Math

수학은 AI, Computational Biology, geometry, generative model, evaluation을 읽기 위한 공통 기반입니다. 이 섹션은 “수학을 깊게 증명하는 곳”이 아니라, 논문과 구현에서 나오는 식을 해석하기 위한 언어를 정리하는 곳입니다.

공식은 모델을 어렵게 보이게 하려는 장식이 아니라, 어떤 quantity를 예측하고 최적화하고 평가하는지 고정하는 도구입니다.

이 블로그에서 Math는 독립적인 수학 교과서가 아니라, AI 문서의 수식을 읽기 위한 최소 공통 언어입니다. 더 깊은 증명보다 “이 식이 어떤 object, distribution, optimization target을 말하는가”를 우선합니다.

핵심 기반

영역용도시작점
Foundationsdefinitions, formulas, reusable checksMath foundations
Formula readingpaper equations, shapes, distributions, objectives, estimators, metricsFormula Reading
Linear algebravectors, matrices, tensor shapes, embeddings, projections, similarity, SVDLinear algebra, Tensor shape notation
Calculus and optimizationderivatives, chain rule, Jacobians, Hessians, constraints, backpropagation mathCalculus and gradients, Chain rule, Constrained optimization
Probabilityrandom variables, distributions, estimators, Bayesian inference, uncertainty, hypothesis testsProbability and statistics
Informationmaximum likelihood, entropy, KL, cross-entropy, generative objectivesInformation and likelihood
Geometrydistance, angle, frame, transformation, group action, invariance/equivarianceGeometry and symmetry, Geometry, Symmetry group
Graphssets, neighborhoods, permutations, graph modelingDiscrete math and graphs
Dynamicsdiscrete updates, ODEs, SDEs, vector fields, generative flows, molecular dynamicsDynamical systems
Numericsfloating point, stable softmax, log-sum-exp, conditioning, precisionNumerical computing
Evaluationmetrics, confidence intervals, calibration, statistical comparisonEvaluation math

수식 체크리스트

수식이 나오면 먼저 아래를 확인합니다.

  • Object: scalar, vector, matrix, sequence, graph, coordinate, distribution 중 무엇인가?
  • Index: batch index, token index, residue index, graph node index, time index가 무엇인가?
  • Distribution: expectation이 data, model, noise, policy, test distribution 중 어디에서 잡히는가?
  • Parameter: 미분 대상이 input , parameter , coordinate , time 중 무엇인가?
  • Objective: 식이 loss, likelihood, score, reward, metric, constraint 중 무엇인가?
  • Symmetry: permutation, translation, rotation, scale 변환에서 보존돼야 할 양이 무엇인가?
  • Estimation: population quantity와 finite-sample estimate가 구분되어 있는가?
  • Depth: post, concept note, paper note 중 어느 수준까지 풀어써야 하는가?

기본 수식 읽기 순서

AI 수식은 아래 순서로 읽으면 대부분 정리됩니다.

  1. Type: scalar, vector, matrix, tensor, graph, distribution 중 무엇인가?
  2. Shape: 각 축이 batch, token, node, coordinate, feature, time 중 무엇인가?
  3. Domain: 변수의 가능한 값과 constraint가 무엇인가?
  4. Operation: linear map, normalization, expectation, derivative, sampling, aggregation 중 무엇인가?
  5. Target: loss, metric, likelihood, score, energy, reward, constraint 중 무엇인가?
  6. Estimate: population quantity인지 finite sample estimate인지 구분합니다.

Formula by Claim Type

ClaimRouteCheck
Architecture formulaLinear layer, Attentionshape, axis, projection, normalization
Learning objectiveLearning Methods, Objective taxonomylabel, mask, contrast, likelihood, denoising, reward
Geometry ruleGeometry and Symmetry, Equivarianceinvariant target인지 equivariant target인지
Evaluation metricEvaluation Math, Estimator vs metricpopulation quantity, finite estimate, selection rule

연결되는 곳

읽기 경로

  1. vector, matrix, axis, projection, distance, embedding score를 읽으려면 Linear algebra, Tensor shape notation, Vector norm and similarity에서 시작합니다.
  2. PCA, low-rank structure, stability, representation analysis에는 Eigenvalue and eigenvector와 Singular value decomposition를 봅니다.
  3. gradient, curvature, backpropagation math에는 Calculus, Matrix calculus, Jacobian and Hessian를 봅니다.
  4. 수식을 AI optimization과 연결할 때는 Empirical risk minimization, stochastic gradient, gradient descent를 봅니다.
  5. likelihood, loss, sampling, estimate에는 Random variable, Probability distribution, Expectation, Monte Carlo estimation를 봅니다.
  6. 실험 결과 해석에는 Covariance and correlation, Normal distribution, Central limit theorem, Statistical estimator, Bayesian inference, Hypothesis testing, Bias-variance tradeoff를 봅니다.
  7. generative model과 representation learning에는 Maximum likelihood와 Entropy and KL divergence를 봅니다.
  8. graph, structure, molecule, protein note에는 Geometry와 Symmetry group를 봅니다.
  9. graph neural network, molecular graph, contact map, retrieval, set function에는 Discrete math and graphs를 봅니다.
  10. diffusion, flow matching, probability-flow ODE, recurrent model, molecular dynamics, iterative refinement에는 Dynamical systems를 봅니다.
  11. exponential, probability, mixed precision, reduction, NaN이 나오면 Numerical computing을 봅니다.
  12. 보고된 결과를 해석할 때는 Confidence interval과 Statistical significance를 봅니다.

경계

  • Math는 derivative, probability, likelihood, geometry, metric, numerical stability처럼 object와 formula를 설명합니다.
  • AI는 loss, optimizer, schedule, architecture, learning method처럼 modeling과 training 선택을 설명합니다.
  • Computational Biology는 protein, ligand, pocket, sequence, structure, assay, split처럼 domain object와 workflow를 설명합니다.
  • Infra는 GPU memory, Slurm, storage, reproducibility, server operation처럼 resource behavior를 설명합니다.

AI 연결 예시

  • Linear layer는 matrix multiplication입니다: Linear layer
  • Attention은 dot product, softmax, weighted sum입니다: Attention
  • Loss는 probability model이나 decision objective를 반영합니다: Loss function
  • Diffusion/flow는 probability path와 vector field를 다룹니다: Diffusion model, Flow matching
  • Equivariance는 group action 아래에서 output이 어떻게 변해야 하는지를 말합니다: Equivariance

논문 수식을 읽을 때

새 논문의 수식은 멋있어 보이는 notation보다 어떤 양을 정의하는지 먼저 봅니다.

수식 종류먼저 확인할 것시작점
Linear operationshape, axis, projection, rank, similarityLinear algebra, Tensor shape notation
Gradient or updateobjective, parameter, gradient estimate, optimizer stateCalculus and gradients
Constraintfeasible set, Lagrangian, penalty, projection, invalid-output handlingConstrained optimization
Probability / expectationrandom variable, conditioning, sampling distributionProbability and statistics
Likelihood / entropy / KLmodeled distribution, target distribution, bound or approximationInformation and likelihood
Symmetry / coordinate ruletransformation group, invariant target, equivariant targetGeometry and symmetry
Graph or set equationnode/edge/set unit, permutation behavior, aggregationDiscrete math and graphs
ODE/SDE/flowstate, time, vector field, integration or sampling pathDynamical systems
Metric / comparisonpoint estimate, uncertainty, paired examples, selection ruleEvaluation math