Neural Networks for Natural Language Processing: using Word2Vec on TrueNorth

Map word2vec to TrueNorth, working with the [ NLP group], see also 2015/word2vec

Team: Daniel Mendat, Guillaume Garreau, Andrew Cassidy, john arthur, Rodrigo Alvarez, Andreas Andreou, and [ NLP group]

Objective

  • Convert input word to vector
  • Find nearby dictionary words in vector space

Methods

  • Dot-product similarity
  • Vector-matrix multiplation

Results

  • Single chip demonstration:
    • 95,000-word dictionary
    • 64-element vectors
    • 3991 TrueNorth cores

Examples

input:telluridebasketballmountainsneuron
carbonvolleyballlandformselectron
coloradohandballrangesprotein
coppersoccerglaciertissue
springsncaaplateaucells

Analogies

  • good : better :: bad : worse
  • france : wine :: germany : beer
  • apple : iphone :: google : search

Attachments