Rae, Jack William;
(2021)
Towards Lifelong Reasoning with Sparse and Compressive Memory Systems.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Jack Rae - Final Thesis.pdf - Accepted Version Download (5MB) | Preview |
Abstract
Humans have a remarkable ability to remember information over long time horizons. When reading a book, we build up a compressed representation of the past narrative, such as the characters and events that have built up the story so far. We can do this even if they are separated by thousands of words from the current text, or long stretches of time between readings. During our life, we build up and retain memories that tell us where we live, what we have experienced, and who we are. Adding memory to artificial neural networks has been transformative in machine learning, allowing models to extract structure from temporal data, and more accurately model the future. However the capacity for long-range reasoning in current memory-augmented neural networks is considerably limited, in comparison to humans, despite the access to powerful modern computers. This thesis explores two prominent approaches towards scaling artificial memories to lifelong capacity: sparse access and compressive memory structures. With sparse access, the inspection, retrieval, and updating of only a very small subset of pertinent memory is considered. It is found that sparse memory access is beneficial for learning, allowing for improved data-efficiency and improved generalisation. From a computational perspective - sparsity allows scaling to memories with millions of entities on a simple CPU-based machine. It is shown that memory systems that compress the past to a smaller set of representations reduce redundancy and can speed up the learning of rare classes and improve upon classical data-structures in database systems. Compressive memory architectures are also devised for sequence prediction tasks and are observed to significantly increase the state-of-the-art in modelling natural language.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Towards Lifelong Reasoning with Sparse and Compressive Memory Systems |
| Event: | UCL |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2021. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science > CoMPLEX: Mat&Phys in Life Sci and Exp Bio |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10126220 |
Archive Staff Only
 |
View Item |
