Search results for “функции потерь”

A survey lecture: machine learning as formulations of optimisation problems, a model, a loss function, a regulariser. Regression and classification, continuous approximations of the threshold loss function, SVM and logistic regression, quantile, robust and SVM regression; the L0, L1, L2, Elastic Net, SFM, RFM regularisers and feature selection; learning to rank: ordinal regression, the pairwise approach and AUC, LambdaRank; estimating a density and a mixture of distributions, clustering as a special case of a mixture, the one-class SVM; semi-supervised learning, the transductive SVM, PU-learning; matrix factorisations, autoencoders, multidimensional scaling, graph factorisations, GraphEDM; transfer learning, multi-task and self-supervised learning, distillation, learning using privileged information, generative adversarial networks; a typology of machine learning problems.

Objects, features and answers; the types of supervised learning problems; parametric models and the learning method; loss functions and empirical risk minimization; the Runge example and overfitting; hold-out, leave-one-out and cross-validation; applied problems from medical diagnosis to sales forecasting; the CRISP-DM standard.

The general scheme of supervised learning for regression, binary and multiclass classification; the margin and continuous approximations of the threshold loss; the linear classifier as a model of a neuron; the stochastic gradient method, momentum, Nesterov's accelerated gradient and heuristics; multicollinearity and L2 regularization; the maximum likelihood principle, its link to empirical risk minimization and the probabilistic meaning of regularization.

Bias–variance analysis: the decomposition of the error into noise, bias and variance, the bias and variance of simple voting, why complex ensembles do not overfit; gradient boosting for an arbitrary loss function as a parametric approximation of a gradient step; stochastic gradient boosting; special cases: regression, AdaBoost, AnyBoost, variants of loss functions; XGBoost: the tree as a linear model, L0 and L2 regularisation, analytical leaf values and the splitting criterion; CatBoost: leakage in the gradients, ordered boosting on subsamples of doubling length, unbiased target statistics for categorical features.

Ensembles of models: base algorithms, the space of estimates, the decision rule and correcting functions; the problem of diversity of base algorithms; bagging, the random subspace method and a common pseudocode of stochastic ensembling; the unbiased out-of-bag estimate and estimating feature importance; converting simple voting into weighted voting; the random forest; boosting for binary classification with abstentions, the exponential approximation of the threshold loss function, the main theorem of AdaBoost with proof, convergence, classical AdaBoost and heuristics; why boosting does not overfit: experiments, bounding the probability of error through margins, the distribution of margins, comparison with bagging; the ComBoost algorithm.

Nonlinear regression: the Newton–Raphson and Gauss–Newton methods as a series of linear regressions; logistic regression and iteratively reweighted least squares (IRLS); the generalized additive model and backfitting; the probabilistic meaning of least squares; the generalized linear model (GLM), the exponential family of distributions and the link function; logistic regression as a special case of GLM; the median, quantile regression and robust loss functions.

Anomaly and novelty detection; heuristic anomaly scores: local density, isolation forest, a mixture with a background component, robust principal component analysis; outlier removal in nonparametric regression, the LOWESS algorithm; anomaly as the value of the loss function; the theory of robust learning: slowly growing functions, iterative reweighting of objects with IRS and IRLS, robust loss functions, the median, the quantile and the censored mean, robust aggregation functions and their smoothing, the IR-ERM algorithm; one-class SVM, learning from positive and unlabeled objects (PU-learning), problems with new and unknown classes.

The optimal separating hyperplane and the width of the separating strip; the hinge loss and L2 regularization; the Karush–Kuhn–Tucker conditions, the dual problem and the types of objects; kernels, rectifying spaces and the constructive building of kernels; SVM as a two-layer neural network; choosing the constant C; SVM regression; L1 regularization, Elastic Net SVM, SFM and RVM.

The neuron as a linear model; implementing Boolean functions and the exclusive-OR problem; two-layer networks as universal approximators, Cybenko's theorem; the multilayer network and the backpropagation method; the adaptive gradient methods RMSProp, AdaDelta, Adam, Nadam; gradient explosion and gradient clipping; Dropout, ReLU activation functions, batch normalization, initialization and network pruning (Optimal Brain Damage).

A closing review of the first semester: empirical risk minimisation and gradient methods; underfitting, overfitting, multicollinearity and regularisation; optimisation formulations of supervised learning (regression, classification, ranking) and unsupervised learning (density, mixtures, clustering, one-class classification, semi-supervised learning, PU-learning, principal component analysis, autoencoders); the six scientific schools of machine learning — symbolism, evolutionism, analogism, Bayesianism, connectionism, compositionism — and their Soviet roots; estimate-calculating algorithms; practice: CRISP-DM, data preprocessing, model evaluation and selection, loss analysis, A/B testing, meta-learning and AutoML.

Sequence-to-sequence problems: the recurrent network, the encoder–decoder architecture and its limitations; the recurrent network with an attention model, interpreting the attention matrix in machine translation and image captioning; similarity functions, the query–key–value attention model, the attention formula, self-attention; multi-head attention and its regularisation, hierarchical attention for document classification, the graph attention model GAT; the transformer for machine translation: the encoder, positional encoding, layer normalisation and skip connections, the decoder with masking, the training criterion, BLEU and WER; the transformer encoder BERT, the MLM and NSP criteria, fine-tuning, BPE and WordPiece tokenisation, model sizes.

