Models of attractor dynamics in the brain

TL;DR

The paper surveys how attractor dynamics in autoassociative networks provide a common computational framework for memory, perception, and decision-making across hippocampus, IT cortex, perception, and working memory. It reviews four cases: discrete hippocampal place-cell attractors with remapping; IT category-specific attractor convergence; unified adaptation-priming dynamics in perception via firing-rate adaptation; and PPC–PFC coupled attractor dynamics underlying sensory-history biases in working memory. The work emphasizes the analytical tractability of attractor models and how they reveal how experience shapes neural representations and behavior. The findings offer mechanistic explanations and predictive frameworks linking neural dynamics to cognitive biases, with implications for understanding cognition and informing theoretical neuroscience.

Abstract

Attractor dynamics are a fundamental computational motif in neural circuits, supporting diverse cognitive functions through stable, self-sustaining patterns of neural activity. In these lecture notes, we review four key examples that demonstrate how autoassociative neural network models can elucidate the computational mechanisms underlying attractor-based information processing in biological neural systems performing cognitive functions. Drawing on empirical evidence, we explore hippocampal spatial representations, visual classification in the inferotemporal cortex, perceptual adaptation and priming, and working-memory biases shaped by sensory history. Across these domains, attractor network models reveal common computational principles and provide analytical insights into how experience shapes neural activity and behavior. Our synthesis underscores the value of attractor models as powerful tools for probing the neural basis of cognition and behavior.

Figures (7)

• Figure 1: Attractor dynamics in hippocampal place cells. (A) Exposure to two distinct "endpoint" environments led to separate, non-overlapping spatial firing patterns in hippocampal place cells (red is higher, blue is lower). (B) When animals were subsequently exposed to morphed intermediate environments, the place cell firing pattern shifted abruptly to resemble those of the closer endpoint environment (see wills2005attractor for details on the measure of similarity of place cells' firing rates between enclosures). (C) Quantification of these transitions confirmed discrete, attractor-like shifts in spatial representations, supporting the existence of attractor dynamics underlying place cell activity. Plots were adapted and reprinted from wills2005attractor with permissions.
• Figure 2: Modeling attractor dynamics in inferior temporal (IT) cortex. (A) Two monkeys were trained to perform a match-to-sample task where the sample could be a morph of the two option images. (B) In IT neurons selective to one of the images ("morph eff") but not the other ("morph ineff"), the activity elicited by intermediate morphs closer to the effective image was more similar to the effective image's activity than predicted by a linear dependence on morph level, resembling attractor dynamics ($300\,$ms post stimulus onset and later; solid and dashed lines indicate higher image-based similarity to the effective or ineffective morphs, respectively). (C) An autoassociative neural network model was constructed to simulate these dynamics, with orthogonal patterns stored as memories in the recurrent IT network. As seen in (D), when memory storage approached network capacity and firing rate adaptation was included, neurons selective for a stored pattern showed similar attractor-like convergence: morphs closer to the memorized pattern elicited neural activity similar to that of the stored pattern, paralleling the experimental observation in (B). Plots were adapted and reprinted from akrami2009converging with permissions.
• Figure 3: Example primer and test stimuli from webster2004adaptation, adapted from the JACNeuF and JACFEE image dataset of biehl1997matsumoto. The primers "happy" and "angry" (left and right, respectively) are presented before the test stimulus, a "neutral" (center). Behavioral results from the study show that subjects' perception of the neutral face is systematically biased by the primer: participants are more likely to judge the neutral face as "happy" or "angry" depending on whether the preceding primer was happy or angry, respectively.
• Figure 4: Design of the task in daelli2010recent probing the influence of primers in perceptual adaptation. In (A), a primer or "adapter" is first presented for a fixed period of $3\,$s, followed by a delay period of a varied length ($50\,$ms for experiments 1 and 3 and $3100\,$ms for experiments 2 and 3). The morphed object or ambiguous stimulus is then presented for $400\,$ms. The task ends after the subject is probed to answer which stimulus category the morphed image belongs to. (B) Example images shown to the subjects with both extremes (A and B) and the morph (A/B).
• Figure 5: Illustrated results from daelli2010recent of the priming effects based on delay duration. The top panel shows the experimental paradigm with short delays ($5$-$100\,$ms) between prime and target, resulting in adaptation aftereffects where ambiguous stimuli are perceived as less similar to the adapter (blue curve shows perceptual shift away from prototype A). The bottom panel shows the same paradigm with longer delays ($2$-$3\,$s), where the effect reverses to priming, with ambiguous stimuli perceived as more similar to the adapter (red curve shows perceptual shift toward prototype A). Both conditions used a $100\,$ms target presentation, demonstrating how temporal dynamics determine whether adaptation or priming dominates perception.