機械学習の社会実装では、予測精度が高くても、機械学習がブラックボックであるために使うことができないということがよく起きます。
このスライドでは機械学習が不得意な予測結果の根拠を示すために考案されたLIMEの論文を解説します。
Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "" Why should i trust you?" Explaining the predictions of any classifier." Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016.
The document discusses distances between data and similarity measures in data analysis. It introduces the concept of distance between data as a quantitative measure of how different two data points are, with smaller distances indicating greater similarity. Distances are useful for tasks like clustering data, detecting anomalies, data recognition, and measuring approximation errors. The most common distance measure, Euclidean distance, is explained for vectors of any dimension using the concept of norm from geometry. Caution is advised when calculating distances between data with differing scales.
園田翔氏の博士論文を解説しました。
Integral Representation Theory of Deep Neural Networks
深層学習を数学的に定式化して解釈します。
3行でいうと、
ーニューラルネットワーク—(連続化)→双対リッジレット変換
ー双対リッジレット変換=輸送写像
ー輸送写像でNeural Networkを定式化し、解釈する。
目次
ー深層ニューラルネットワークの数学的定式化
ーリッジレット変換について
ー輸送写像について
Reading Seminar (140515) Spectral Learning of L-PCFGsKeisuke OTAKI
1. The document presents a spectral learning method for latent-variable PCFGs (L-PCFGs) that uses tensor factorization.
2. It defines observable representations based on features of tree structures that can be computed from training data alone, without hidden variables.
3. The tensor parameter C of the L-PCFG can be recovered from the observable representations, allowing for spectral learning of the L-PCFG from a treebank via tensor methods.
機械学習の社会実装では、予測精度が高くても、機械学習がブラックボックであるために使うことができないということがよく起きます。
このスライドでは機械学習が不得意な予測結果の根拠を示すために考案されたLIMEの論文を解説します。
Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "" Why should i trust you?" Explaining the predictions of any classifier." Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016.
The document discusses distances between data and similarity measures in data analysis. It introduces the concept of distance between data as a quantitative measure of how different two data points are, with smaller distances indicating greater similarity. Distances are useful for tasks like clustering data, detecting anomalies, data recognition, and measuring approximation errors. The most common distance measure, Euclidean distance, is explained for vectors of any dimension using the concept of norm from geometry. Caution is advised when calculating distances between data with differing scales.
園田翔氏の博士論文を解説しました。
Integral Representation Theory of Deep Neural Networks
深層学習を数学的に定式化して解釈します。
3行でいうと、
ーニューラルネットワーク—(連続化)→双対リッジレット変換
ー双対リッジレット変換=輸送写像
ー輸送写像でNeural Networkを定式化し、解釈する。
目次
ー深層ニューラルネットワークの数学的定式化
ーリッジレット変換について
ー輸送写像について
Reading Seminar (140515) Spectral Learning of L-PCFGsKeisuke OTAKI
1. The document presents a spectral learning method for latent-variable PCFGs (L-PCFGs) that uses tensor factorization.
2. It defines observable representations based on features of tree structures that can be computed from training data alone, without hidden variables.
3. The tensor parameter C of the L-PCFG can be recovered from the observable representations, allowing for spectral learning of the L-PCFG from a treebank via tensor methods.
Tensor Decomposition and its ApplicationsKeisuke OTAKI
This document discusses tensor factorizations and decompositions and their applications in data mining. It introduces tensors as multi-dimensional arrays and covers 2nd order tensors (matrices) and 3rd order tensors. It describes how tensor decompositions like the Tucker model and CANDECOMP/PARAFAC (CP) model can be used to decompose tensors into core elements to interpret data. It also discusses singular value decomposition (SVD) as a way to decompose matrices and reduce dimensions while approximating the original matrix.
1. The document describes a general boosting procedure for combining weak learners to create a strong learner.
2. It involves initializing the model, learning weak learners, calculating error rates, adjusting the distribution of the training data, and combining weak learners.
3. It also describes the AdaBoost algorithm which implements this general boosting procedure and learns weak learners in sequence while focusing more on examples that previous learners got wrong.
This document discusses various machine learning topics including supervised learning techniques like support vector machines, decision trees, and neural networks. It also discusses unsupervised learning techniques like clustering algorithms. It provides short code examples for algorithms like quicksort in Haskell and OCaml. Finally, it introduces other concepts like probably approximately correct learning and boosting.
The document discusses wavelet transforms and related concepts like mother wavelets, scaling functions, and two-scale relationships. It covers definitions of wavelet transforms and wavelets, properties of wavelets like orthogonality, and applications of wavelet transforms such as signal analysis and image compression. Sections 2.1 through 2.11 each explore an aspect of wavelet transforms and wavelets.
The document discusses linear regression and regularization. It introduces linear regression models using basis functions and describes optimization methods like gradient descent. It then covers regularization, discussing how adding constraints like lasso and ridge regression can improve generalization by reducing overfitting. The document presents evidence selection methods like Bayesian model averaging to integrate over models. Finally, it describes applying empirical Bayes methods to estimate hyperparameters like α and β by maximizing the marginal likelihood.
The document discusses greedy algorithms. It defines greedy algorithms as choosing the locally optimal choice at each step in the hope of finding a global optimum. The document outlines the steps of greedy algorithms as having optimal substructure, making greedy choices at each step, and being iterative or recursive. It provides examples of activity selection problems and the 0-1 knapsack problem to illustrate greedy algorithms.