Unsupervised learning in NLP tasks on using Vector-Symbolic
representations on phasor-based associative memory
The goal is to adapt algorithms of unsupervised learning using Self-Organized Maps
on symbolic streams, e.g. texts, for phasor-based computation with spikes on Loihi. The project
will be based on the recently developed in my group holistic vector-symbolic realization of SOMbased learning of n-grams, permuted, words, variable length sequences, which demonstrated
significant gains in performance, runtime and energy consumption on conventional computing
architectures. The main hypothesis to test during the project is that the vector symbolic SOM
pipeline could be implemented using the functionality of phasor-based realization of associative
memory on Loihi. Along the way of project execution, it is expected to gain important theoretical
insights on how to efficiently implement unsupervised SOM-like learning on spiking hardware.
The results of the project and the theoretical insights are expected to be significant for a wide
range of practical applications
Funder: Intel 1 300 000 SEK
Contact
Evgeny Osipov
- Professor
- 0920-491578
- evgeny.osipov@ltu.se
- Evgeny Osipov
Updated: