RT @HHMIJanelia: 💻 A new class of simplified ‘minimodels’ developed by @Du_Fengtong @NunezKant @marius10p @computingnature can predict the….

RT @HHMIJanelia: A new tissue expansion method developed by the Wang (@mengwang939) & Tillberg (@TillbergPaul) labs & the @LingjunLi2 lab @….

I love this new technique developed my colleague James Liu and his lab at Janelia.

I’m excited about this highly collaborative project by @sunw37, @JohanWinn, and many other talented contributors. Read more about how our @HHMIJanelia team imaged thousands of neurons in mouse hippocampus as mice learned cognitive tasks over several days:.

11/11 In addition to the co-first authors @sunw37 and @JohanWinn, many others made critical contributions. Big thanks to co-authors Maanasa, @Chongxi_lai, Koichiro, Michalis, Rachel, and James, as well as many others recognized in the Acknowledgments!.

10/11 We hope this work provides useful data for gaining insight into the formation of cognitive maps and inspires novel AI learning algorithms, for example, in unsupervised learning and model-based RL. #NeuroAI.

9/11 Our work provides new data supporting previous work on multiple hippocampal representations for potentially ambiguous observations -- e.g., Eichenbaum, @dileeplearning, @behrenstimb, @jcrwhittington. See the preprint for many other references to foundational work!.

8/11 What does this teach us about hippocampal computation and learning? Both an HMM-based CSCG model @dileeplearning and RNNs with Hebbian learning, recapitulated key features of the “Orthogonalized State Machine (OSM)”, while LSTMs and Transformers did not natively do so.

7/11 TLDR: We found that the learned hippocampal task representations resembled a state machine, generated by progressive decorrelation of task-relevant states. Introducing variations of the task showed flexible usage of this learned state machine.

6/11 By using Nonlinear dimensionality reduction (UMAP) as an intuitive visualization tool we found distinct manifold topology stages for the two trial types, which gradually evolved during learning (more extensive analysis in the paper).

5/11 Place tuning of individual cells between trial types showed progressive decorrelation, resulting in eventual orthogonalization at key track segments.

4/11 During training, animals went through stereotypical behavioral stages: 1) licking randomly, 2) licking at the two potential reward zones, 3) stopping licking if it received a reward at the Near zone, 4) correctly using the indicator to predict the reward location (expert).

3/11 Mice learned to collect water rewards at either Near or Far reward zones in two VR tracks based on distinct indicator cues (Stripes or Dots) at the track start. Expert mice (shown below) only licked at the correct reward zones to trigger a water reward.

2/11 We investigated how activity in the hippocampus changes over days to weeks while naïve mice learned a VR task. We tracked thousands of CA1 cells with a 2p-RAM mesoscope designed by @sofroniewn, Dan Flickinger, @svoboda314. We ensured stability using hard/software solutions.

1/11 How are cognitive maps formed during learning? . Very excited to share new work by outstanding co-first authors @sunw37 and @JohanWinn and other wonderful collaborators @HHMIJanelia

(10/10) Huge thanks again to Xinyu and Ching-Lung @hiallen72 for this beautiful work. Also, congrats to Ching-Lung for starting his new lab at Academia Sinica! @SinicaFans @IbmsSinica.

(9/10) Implication: How does the hippocampus infer separate latent states, i.e. contexts? @honisanders @gershbrain Either by massive sensory changes signaling a different environment or subtle changes but with a strong behavioral contingency. Related.

(8/10) Our model for two-step learning in the hippocampus: Gradual learning creates separate CA3 codes for the two behavioral contexts. Rapid plasticity in CA1 reads out the conjunctive code (position and choice), allowing flexible routing of the code to downstream brain areas.

(7/10) In striking contrast, when the same experiment was done in mice that had not been trained to associate the initial cues with the rewarded arm of the Y-maze, inducing plasticity created conventional place cells rather than splitter cells.

(6/10) Examination of intracellular Vm showed that the postsynaptic potential had all-or-none dynamics in the delay zone between trial-types. This indicates that the trial-type-specific information had been inherited from their inputs (presumably CA3).