The problem of estimating generalization ability; the error matrix and the probability of overfitting; the case of a single algorithm and the hypergeometric distribution; the principle of uniform convergence of frequencies and Boole's inequality; the Vapnik–Chervonenkis bound, the growth function and the capacity of a family of algorithms; inverting the bound and the complexity penalty; structural risk minimization; the splitting and similarity effects; the Occam's razor bound and its link to regularization.

The learning-to-rank problem: recovering a partial order relation, the pointwise, pairwise and listwise approaches, the linear ranking model, applications; ordinal regression and the ordinal SVM (SVOR); the pairwise criterion, Ranking SVM, RankNet, the transition to the listwise method LambdaRank; ranking search results: query–document objects, assessor ratings, text, link and click features, TF-IDF and its probabilistic meaning, Okapi BM25, PageRank; quality criteria: precision and recall, MAP, the share of defective pairs and its link with AUC, DCG and NDCG, the pFound user behaviour model; the DSSM neural model: a siamese network, word hashing by letter trigrams, training on clicks with negative sampling.

The probabilistic formulation of classification: prior and posterior class probabilities, class likelihood functions, the maximum a posteriori principle; the optimal Bayes classifier and losses from errors; empirical estimation of prior probabilities and class densities; the naive Bayes classifier, features with exponential-family densities and the linear naive Bayes classifier; a survey of Bayesian classifiers: the Parzen window method, the quadratic discriminant, Fisher's linear discriminant, Gaussian mixtures with diagonal covariance matrices, and the radial basis function network.

Reinforcement learning as learning from the environment's response: the multi-armed bandit problem, estimating the value of an action, the greedy and ε-greedy strategies, the exploration–exploitation trade-off, the UCB method, the exponential moving average, the pursuit method and reinforcement comparison; an environment with states, the Markov decision process, how it differs from ordinary machine learning problems, the discounted return, the state and action value functions, the Bellman equation, the SARSA and Q-learning methods, on-policy and off-policy; deep Q-learning DQN on the example of Atari games, experience replay; gradient optimisation of the policy (policy gradient) and its connection with log-likelihood maximisation; model-based and model-free approaches, a regression model of the reward with an upper confidence bound, LinUCB, recommending news articles, evaluating a policy on historical data.

Estimating a probability density from a sample: the maximum likelihood principle, the multivariate Gaussian density and its geometric meaning, multicollinearity and regularisation of the covariance matrix; separating a mixture of distributions, the EM algorithm as a method of simple iterations, the Gaussian mixture (GMM); the nonparametric Parzen–Rosenblatt estimate, the choice of kernel and window width; Bayesian classification theory: the generative and discriminative approaches, the optimal Bayes classifier, the Parzen window method, the quadratic discriminant, Fisher's linear discriminant, the radial basis function network and its training by the EM algorithm.

Measures of classification quality: type I and type II errors, the ROC curve and AUC, gradient maximization of AUC, log-loss, precision and recall, micro- and macro-averaging, the F-measure; internal and external criteria of generalization ability; hold-out, leave-one-out, q-fold and t×q-fold cross-validation, stratification; model consistency; regularization, AIC, BIC and the Vapnik–Chervonenkis bound; multi-criteria model selection; feature selection: full search, Add, Add-Del, depth-first search, beam search (GMDH), the genetic algorithm and random search with adaptation.

Reminder: the neuron model, the multilayer network, stochastic gradient, momentum, vanishing and exploding gradients, ReLU, dropout, batch normalisation; depth matters more than width; feature generation and end-to-end learning; convolutional neural networks: convolution and pooling layers, the standard CNN scheme, AlexNet and ImageNet, residual networks ResNet and Highway Networks; convolutional networks for speech, texts and games; recurrent neural networks, training by BPTT, long short-term memory networks LSTM, the peephole variant, GRU.

Multiple linear regression and least squares, the pseudoinverse and the geometric meaning of least squares; the solution through the singular value decomposition; multicollinearity and the condition number; ridge regression, choosing the regularization parameter and the effective dimension; LASSO and non-smooth regularizers for feature selection; principal component analysis as a low-rank matrix decomposition; the spectral method of least squares.

Time series forecasting problems: the time series model, the forecasting horizon, the least squares method; econometrics as the main source of problems — trends, seasonalities, change points; examples of forecasting sales volumes and electricity prices; the linear autoregression model; a brief survey of forecasting methods; adaptive short-term forecasting methods: the exponential moving average and the smoothing parameter, the models of Holt, Theil–Wage, Winters with linear and exponential trend, adaptive autoregression; the tracking signal and a statistical test of model adequacy, the Trigg–Leach model; adaptive selection and adaptive composition of models, the LAVR method — local adaptation of weights with regularisation, the monotonicity constraint as a regulariser; experiments on the sales forecasting problem.