)Details
Algorithm
O(|A|) argmax table lookupreinforcement-learning

Tabular Q-Learning

Q-table values directly express expected cumulative discounted future returns per action.

Implementing Tools:
GymnasiumTorchrl
reinforcement learningdiscrete control
Paper (1992)Details
Algorithm
O(|A|) lookupreinforcement-learning

SARSA (State-Action-Reward-State-Action)

Learned Q-values reflect the safety penalties of the active exploration policy.

Implementing Tools:
GymnasiumTorchrl
reinforcement learningsafe exploration
#state-action-reward-state-action-sarsaDetails
Showing 19–36 of 100 Method Families