Weinan Sun
@sunw37.bsky.social
Neuroscience, Artificial Intelligence, and Beyond.
Assistant professor at Cornell Dept. of Neurobiology & Behavior
Assistant professor at Cornell Dept. of Neurobiology & Behavior
thank you Dileep for the huge inspiration!
February 13, 2025 at 1:40 AM
thank you Dileep for the huge inspiration!
sounds great! hope to catch up soon :)
February 12, 2025 at 11:42 PM
sounds great! hope to catch up soon :)
Thank you, Dan! Hope you are well!
February 12, 2025 at 9:30 PM
Thank you, Dan! Hope you are well!
12/12 This work wouldn’t be possible without co-first author @johanwinn.bsky.social, mentor @nspruston.bsky.social, and co-authors Maanasa, Chongxi, Koichiro, Arco, Michalis, Rachel, @computingnature.bsky.social, Dan, and James, plus many others (see Acknowledgments), and support from @HHMIJanelia!
February 12, 2025 at 7:36 PM
12/12 This work wouldn’t be possible without co-first author @johanwinn.bsky.social, mentor @nspruston.bsky.social, and co-authors Maanasa, Chongxi, Koichiro, Arco, Michalis, Rachel, @computingnature.bsky.social, Dan, and James, plus many others (see Acknowledgments), and support from @HHMIJanelia!
11/12 In summary, our work offers a dataset on how hippocampal cognitive maps form, revealing how brains build mental models and shedding light on learning algorithms—which could guide biologically inspired AI that reason using naturally formed internal world models.
February 12, 2025 at 7:36 PM
11/12 In summary, our work offers a dataset on how hippocampal cognitive maps form, revealing how brains build mental models and shedding light on learning algorithms—which could guide biologically inspired AI that reason using naturally formed internal world models.
10/12 Our data support prior work on multiple hippocampal representations for ambiguous inputs (e.g., Eichenbaum, @dileeplearning.bsky.social, @behrenstimb.bsky.social, @jcrwhittington.bsky.social). See our paper and preprint for more foundational references!
February 12, 2025 at 7:36 PM
10/12 Our data support prior work on multiple hippocampal representations for ambiguous inputs (e.g., Eichenbaum, @dileeplearning.bsky.social, @behrenstimb.bsky.social, @jcrwhittington.bsky.social). See our paper and preprint for more foundational references!
9/12 What do our findings reveal about hippocampal computation? We tested several models—but only Clone-Structured Causal Graph (CSCG) @dileeplearning.bsky.social matched the orthogonalized states and learning trajectory, highlighting hidden-state inference as a key learning principle.
February 12, 2025 at 7:36 PM
9/12 What do our findings reveal about hippocampal computation? We tested several models—but only Clone-Structured Causal Graph (CSCG) @dileeplearning.bsky.social matched the orthogonalized states and learning trajectory, highlighting hidden-state inference as a key learning principle.
8/12 At the single-cell level, cells in expert mice showed a continuum: some cells were “place-like,” while others became “splitters” firing selectively for Near vs. Far trials. Single-cell responses evolve across sessions, driving overall decorrelation. Explore our data: cognitivemap.janelia.org
February 12, 2025 at 7:36 PM
8/12 At the single-cell level, cells in expert mice showed a continuum: some cells were “place-like,” while others became “splitters” firing selectively for Near vs. Far trials. Single-cell responses evolve across sessions, driving overall decorrelation. Explore our data: cognitivemap.janelia.org
7/12 The final representation resembled a state machine. In ambiguous areas, CA1 activity split into distinct latent states for Near vs. Far corridors. When cues or track lengths changed, the map flexibly adapted.
February 12, 2025 at 7:36 PM
7/12 The final representation resembled a state machine. In ambiguous areas, CA1 activity split into distinct latent states for Near vs. Far corridors. When cues or track lengths changed, the map flexibly adapted.
6/12 Using UMAP on our data, the CA1 manifold evolved over learning. It started as an unstructured cluster, then formed a hub-and-spoke, and finally a ring that split into branches by trial type—refining similar inputs into distinct states that eventually capture the underlying task structure.
February 12, 2025 at 7:36 PM
6/12 Using UMAP on our data, the CA1 manifold evolved over learning. It started as an unstructured cluster, then formed a hub-and-spoke, and finally a ring that split into branches by trial type—refining similar inputs into distinct states that eventually capture the underlying task structure.