Using Context to Disambiguate Similar Signals Using Conceptors

Chatterji, Satchit (2022) Using Context to Disambiguate Similar Signals Using Conceptors. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelor_Thesis_S3889807.pdf Download (944kB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (130kB)

Abstract

Humans robustly and ubiquitously use symbolic context to distinguish between noisy or ambiguous percepts. However, no consensus exists on how to properly integrate top-down/symbolic and bottom-up/sub-symbolic information flow in artificial neural networks. One approach to this `bidirectional information flow' problem in recurrent neural networks (RNNs) is the conceptor -- a matrix that characterises the linear space of neuronal activations in response to a signal. These may be abstractly combined using Boolean logical connectives. However, they may alternatively be interpreted as fuzzy operators, and conceptors themselves corresponding to fuzzy sets. The current work introduces a bidirectional classification model that allows for top-down information to affect bottom-up processes in an RNN using conceptors. Building on conceptor-based classification, three biasing methods are discussed, based on (i) crisp symbols, (ii) probability-quantified uncertainty, and (iii) fuzzy membership-quantified uncertainty. As a demonstration, phonemes are transcribed using the TIMIT dataset. With the goal of using contextual information to help disambiguate similar percepts, the discriminatory power of these biasing paradigms are studied with respect to similar-sounding phonemes.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Jaeger, H. and Pourcel, G. A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	17 Aug 2022 10:07
Last Modified:	17 Aug 2022 10:07
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/28406

Actions (login required)

View Item