“Analysing Iterative Machine Learning Algorithms with Information Geometric Methods”

¶ machine learning and on-line algorithms · a connection between machine learning, statistics and geometry ¸ stochastic gradient algorithm in Riemannian spaces non-Euclidean example: à multinomial distribution leading to the simplex UNE Seminar: A Geometric View of Non-Linear On-Line Learning Algorithms 4 What’s machine learning? à Example: play a prediction game • given an input xt predict an output yt • e.g., xt is a pair of temperature and humidity of day t and yt is {0, 1}, where 1 means that day t is a rainy day and 0 it is not; • e.g., xt are scanned 256 × 256-images of digits and yt are numbers 0, 1,..., 9 — pattern recognition (recognition of handwritten digits: post code, zip) Figure 1: pixel image of a handwritten digit UNE Seminar: A Geometric View of Non-Linear On-Line Learning Algorithms