Abstract

Please join us for our fortnightly Computational Neuroscience journal club on Tuesday 14th November at 2pm UK time in the CBL seminar room, or online on zoom.

The title is ‘Low rank RNNs’, presented by Youjing Yu and Yashar Ahmadian.

Zoom information: https://eng-cam.zoom.us/j/84204498431?pwd=Um1oU284b1YxWThObGw4ZU9XZitWdz09 Meeting ID: 842 0449 8431 Passcode: 684140

Summary:

RNNs are super cool – they can be trained to perform a range of tasks, they incorporate temporal information and, most important of all, they can model the brain’s processes quite well. However, it is very difficult to understand how the connectivity in RNNs give rise to the input-output mappings that they are able to perform. In this talk, we look at Ostojic Lab’s work on a special family of RNNs – low-rank RNNs. In the first part I will give an introduction to the theoretical framework of RNN , before explaining how the (sub)population statistics give rise to dynamics [1, 2]. In the second part Yashar will take over to explain how we can reverse-engineer the trained full RNN networks to produce the low-rank versions that matches the performance of the full-rank RNNs. Finally, Yashar will talk about how RNNs are able to perform context-dependent flexible tasks [3].

[1] Mastrogiuseppe, Francesca, and Srdjan Ostojic. 2018. “Linking Connectivity, Dynamics, and Computations in Low-Rank Recurrent Neural Networks” 99 (3): 609-623.e29. https://doi.org/10.1016/j.neuron.2018.07.003. [2] Beiran, Manuel, Alexis Dubreuil, Adrian Valente, Francesca Mastrogiuseppe, and Srdjan Ostojic. 2021. “Shaping Dynamics with Multiple Populations in Low-Rank Recurrent Networks.” Neural Computation 33 (6): 1572–1615. https://doi.org/10.1162/neco_a_01381. [3] Dubreuil, Alexis, Adrian Valente, Manuel Beirán, Francesca Mastrogiuseppe, and Srdjan Ostojic. 2022. “The Role of Population Structure in Computations through Neural Dynamics.” Nature Neuroscience 25 (6): 783–94. https://doi.org/10.1038/s41593-022-01088-4.