A multilevel account of hippocampal function in spatial and concept learning: Bridging models of behavior and neural assemblies

Abstract

A complete neuroscience requires multilevel theories that address phenomena ranging from higher-level cognitive behaviors to activities within a cell. We propose an extension to the level of mechanism approach where a computational model of cognition sits in between behavior and brain: It explains the higher-level behavior and can be decomposed into lower-level component mechanisms to provide a richer understanding of the system than any level alone. Toward this end, we decomposed a cognitive model into neuron-like units using a neural flocking approach that parallels recurrent hippocampal activity. Neural flocking coordinates units that collectively form higher-level mental constructs. The decomposed model suggested how brain-scale neural populations coordinate to form assemblies encoding concept and spatial representations and why so many neurons are needed for robust performance at the cognitive level. This multilevel explanation provides a way to understand how cognition and symbol-like representations are supported by coordinated neural populations (assemblies) formed through learning.

Fig. 1.. Multilevel explanation of concept learning in the brain: decomposition of a cognitive model into neural flocks.

(A) Levels of mechanisms for neuroscience. Behavior is the phenomenon of interest, explained by a task-performing algorithm (cognitive model), which can be decomposed into lower-level mechanisms. (B) Our behavior of interest is concept learning and categorization. (C) Behavior is explained by the cognitive model. After the stimulus is encoded, attention is applied, and neuron-like units activate according to their similarity to the input. These activations are transmitted through learned association weights to generate an output (e.g., category decision). Dotted circles are abstract clusters. (D) Decomposition of the clusters into neuron-like units and the flocking learning rule. Clusters from (C) are decomposed into neuron-like units [gray circles in (C) represent units, and dashed circles highlight units in the same flock or virtual cluster]. Left to right: k winners (blue) move toward the stimulus (“S”), followed by a recurrent update, where units move toward their centroid (“C”). k neuron-like units become similarly tuned over time, forming a neural flock.

Fig. 2.. Formation of concept and spatial representations by neural flocking.

(A) The model learns distinct representations for apples and oranges. k winners (i.e., most activated units) adjust their receptive fields toward the current stimulus, followed by a recurrent update toward their centroid. (B) The second update is sufficient to solve the coordination problem allowing SUSTAIN-d to form neural flocks or virtual clusters (which, in this example, represent the concepts apple and orange). (C) Spatial representation formation. Left: An agent (e.g., a rodent) forages in an environment. Right: Development of spatial representations. SUSTAIN-d’s neuron-like units are initially uniformly tuned to locations. At each time step, the k winners move toward the stimulus (e.g., sensory information at the current location) and each other (i.e., neural flocking). This learning dynamic creates flocks or virtual clusters of units with similar spatial tuning, akin to place-cell assemblies. These flocks tile the environment. (D) Examples of grid cell-like activity patterns and corresponding spatial autocorrelograms after learning. See fig. S1A for more examples and fig. S1B for distributions of grid scores).

Fig. 3.. SUSTAIN-d’s brain-scale population of neuron-like units collectively displays the same behavior as the high-level cognitive model that it decomposes, while making additional predictions about robustness in neural computation.

(A) Six concept learning structures (29). Bottom: In each box, stimuli in the left and right columns are in different categories. Top: Cubes represent each stimulus in binary stimulus feature space (color, shape, and size) for each structure. (B) Learning curves from human behavior (65) (left) and model fits (right). Probability of error is plotted as a function of learning block for each structure. (C) Neuron-like units form neural flocks or virtual clusters (e.g., type I in blue and type VI in brown; see fig. S2 for all types) that parallel clusters in the higher-level cognitive model. The number of units are subsampled from the whole population for better visualization. (D) The more neuron-like units, the more robust the model is when confronted by failure modes (e.g., cell death, noise, and synaptic transmission failure). (E) The stronger the recurrence during learning, the better the noise tolerance. See fig. S3 for more examples.

Fig. 4.. Further decomposing SUSTAIN-d to capture known functional differences along the anterior-posterior axis of the hippocampus.

(A) Illustration of the anterior-posterior (blue-yellow) gradient in human hippocampus (HPC). Place fields in the anterior hippocampus are broader, and posterior hippocampus place fields are more narrow. Likewise, anterior hippocampus is strongly activated by concept learning structures that follow broad, general rules (left), and posterior hippocampus is more strongly engaged by irregular rule-plus-exception structures where specific instances are important. (B) SUSTAIN-d is further decomposed into a bank of units with broader tuning to model anterior hippocampus (blue) and a narrowly tuned bank of units to model posterior hippocampus (yellow). Both banks contribute to the output and compete to exert control over the category decision. (C) Model output (left) combines the anterior (middle) and posterior (right) neuron-like units’ output. The anterior units dominate for simple category structures, whereas the posterior units dominate for irregular structures.