Chap5.6-MixDensityNetworks

Chap5.6-MixDensityNetworks - Machine Learning Srihari...

Info iconThis preview shows pages 1–7. Sign up to view the full content.

View Full Document Right Arrow Icon

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Machine Learning Srihari Mixture Density Networks Sargur Srihari 1 Machine Learning Srihari Mixture Density Networks • In some problems distribution can be multimodal • Particularly in inverse problems • Gaussian assumption can lead to poor results • In regression p( t|x ) i s typically assumed to be Gaussian • i.e. , p (t|x)= N (t| y (x,w) , β-1 ) 2 Machine Learning Srihari Kinematics of a robot arm • Robot arm with two links • This is a regression problem with two inputs (desired location of arm) and two outputs (angles for links) 3 Forward problem: Find end effector position given joint angles Has a unique solution Inverse kinematics has two solutions: Elbow-up and elbow-down Machine Learning Srihari Forward and Inverse Problems 4 • Forward problems correspond to causality in a physical system have a unique solution • If forward problem is a many-to-one mapping, inverse has multiple solutions • Inverse classi¡cation problem symptoms caused by disease Machine Learning Srihari Data Set for Forward and Inverse Problems 5 • Least squares corresponds to • Maximum likelihood under a Gaussian assumption • Leads to a poor result for highly non-Gaussian inverse problem • Seek a general framework for modeling conditional probability distributions • Achieved by using a mixture model p( t|x ) Forward problem data set: x is sampled uniformly over (0,1) to give values { x n } Target t n obtained by function x n +0.3sin(2 π x n ) Then add noise over (-0.1,0.1) Red curve is result of ¡tting a two-layer neural network by minimizing sum-of-squared error Corresponding inverse problem by reversing x and t Very poor ¡t to data Machine Learning Srihari A Mixture Density • Model with K components • Components can be Gaussian for...
View Full Document

This document was uploaded on 02/25/2012.

Page1 / 20

Chap5.6-MixDensityNetworks - Machine Learning Srihari...

This preview shows document pages 1 - 7. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online