Rapid spatial cognition in mice, with and without neocortex and hippocampus

Key results

1. 01

   ShownNaive wildtype mice learned a 9-hole path maze, with a four-way choice at each hole, within about 20 rewards, improving in three phases. Part of the learning, a bias against re-crossing the hole just passed, was acquired before the first reward.

   25 mice on Mask A: reward intervals fell to ~0.11 of the starting value; traverse 22 took ~0.2 of the initial duration while running speed rose only ~1.5×. First-traverse corridor error rate was 0.26 vs 0.44 for a memoryless random walker; estimated forward bias rose from ~0.35 to ~0.6, crossing the walker's 0.5 roughly one quarter into the first journey. · 2.2 Rapid learning of a 9-turn maze in three…; 3.1 Rapid rule learning and generalization; 5.7.1 First-order Markov model with forward bias

2. 02

   SuggestedMice lacking neocortex and hippocampus (Emx1-Cre conditional Pals1 knockouts) learned the 9-hole maze, retained it across weeks and transferred experience to new masks. Their deficit centred on slow, repetitive exploration before reward discovery, and fewer reached the learning criterion.

   9 of 16 acortical mice learned Mask A; 4 learners vs 27 controls needed ~2.4× as long for four rewards, with D0 ~2.6× and comparable late duration and learning rate. Across genotypes, median traverse duration in later sessions was ~0.4 of Day 1, and stayed below Day 1 even after gaps approaching 60 days. New 9-turn masks: D0 ~0.3, E0 ~0.7 of first maze. Mask D: 5 of 13 acortical vs 8 of 8 siblings learned. · 2.5 Acortical mice rapidly learned the maze…; 2.6 Acortical mice retained maze solutions over…; 2.7 Acortical mice generalized in new mazes

3. 03

   SuggestedIn a cyclic maze of two densely interconnected subgraphs joined by a bottleneck corridor (Mask D), mice learned to favour the bottleneck at its gateway within 1-3 rewards, separately for each direction. Choices inside the subgraphs stayed largely unbiased and routes never settled.

   6 of 7 mice obtained rewards; traverses improved ~5× over the first 20. Gateway bottleneck choice (chance 0.2) rose within ~3 rewards outbound and was above chance on the first homebound journey; off-path choices stayed near 1/3 chance for most start corridors, with no trend across rewards. Graph error rates fell only to ~0.7 of initial. First reward followed the bottleneck decision by tens of seconds to minutes. · 2.4 Rapid credit assignment to a remote decision…; 3.2 Rapid learning of remote decisions; 5.6.6 Directional transitions in the Mask D maze

4. 04

   ShownWildtype mice retained Mask A overnight and brought experience to new 9-turn mazes. Transfer lowered both starting and late values of duration and turn error while learning rates stayed similar to Day 1, and an advantage from the familiar turn sequence was short-lived.

   Day 2.1 Mask A vs Day 1 (8 mice): D0 ~0.3, E0 ~0.5, Dlate ~0.4, Elate ~0.6. Days 2.2-2.4: D0 ~0.6, E0 ~0.7, Dlate ~0.7 and Elate ~0.6 of Day 1, with similar learning rates; first-traverse corridor error in a new maze ~0.6 of Day 1. On Day 2.1, Mask B (Mask A's turn sequence) vs Mask C gave D0 ~0.6, an advantage gone by Days 2.2-2.4. · 2.3 Overnight memory and generalization to new…; Figure legends

5. 05

   SuggestedIn simulations on the Mask D graph, a model-based circuit that learns a map from unrewarded exploration (Endotaxis) reproduced one-shot bottleneck preference, whereas tabular TD(0) Q-learning propagated value slowly. On Mask A, mouse errors also contradicted model-free predictions.

   Before reward both agents match the memoryless walker (bottleneck choice 0.2); after one reward Endotaxis is perfect, while Q-learning (20 agents) lags, depending on self-play routes. On one mouse's first two journeys, Endotaxis learned adjacency from sorties, then signalled the bottleneck above control corridors. In early outbound Mask A traverses, mouse turn errors were lower at holes farthest from reward. · 2.8 Candidate algorithms for few-shot spatial…; 3.2 Rapid learning of remote decisions; 5.7.2 The Endotaxis Model; 5.7.3 Model-free Reinforcement Learning