The prefrontal cortex has long been thought to subserve both working memory (the holding of information online for processing) and “executive ” functions (deciding how to manipulate working memory and perform processing). Although many computational models of working memory have been developed, the mechanistic basis of executive function remains elusive, often amounting to a homunculus. This paper presents an attempt to deconstruct this homunculus through powerful learning mechanisms that allow a computational model of the prefrontal cortex to control both itself and other brain areas in a strategic, task-appropriate manner. These learning mechanisms are based on subcortical structures in the midbrain, basal ganglia and amygdala, which together form an actor/critic architecture. The critic system learns which prefrontal representations are task-relevant and trains the actor, which in turn provides a dynamic gating mechanism for controlling working memory updating. Computationally, the learning mechanism is designed to simultaneously solve the temporal and structural credit assignment problems. The model’s performance compares favorably with standard backpropagation-based temporal learning mechanisms on the challenging 1-2-AX working memory task, and other benchmark working memory tasks.

Abstract not found

Abstract not found

The prefrontal cortex (PFC) has long been thought to subserve both working memory and “executive” function, but the mechanistic basis of their integrated function has remained poorly understood, often amounting to a homunculus. This paper reviews the progress in our lab and others pursuing a long-term research agenda to deconstruct this homunculus by elucidating the precise computational and neural mechanisms underlying these phenomena. We outline six key functional demands underlying working memory, and then describe the current state of our computational model of the PFC and associated systems in the basal ganglia (BG). The model, called PBWM (prefrontal-cortex, basal-ganglia working memory model), relies on actively maintained representations in the PFC, which are dynamically updated/gated by the BG. It is capable of developing human-like performance largely on its own by taking advantage of powerful reinforcement learning mechanisms, based on the midbrain dopaminergic system and its activation via the BG and amygdala. These learning mechanisms enable the model to learn to control both itself and other brain areas in a strategic, task-appropriate manner. The model can learn challenging working memory tasks, and has been corroborated by several important empirical studies.

The authors present their primary value learned value (PVLV) model for understanding the rewardpredictive firing properties of dopamine (DA) neurons as an alternative to the temporal-differences (TD) algorithm. PVLV is more directly related to underlying biology and is also more robust to variability in the environment. The primary value (PV) system controls performance and learning during primary rewards, whereas the learned value (LV) system learns about conditioned stimuli. The PV system is essentially the Rescorla–Wagner/delta-rule and comprises the neurons in the ventral striatum/nucleus accumbens that inhibit DA cells. The LV system comprises the neurons in the central nucleus of the amygdala that excite DA cells. The authors show that the PVLV model can account for critical aspects of the DA firing data, making a number of clear predictions about lesion effects, several of which are consistent with existing data. For example, first- and second-order conditioning can be anatomically dissociated, which is consistent with PVLV and not TD. Overall, the model provides a biologically plausible framework for understanding the neural basis of reward learning.

In order to perform object recognition it is necessary to learn representations of the underlying components of images. Such components correspond to objects, object-parts, or features. Nonnegative matrix factorisation is a generative model that has been specifically proposed for finding such meaningful representations of image data, through the use of non-negativity constraints on the factors. This article reports on an empirical investigation of the performance of non-negative matrix factorisation algorithms. It is found that such algorithms need to impose additional constraints on the sparseness of the factors in order to successfully deal with occlusion. However, these constraints can themselves result in these algorithms failing to identify image components under certain conditions. In contrast, a recognition model (a competitive learning neural network algorithm) reliably and accurately learns representations of elementary image features without such constraints.

The first two authors contributed equally to this research Revised manuscript, submitted to Neural Computation

& We address the connection between conceptual knowledge and cognitive control using a neural network model. This model extends a widely held theory of cognitive control [Cohen, J. D., Dunbar, K., & McClelland, J. L. On the control of automatic processes: A parallel distributed processing model of the Stroop effect. Psychological Review, 97, 332–361, 1990] so that it can explain new empirical findings. Leveraging other computational modeling work, we hypothesize that representations used for task control are recruited from preexisting representations for categories, such as the concept of color relevant to the Stroop task we model here. This hypothesis allows the model to account for otherwise puzzling fMRI results, such as increased activity in brain regions processing to-be-ignored information. In addition, biologically motivated changes in the model’s pattern of connectivity show how global competition can arise when inhibition is strictly local, as it seems to be in the cortex. We also discuss the potential for this theory to unify models of task control with other forms of attention. &

A large and influential class of neural network architectures use post-integration lateral inhibition as a mechanism for competition. We argue that these algorithms are computationally deficient in that they fail to generate, or learn, appropriate perceptual representations under certain circumstances. An alternative neural network architecture is presented in which nodes compete for the right to receive inputs rather than for the right to generate outputs. This form of competition, implemented through pre-integration lateral inhibition, does provide appropriate coding properties and can be used to efficiently learn such representations. Furthermore, this architecture is consistent with both neuro-anatomical and neuro-physiological data. We thus argue that pre-integration lateral inhibition has computational advantages over conventional neural network architectures while remaining equally biologically plausible.

We investigate how different forms of plasticity shape the dynamics and computational properties of simple recurrent spiking neural networks. In particular, we study the effect of combining two forms of neuronal plasticity: spike timing dependent plasticity (STDP) that changes synaptic strength and intrinsic plasticity (IP) that changes the excitability of individual neurons to maintain homeostasis of their activity. We find that the interaction of these forms of plasticity gives rise to interesting network dynamics characterized by a comparatively large number of stable limit cycles. We study the response of such networks to external input and find that they exhibit a fading memory of recent inputs. We then demonstrate that the combination of STDP and IP shapes the network structure and dynamics in ways that allow the discovery of patterns in input time series and lead to good performance in time series prediction. Our results underscore the importance of studying the interaction of different forms of plasticity on network behavior.