The geometry of neural activity: low-dimensional dynamics and high-dimensional representations


Joao Barbosa | École normale Supérieure, Paris, France
Adrian Valente | École normale Supérieure, Paris, France
Yuxiu Shao | École normale Supérieure, Paris, France
Ljubica Cimesa | École normale Supérieure, Paris, France


The dimensionality of neural representations is at the center of an intense debate. Theoretically, a low-dimensional representation buys you robustness and generalizability. On the other hand, high-dimensional representations are more flexible, because they are more easily decoded by downstream areas. This theoretical tension is reflected in paradoxical findings in neuroscience. With reconciliatory potential, some have argued that the widespread low-dimensional representations in neurophysiological recordings could simply reflect the dimensionality of the task the animals are engaged with.

This workshop will bring together evidence from both views, with the hope of clarifying their interaction in particular in three broad fields:
1. Decision Making
2. Learning in Artificial Neural Networks
3. Motor Learning

With the help of the audience and the speakers, we would like to address the following general questions:
1. Does the brain use both high and low dimensional representations, and if so, for what purposes?
2. Can the paradoxical empirical findings be reconciled? For example, seeing neural representations as low-dimensional with respect to the neural space (~10^10) but high dimensional with respect to the task variables (<10)?
3. How does dimensionality change during learning, both in biological and artificial networks?

Schedule (CEST)




Decision Making


Carsen Stringer | HHMI Janelia Research Campus, Ashburn, USA
Rastermap: Extracting structure from high-dimensional neural data


Mattia Rigotti | IBM Research AI, USA
Connectivity constraints on the dimensionality of neural representations


Jonathan Pillow | Princeton University, USA
Characterizing the multi-dimensional dynamics of decision-making in PFC


Wrap up & 30 min break


Learning in Artificial Networks


Friedrich Schuessler | Technion, Haifa, Israel
Rich and lazy learning in RNNs for neuroscience?


SueYeon Chung | Columbia University, New York, USA
Perceptual manifolds in biological and artificial neural networks


Wrap up & 60 min break


Motor Learning


Sara Solla | Northwestern University, Evanston, USA
Low Dimensional Manifolds for Neural Population Dynamics


Alex Cayco-Gajic | École Normale Supérieure, Paris, France
High-dimensional structure of cerebellar representations


Abigail Russo | Princeton University, USA
Quantification of population structure for the study of neural dynamics in motor cortices


Discussion Panel