Randall C. O'Reilly

A major outstanding question in neuroscience is whether the neocortex uses the same powerful learning algorithm as current AI models: error backpropagation. One way this could be accomplished is as a function of the temporal derivative (i.e., differences in neural activity states over time), which can closely approximate the backpropagated error gradient. We tested the hypothesis that the direction of synaptic plasticity is a function of the temporal derivative in synaptic activity over the course of a 200 ms (5 Hz) theta cycle. Using mouse hippocampal slices, we drove presynaptic activity across the two 100 ms halves of a 200 ms window at either 25 Hz or 50 Hz, combined with corresponding low and high magnitudes of postsynaptic depolarization, testing all four 2×2 combinations of these low and high activity levels, while measuring the resulting effects on synaptic efficacy (as measured by EPSP amplitude to standard test probes). Consistent with the computational hypothesis, a positive temporal derivative (low to high) resulted in LTP (increased synaptic strength), while a negative temporal derivative (high to low) resulted in LTD. Critically, both no-change conditions (stable low or high across 200 ms) resulted in no net synaptic change, even though the high no-change condition had the highest overall synaptic activity levels. Possible biochemical mechanisms that could support these results are discussed.

The entorhinal cortex (EC) provides the major input to the hippocampus (HPC). Numerous computational models on the EC propose that its grid cells serve as a spatial metric, supporting path integration and efficient generalization. However, little is known about how these cells could contribute to episodic memory, which emphasizes episode-specific representations that align with pattern separation. Taking into consideration anatomical specifications of EC inputs to the HPC and computational principles underlying the EC-HPC memory system, we argue that EC layer IIa (EC2a) supports pattern separation and EC layer IIb/III (EC2b/3) supports generalization. Utilizing recurrent inhibition and the nature of single EC2a neurons binding converging inputs from the neocortex (i.e., conjunctive coding), we built a biologically-based neural network model of the EC-HPC system for episodic memory. By examining how EC2a transformed its cortical inputs and output them to the trisynaptic pathway (EC2a - Dentate Gyrus - CA3 - CA1), we found that instead of systematically generalizing across similar inputs, recurrent inhibition and conjunctive coding in EC2a neurons support strong pattern separation and increase mnemonic discrimination. Furthermore, lesioning EC2a neurons in the model resembled memory impairments found in people with Alzheimer’s Disease, suggesting an intertwined relationship between memory and the majority of pure grid cells in the EC. On the other hand, the topographically organized monosynaptic pathway (EC2b/3 - CA1) is computationally more suitable for efficient factorization and generalization. This model provides novel anatomically-based predictions regarding the computational roles of EC cells in pattern separation and generalization, which together form a critical computational framework for both episodic memory and spatial navigation.

Recent work has shown that the brain abstracts non-spatial relationships between entities or task states into representations called cognitive maps. Here, we investigated how cognitive control enables flexible top-down selection of goal-relevant information from multidimensional cognitive maps retrieved from memory. We examined the relationship between cognitive control and representational geometry by conducting parallel analyses of fMRI data and recurrent neural network (RNN) models trained to perform the same task. We found both stable map-like representations in a medial temporal lobe and orbitofrontal cortical network that reflected both task-relevant and irrelevant dimensions and dynamic, orthogonal representations of only relevant task dimensions in a frontoparietal network. These representational motifs also emerged with distinct temporal profiles over the course of training in the RNN, with map-like representations appearing first. We further show that increasing control demands due to incongruence (conflicting responses) between current task-relevant and irrelevant dimensions impact the geometry of subjective representations, and the degree of this effect further accounts for individual differences in cognitive control. Taken together, our findings show how complementary representational geometries balance stability and behavioral flexibility, and reveal an intricate bidirectional relationship between cognitive control and cognitive map geometry.

While many theories assume that sleep is critical in stabilizing and strengthening memories, our recent behavioral study (Liu & Ranganath, 2021) suggests that sleep does not simply stabilize memories. Instead, it plays a more complex role, integrating information across two temporally distinct learning episodes. In the current study, we simulated the results of Liu and Ranganath (2021) using our biologically plausible computational model, TEACH, developed based on the complementary learning systems (CLS) framework. Our model suggests that when memories are activated during sleep, the reduced influence of temporal context establishes connections across temporally-separated events through mutual training between the hippocampus and neocortex. In addition to providing a compelling mechanistic explanation for the selective effect of sleep, this model offers new examples of the diverse ways in which the cortex and hippocampus can interact during learning.

Neural networks struggle in continual learning settings from catastrophic forgetting: when trials are blocked, new learning can overwrite the learning from previous blocks. Humans learn effectively in these settings, in some cases even showing an advantage of blocking, suggesting the brain contains mechanisms to overcome this problem. Here, we build on previous work and show that neural networks equipped with a mechanism for cognitive control do not exhibit catastrophic forgetting when trials are blocked. We further show an advantage of blocking over interleaving when there is a bias for active maintenance in the control signal, implying a tradeoff between maintenance and the strength of control. Analyses of map-like representations learned by the networks provided additional insights into these mechanisms. Our work highlights the potential of cognitive control to aid continual learning in neural networks, and offers an explanation for the advantage of blocking that has been observed in humans.

