[ContinualAI Seminars]: "Continual learning of sequences through Generative Replay: A Hippocampus modeling story"
7 views
Skip to first unread message
Keiland Cooper
Oct 26, 2022, 6:10:15 AM10/26/22
to Continual Learning & AI News
Hi folks,
Join us this Thursday 10-27-2022, 15:30 PM UTC, for the ContinualAI Seminar, where Ivilin Peev Stoianov (National Research Council of Italy, Institute of Cognitive Sciences and Technologies) will present the paper:
Title: “Continual learning of sequences through Generative Replay: A Hippocampus modeling story”
Link: https://www.sciencedirect.com/science/article/pii/S0301008222001150
Abstract: Experience replay mitigates catastrophic forgetting in continual learning by reactivating previous observations. However, memorizing unbounded amount of experiences could be costly, especially for biological systems, and there is no evidence for a verbatim memory holding such traces in living organisms Nevertheless, there is evidence for experience replays in the hippocampal formation during rest, e.g., in rodents, in the context of spatial navigation. We investigated this issue computationally, with a novel hierarchical generative model that leverages on clustering through mixtures to organize and replay experienced traces without verbatim memory, thus overcoming the issue of catastrophic forgetting and supporting continual learning with a biologically plausible method. The model assumes a minimal three-level structure operating on a common representational space. The lowest level encodes individual items (localized activation in the space); sequences of items, or traces are computed through leaky integration at the second layer; finally, sets of compatible sequences – e.g., belonging to the same context - are clustered in "maps" at the third layer. The generative structure natively supports the stochastic activation of sequences compatible with the previous experiences, also called "generative replay". We show that the model learns and efficiently retains multiple spatial navigation trajectories, by organizing them into spatial maps. Furthermore, the model reproduces flexible and prospective aspects of hippocampal dynamics that are challenging to explain within existing frameworks. The mode thus reconciles multiple roles of the hippocampal formation in map-based navigation, episodic memory and imagination. More broadly, the model paves the way for continual learning of general spaces through hierarchical mixtures.
---------------
- YouTube link:
https://www.youtube.com/watch?v=4thByqjQU-k
- Microsoft Teams:
https://bit.ly/clai-seminars
- YouTube recordings of the previous sessions:
https://www.youtube.com/c/ContinualAI
---------------
Feel free to share this email to anyone interested and invite them to subscribe this mailing-list here: https://groups.google.com/g/continualai
Please also contact me if you want to speak at one of the next sessions!
Looking forward to seeing you all there!
All the best,
Keiland Cooper
University of California
ContinualAI Co-founder
http://kwcooper.xyz
Join us this Thursday 10-27-2022, 15:30 PM UTC, for the ContinualAI Seminar, where Ivilin Peev Stoianov (National Research Council of Italy, Institute of Cognitive Sciences and Technologies) will present the paper:
Title: “Continual learning of sequences through Generative Replay: A Hippocampus modeling story”
Link: https://www.sciencedirect.com/science/article/pii/S0301008222001150
Abstract: Experience replay mitigates catastrophic forgetting in continual learning by reactivating previous observations. However, memorizing unbounded amount of experiences could be costly, especially for biological systems, and there is no evidence for a verbatim memory holding such traces in living organisms Nevertheless, there is evidence for experience replays in the hippocampal formation during rest, e.g., in rodents, in the context of spatial navigation. We investigated this issue computationally, with a novel hierarchical generative model that leverages on clustering through mixtures to organize and replay experienced traces without verbatim memory, thus overcoming the issue of catastrophic forgetting and supporting continual learning with a biologically plausible method. The model assumes a minimal three-level structure operating on a common representational space. The lowest level encodes individual items (localized activation in the space); sequences of items, or traces are computed through leaky integration at the second layer; finally, sets of compatible sequences – e.g., belonging to the same context - are clustered in "maps" at the third layer. The generative structure natively supports the stochastic activation of sequences compatible with the previous experiences, also called "generative replay". We show that the model learns and efficiently retains multiple spatial navigation trajectories, by organizing them into spatial maps. Furthermore, the model reproduces flexible and prospective aspects of hippocampal dynamics that are challenging to explain within existing frameworks. The mode thus reconciles multiple roles of the hippocampal formation in map-based navigation, episodic memory and imagination. More broadly, the model paves the way for continual learning of general spaces through hierarchical mixtures.
---------------
- YouTube link:
https://www.youtube.com/watch?v=4thByqjQU-k
- Microsoft Teams:
https://bit.ly/clai-seminars
- YouTube recordings of the previous sessions:
https://www.youtube.com/c/ContinualAI
---------------
Feel free to share this email to anyone interested and invite them to subscribe this mailing-list here: https://groups.google.com/g/continualai
Please also contact me if you want to speak at one of the next sessions!
Looking forward to seeing you all there!
All the best,
Keiland Cooper
University of California
ContinualAI Co-founder
http://kwcooper.xyz
Reply all
Reply to author
Forward
0 new messages