Janis Keck
LightNews LightNews
keckjanis.bsky.social
Janis Keck
@keckjanis.bsky.social
64 followers 150 following 27 posts
phd student in computational neuroscience, interested in geometric principles of biological & artificial learning || looking for postdoc positions
Posts Replies Media Videos
keckjanis.bsky.social
Symmetric learning rules might also explain why place cells in CA3 show less backward shift
(an indicator for the direction of prediction in the representation) than those in CA1, an empirical effect which we may reproduce in our model. 6/n
These plots shows the distribution of shifts of the centre of mass of place fields CA3 and CA1 in a computational model, using only asymmetric learning rules (top) or using a symmeric rule for CA3 and a symmetric rule for CA1 (bottom). One observes that in the latter case, CA3 place fields shift less than their CA1 counterparts, while when only using asymmetric learning rules, both shift equally. Also, exemplar place cells from the model and their shift are plotted.
June 25, 2025 at 1:19 PM
keckjanis.bsky.social
This suggests symmetric learning rules provide a powerful inductive bias for problems with inherent symmetry, like navigation in metric spaces. Interestingly, in an asymmetric environment (directed graph), the classic SR performs better. 5/n
This plot shows navigation performance in a reinforcement learning experiment on a directed graph. Number of steps (indicating performance) is shown on the y-axis, and episode on the x-axis. There is a vertical bar, indicating where a new target is introduced, and a horizontal bar indicating optimal performance. There are two learning curves, one for symmetric and one for asymmetric agents. Before introduction of the new target, both agents approach the optimal performance. After the new target is introduced, the asymmetric agents perform better.
June 25, 2025 at 1:19 PM
keckjanis.bsky.social
Somewhat surprisingly, these symmetrized successor representations may still be used to solve navigation tasks.
Crucially, when we test these SRs on novel navigation targets without relearning, the symmetrized version leads to better generalization than the classical one.

4/n
This plot shows navigation performance in a reinforcement learning experiment in grid worlds. Number of steps (indicating performance) is shown on the y-axis, and episode on the x-axis. There is a vertical bar, indicating where a new target is introduced, and a horizontal bar indicating optimal performance. There are two learning curves, one for symmetric and one for asymmetric agents. Before introduction of the new target, both agents approach the optimal performance. After the new target is introduced, the symmetric agents perform better.
June 25, 2025 at 1:19 PM
keckjanis.bsky.social
As one might expect, such symmetric learning rules will lead to successor representations which are also symmetrized. This means they reflect a transition structure different from the one the agent is actually experiencing. 3/n
This figure shows a cartoonish depiction of a rat running on a circular track, with the state on the track depicted by angle. There are additional plots showing exemplar sequences of states in three scenarios: an unbiased random walk (top row), a clockwise biased walk (middle row) and a counterclockwise biased wolk (bottom row). Additional plots show the successor representations that are learned in these scenarios with the symmetric and the asymmetric rule. Importantly, the symmetric SR looks identical in all three scenarios, while the asymmetric one picks up the directional bias.
June 25, 2025 at 1:19 PM
keckjanis.bsky.social
We started with a simple question: What happens to successor representations if we learn them with a learning rule that is symmetric in time, that is, invariant to the order of inputs? This question is particularly relevant for hippocampal subregions like CA3, where such rules might be at play. 2/n
Cartoonish plots are shown, illustrating the timecourse of the change in a synaptic weight and the activity of a pre- and a postsynaptic neuron. There are two panels, one for asymmetric and one for symmetric learning rule. In each panel, the activities of the neurons are once shown in forward and once in backward order. One can see that in the case of the symmetric learning rule, changing the order of activations simply changes the order of synaptic changes accordingly, while in the asymmetric learning rule no such simple relation holds.
June 25, 2025 at 1:19 PM