The hippocampus plays a critical role in the rapid learning of new episodic memories. Many computational models propose that the hippocampus is an autoassociator that relies on Hebbian learning (i.e., “cells that fire together, wire together”). However, Hebbian learning is computationally suboptimal as it does not learn in a way that is driven toward, and limited by, the objective of achieving effective retrieval. Thus, Hebbian learning results in more interference and a lower overall capacity. Our previous computational models have utilized a powerful, biologically plausible form of error-driven learning in hippocampal CA1 and entorhinal cortex (EC) (functioning as a sparse autoencoder) by contrasting local activity states at different phases in the theta cycle. Based on specific neural data and a recent abstract computational model, we propose a new model called Theremin (Total Hippocampal ERror MINimization) that extends error-driven learning to area CA3—the mnemonic heart of the hippocampal system. In the model, CA3 responds to the EC monosynaptic input prior to the EC disynaptic input through dentate gyrus (DG), giving rise to a temporal difference between these two activation states, which drives error-driven learning in the EC→CA3 and CA3↔CA3 projections. In effect, DG serves as a teacher to CA3, correcting its patterns into more pattern-separated ones, thereby reducing interference. Results showed that Theremin, compared with our original Hebbian-based model, has significantly increased capacity and learning speed. The model makes several novel predictions that can be tested in future studies.

A hallmark of human intelligence is the ability to adapt to new situations by applying learned rules to new content (systematicity) and thereby enabling an open-ended number of inferences and actions (generativity). Here, we propose that the human brain accomplishes these feats through pathways in the parietal cortex that encode the abstract structure of space, events, and tasks and pathways in the temporal cortex that encode information about specific people, places, and things (content). Recent neural network models show how the separation of structure and content might emerge through a combination of architectural biases and learning, and these networks show dramatic improvements over previous models in the ability to capture systematic, generative behavior. We close by considering how the hippocampal formation may form integrative memories that enable rapid learning of new structure and content representations.

The hippocampus plays a critical role in the rapid learning of new episodic memories. Many computational models propose that the hippocampus is an autoassociator that relies on Hebbian learning (i.e., “cells that fire together, wire together”). However, Hebbian learning is computationally suboptimal as it modifies weights unnecessarily beyond what is actually needed to achieve effective retrieval, causing more interference and resulting in a lower learning capacity. Our previous computational models have utilized a powerful, biologically plausible form of error-driven learning in hippocampal CA1 and entorhinal cortex (EC) (functioning as a sparse autoencoder) by contrasting local activity states at different phases in the theta cycle. Based on specific neural data and a recent abstract computational model, we propose a new model called Theremin (Total Hippocampal ERror MINimization) that extends error-driven learning to area CA3 — the mnemonic heart of the hippocampal system. In the model, CA3 responds to the EC monosynaptic input prior to the EC disynaptic input through dentate gyrus (DG), giving rise to a temporal difference between these two activation states, which drives error-driven learning in the EC→CA3 and CA3↔CA3 projections. In effect, DG serves as a teacher to CA3, correcting its patterns into more pattern-separated ones, thereby reducing interference. Results showed that Theremin, compared with our original model, has significantly increased capacity and learning speed. The model makes several novel predictions that can be tested in future studies. Author Summary: Exemplified by the famous case of patient H.M. (Henry Molaison) who had his hippocampus surgically removed, the hippocampus is critical for learning and remembering about everyday events — what is typically called “episodic memory.” The dominant theory for how it learns is based on the intuitive principle stated by Donald Hebb in 1949, that neurons that “fire together, wire together” — when two neurons are active at the same time, the strength of their connection increases. We show in this paper that using a different form of learning based on correcting errors (error-driven learning) results in significantly improved episodic memory function in a biologically-based computational model of the hippocampus. This model also provides a significantly better account of behavioral data on the testing effect, where learning by testing with partial cues is better than learning with the complete set of information.

How do humans learn from raw sensory experience? Throughout life, but most obviously in infancy, we learn without explicit instruction. We propose a detailed biological mechanism for the widely embraced idea that learning is driven by the differences between predictions and actual outcomes (i.e., predictive error-driven learning). Specifically, numerous weak projections into the pulvinar nucleus of the thalamus generate top–down predictions, and sparse driver inputs from lower areas supply the actual outcome, originating in Layer 5 intrinsic bursting neurons. Thus, the outcome representation is only briefly activated, roughly every 100 msec (i.e., 10 Hz, alpha), resulting in a temporal difference error signal, which drives local synaptic changes throughout the neocortex. This results in a biologically plausible form of error backpropagation learning. We implemented these mechanisms in a large-scale model of the visual system and found that the simulated inferotemporal pathway learns to systematically categorize 3-D objects according to invariant shape properties, based solely on predictive learning from raw visual inputs. These categories match human judgments on the same stimuli and are consistent with neural representations in inferotemporal cortex in primates.