Brain ML

We studied a brain inspired neural approach to perform multi-modal clustering.

Date: 2018–2019

Advisor: Pawel Herman, KTH, Stockholm

The aim of this Master’s research project was to study a Deep neural network architecture inspired by some current knowldege of the human brain. The targetted downstream task was the unsupervised clustering of images and associated captions.

Abstract: The human neocortex uses efficient learning architectures and data representationsthat allow for generalization and, particularly, multi-modality, i.e. semantic con-nections between different spaces or domains (language and vision for instance).The idea developed in this paper is to try to model this behavior through an au-toassociative memory and feedforward representation learning networks inspiredfrom the way neural information is processed in the neocortex. We also aim toprove the necessity of sparsifying information to gain generality and robustness,as it seems to be the case in some of the neural information processing pathwaysin the human brain. To this end, we produce a sparse representation of an imageand its associated caption thanks to a sparse autoencoder and we send it to anattractor model to perform unsupervised learning on the fused representations. Thechallenge is to obtain basins of attractors which are wide enough to be able toperform clustering.

A report summarizing our work can be found here.