GATE: Adaptive Learning with Working Memory by Information Gating in Multi-lamellar Hippocampal Formation

TL;DR

The paper tackles how the hippocampal formation supports both working memory and rapid generalization. It introduces GATE, a brain-inspired model with a three-dimensional, dorsoventral multi-lamellar architecture that couples EC3 persistent activity, a re-entrant EC3–CA1–EC5–EC3 loop, and CA1 gating to form flexible cognitive maps. GATE reproduces a range of hippocampal cell types (e.g., splitter, lap, evidence, trace, and delay-active cells) and demonstrates that information can be processed from externally driven details to abstract task representations, enabling fast generalization across cue changes, environments, or tasks. The framework provides experimentally testable predictions and offers a principled approach for building brain-inspired systems capable of flexible memory and rapid adaptation.

Abstract

Hippocampal formation (HF) can rapidly adapt to varied environments and build flexible working memory (WM). To mirror the HF's mechanism on generalization and WM, we propose a model named Generalization and Associative Temporary Encoding (GATE), which deploys a 3-D multi-lamellar dorsoventral (DV) architecture, and learns to build up internally representation from externally driven information layer-wisely. In each lamella, regions of HF: EC3-CA1-EC5-EC3 forms a re-entrant loop that discriminately maintains information by EC3 persistent activity, and selectively readouts the retained information by CA1 neurons. CA3 and EC5 further provides gating function that controls these processes. After learning complex WM tasks, GATE forms neuron representations that align with experimental records, including splitter, lap, evidence, trace, delay-active cells, as well as conventional place cells. Crucially, DV architecture in GATE also captures information, range from detailed to abstract, which enables a rapid generalization ability when cue, environment or task changes, with learned representations inherited. GATE promises a viable framework for understanding the HF's flexible memory mechanisms and for progressively developing brain-inspired intelligent systems.

Figures (4)

• Figure 1: Working memory tasks, hippocampus formation structure and GATE model. (a) Semantic paradigm of working memory tasks in language understanding. To determine whether the word “Bank” refers to financial bank (indicated by cash) or river bank (indicated by boat rolling), one needs to keep the context in mind. (b-h) Task descriptions. (b) CS+- task: Two cues are presented (randomly one per trial), each indicating a specific correct choice at the track's end. (c) Near/far task: Similar to CS+ but requires actions at different locations. (d) CS1234 task: Two of four cues are actionable; the others are not. (e) Lap task: The agent resets, completes four laps, and acts at the end of the fourth lap. The environment remains unchanged across laps. (f) Evidence task: The agent identifies which of two cues occurs more frequently in a Poisson sequence. (g) Trace task: The agent acts after a fixed delay following a random cue. (h) Sequence task: The agent determines which of two repeated cues appeared earlier in a three-cue sequence. (i) Semantic HF connectivity. Hippocampus (HPC) and EC form a re-entrant loop: EC3 → CA1 → EC5 → EC3. CA3 and EC3 inputs dominate CA1 basal and apical dendrites, respectively. Adjacent lamellas connect dorsoventrally. Sensory input drives dorsal EC3, and ventral CA1 outputs actions. (j) Work flow of GATE in Bank task. EC3 processes sensory input (e.g., “Cash”) and modulates CA1 readout (e.g., activates after “Cash”, not “Boat”). CA3 gates CA1 timing; EC5 integrates CA1 signals and regulates EC3 memory states (write, retain, erase). Correct predictions (e.g., “Financial Bank”) yield rewards.
• Figure 2: Single-lamellar model learns to maintain information. (a) EC3 0/1 state transition. (b) EC3 neuron shows Markov chain property. Upper, activity of a simulated EC3 neuron in response to a pulse input at (45, 55), black indicates on state. Lower, mean neuron rate across trials (shadow=SEM), similar to grienberger_22. (c) Working flow of single-lamellar model, forming a re-entrant Loop. (d) EC3 output governed by $P_{01}$, $P_{10}$, $r_\infty$, $\tau$ basing on EC3 input. Shadowed areas highlight stages of information processing. (e) Sensory stimulus generate different EC3 activity. Upper, mean rate of EC3 subgroup. Lower, sensory stimulus in cue A trial (blue) and B trial (gray). (f) CA1 neuron model. Left, semantic CA1 neuron structure. Right, semantic input and output of CA1 neuron. EC3 drives CA1 apical tuft, CA3 drives CA1 basal dendrites, creating a CA1 Splitter cell zhao_22. (g) Training accuracy of Near/far task (shadow=SEM). (h) The model learns both conventional place cell and splitter cell. Red line indicates x=y. (i) CA1 activity by trial type. Left, firing rate, red arrows indicates representative cells shown on right, which shows similar responds to zhao_22. (j) Agent actions during Near/far task training. Bright area highlights “lick” action.
• Figure 3: Multi-lamellar model learns complex working memory tasks. (a) Semantic paradigm of multi-lamellar model. (b) Evaluation accuracy in task Lap, Sequence, and Evidence. Shadow area indicates SEM. (c) Delay-active cells (trace cells) in Trace task, sorted by max activity location, similar to masuda_20 . CS, conditional stimulus zone (black bar); US, unconditional stimulus zone (brown bar). (d) Cell representation in Lap task. Orange boxes and dashed arrows indicate representative cells on left, similar to sun_20. (e) Evidence cell in Evidence task, similar to nieh_21 (f) Cell representation transition during training, illustrating “switch” pattern similar to zheng_24. (g) Classification accuracy of neuron population from dorsal, intermediate and ventral CA1 in CS1234 task. Dashed green line indicates random accuracy. Gray shadow area, cue zone; blue shadow area, action zone. Ventral CA1 cannot distinguish stimulus identity at action zone, which is similar to biane_23. (h) Dorsal (left), intermediate (middle) and ventral (right) CA1 population representation in CS1234 task. Upper, MDS results shows cues that correspond to the same task outcome (CS1+ and CS2+; CS3- and CS4-) become closer. Lower, representative neuron activity in different cue trial. (i) Neural manifold approach to task topology during training in Near/far task. 20 trails are grouped as one session. Trials start from the black cross, and go clockwise. Note that the representation gradually “de-correlates” at the action zone (black arrow), resembling a “split-shank wedding ring”, similar to sun_23.
• Figure 4: Working memory enables generalization. (a) Semantic four generalization paradigms. PF, place field. (b) GATE learns faster and faster during several times of generalization (gen). Left, Type 1 generalization; right, Type 2 generalization. Upper row, representative loss curve (loss over 0.4 is not shown for clarity). Middle, epochs number used when classification loss achieve 0.01, otherwise stops training if epoch exceeds 300. Lower, splitness index correlation between generalization. (c) Place field does not significantly change during type 1 generalization. Left, two representative neuron activity. Right, neuron activity correlation during generalization or shuffle. (d) Cell representation tend to inherit during generalization. Left, scatter of splitness in training vs. generalization, or training vs. shuffle, red line indicates x=y. Right, place cell splitness during generalization or shuffle. (e) splitness correlation between first or second generalization in dCA1 or intermediate CA1 (iCA1), illustrating CA1 splitness differs in distinct lamella. (f) Representative loss curve in Type 3 (left) and 4 (right) generalization. In all bar-plots, n=30; two-sided Mann-Whittney U-test; *$P$<0.05, **$P$<0.01, ***$P$<0.001, n.s., not significant.