Distinguishing examples while building concepts in hippocampal and artificial networks

Abstract

The hippocampal subfield CA3 is thought to function as an autoassociative network that stores sensory information as memories. This information arrives via the entorhinal cortex (EC), which projects to CA3 directly as well as indirectly through the dentate gyrus (DG). DG sparsifies and decorrelates the information before also projecting to CA3. The computational purpose for receiving two encodings of the same sensory information has not been firmly established. We model CA3 as a Hopfield-like network that stores both correlated and decorrelated encodings and retrieves them at low and high inhibitory tone, respectively. As more memories are stored, the dense, correlated encodings merge along shared features while the sparse, decorrelated encodings remain distinct. In this way, the model learns to transition between concept and example representations by controlling inhibitory tone. To experimentally test for the presence of these complementary encodings, we analyze the theta-modulated tuning of phase-precessing place cells in rat CA3. In accordance with our model's prediction, these neurons exhibit more precise spatial tuning and encode more detailed task features during theta phases with sparser activity. Finally, we generalize the model beyond hippocampal architecture and find that feedforward neural networks trained in multitask learning benefit from a novel loss term that promotes hybrid encoding using correlated and decorrelated representations. Thus, the complementary encodings that we have found in CA3 can provide broad computational advantages for solving complex tasks.

Memory out of context: Spacing effects and decontextualization in a computational model of the medial temporal lobe.

“…Evidence for rapid trisynaptic learning comes from findings such as the high inhibition featured in DG, which allows for a separation between highly similar pattern inputs (Leutgeb & Leutgeb, 2007; Vazdarjanova & Guzowski, 2004), the importance of area CA3 in learning new paired associates (Rajji et al, 2006), and the role of these regions together in discriminating between highly similar information in memory (Bakker et al, 2008). Note that here we suggest an important role for a disynaptic, ECin → CA3 → CA1 → ECout subpathway, following modeling results that this pathway can support generalization (Kang & Toyoizumi, 2024; Kowadlo et al, 2019) and learn via EDL (Zheng et al, 2022). The pathway from ECin → CA1 → ECout constitutes the monosynaptic pathway of the HC (Schapiro et al, 2017), which allows CA1 to directly encode target ECin activity (Grienberger et al, 2022) and sends activity from the HC back into the cortex.…”

Memory out of context: Spacing effects and decontextualization in a computational model of the medial temporal lobe

“…Evidence for rapid trisynaptic learning comes from findings such as the high inhibition featured in DG, which allows for a separation between highly similar pattern inputs (Leutgeb & Leutgeb, 2007; Vazdarjanova & Guzowski, 2004), the importance of area CA3 in learning new paired associates (Rajji, Chapman, Eichenbaum, & Greene, 2006), and the role of these regions together in discriminating between highly similar information in memory (Bakker, Kirwan, Miller, & Stark, 2008). Note that here we suggest an important role for a disynaptic, ECin → CA3 → CA1 → ECout sub-pathway, following modeling results that this pathway can support generalization (Kang & Toyoizumi, 2023; Kowadlo, Ahmed, & Rawlinson, 2019) and learns via EDL (Zheng et al, 2022). The pathway from ECin → CA1 → ECout constitutes the monosynaptic pathway of the hippocampus (Schapiro et al, 2017), which allows CA1 to directly encode target ECin activity (Grienberger, Magee, & Duncan, 2022) and sends activity from the hippocampus back into cortex.…”