Finding synthesis among neurocomputational accounts of working memory

Finding synthesis among neurocomputational accounts of working memory Q&A

We will present work that starts to build a bridge between formal models of visual working memory for simultaneously presented arrays on the one hand, and models for explaining recall of sequentially presented lists on the other. We start from our Interference Model (Oberauer & Lin, 2017) and extend it to the recall of sequentially presented lists of visual objects. The model builds on the assumption that access access to items in working memory relies on cue-based retrieval, which engenders interference as a main source of performance limitation. To account for serial-position effects, we add two new assumptions: With each new item working memory is updated by automatically down-grading previous contents. To counteract this process, encoding of early list items receives an attentional boost, which declines for successive items, creating a primacy gradient. We present a Bayesian hierarchical version of the model, and fits to data of a continuous-reproduction visual-working memory experiment in which up to six items were presented sequentially, and tested in a random order. The model accounts for the distributions of errors, the set-size effect, and serial-position effects over input and output positions.

Working memory is fundamental to cognition, allowing one to hold information “in mind.” A defining characteristic of working memory is its flexibility: we can hold anything in mind. However, typical models of working memory rely on finely tuned, content-specific attractors to persistently maintain neural activity and therefore do not allow for the flexibility observed in behavior. Here, we present a flexible model of working memory that maintains representations through random recurrent connections between two layers of neurons: a structured “sensory” layer and a randomly connected, unstructured layer. As the interactions are untuned with respect to the content being stored, the network maintains any arbitrary input. However, in our model, this flexibility comes at a cost: the random connections overlap, leading to interference between representations and limiting the memory capacity of the network. Additionally, our model captures several other key behavioral and neurophysiological characteristics of working memory.

Recent debate regarding the limits of working memory has focused on whether memory resources are better characterized as discrete or continuous, with models of each type competing to best capture the errors humans make in recall. I will argue that this apparent dichotomy is largely illusory, and that the critical distinction is instead between deterministic and stochastic mechanisms of WM, with only the latter being compatible with observed human performance and the underlying biological system. I will show that reconceptualizing existing models in terms of sampling reveals strong commonalities between supposedly opposing accounts. A probabilistic limit on how many items can be successfully recalled from WM is an emergent property of continuous models, despite these models having no explicit mechanism to enforce such a limit. Furthermore, adding stochasticity in the number of samples to a discrete model puts its ability to describe behaviour on a par with continuous models. Finally, stochastic sampling has a theoretical connection with biologically plausible implementations of WM based on the inherently stochastic activity of neural populations.

Early theories of working memory (WM) (Atkinson & Shiffrin, 1968; Baddeley & Hitch, 1974; Baddeley, 2000, Ericsson & Kintsch,1995) have discussed the essential role of visual knowledge and long-term memory in WM performance. Yet, there is no computational model to show how WM representations can be built from the visual knowledge using neurocomputational mechanisms. We propose a model of WM-visual knowledge that uses the latent representations in a visual knowledge network to encode information in memory. This neurally plausible model (named MLR) represents visual knowledge using a variational autoencoder (VAE; Kingma & Welling, 2013) that learns to compress and reconstruct the pixle-wise visual stimuli. We modified the VAE to represent shape and color in separate maps, to be able to test theories of shape-color bindings. The MLR model uses a binding pool (BP; Swan & Wyble, 2014) to flexibly store information in a shared neural resource, such that attributes of an item (e.g., shape and color) could be represented in one memory system, and bound together according to task demands. The storage of information in working memory is accomplished through self-sustaining activity patterns, and synaptic weights are fixed during storage and retrieval. We showed that MLR can efficiently store and retrieve familiar items when their compressed representation in visual knowledge is encoded. On the other hand, novel items could be encoded and retrieved from less compressed representations. Additionally, the MLR model could extract the categorical information of an item and store it in memory along with the visual information. Finally, our model provides an explanation of how WM is linked to visual knowledge, to store familiar stimuli efficiently, while also being able to build on-the-fly memories of novel stimuli.

A central feature of WM is its ablility to not only store information, but also utilise it. Most models of WM don't specify how information is read out and used. This shifts the burden to other systems. But recent work suggests that WM is task dependent, dynamic, and itself involves manipulation. We propose that WM representations should themselves be able to make decisions and select actions. WM should be effectual. Accordingly, one recent perspective suggests that WM holds a set of rules for transforming stimuli to responses. For example in change detection tasks, WM may function as a "matched filter", and in graded judgements, as a decision-making circuit. Can we model this? We propose that this functionality corresponds to the ability to set up multiple new attractor states. We show that a minimal array of Hebbian units can rapidly carve out an attractor that binds the features of an item. A matching input can trigger competition between these attractors, with the winner potentially triggering an action. The dynamics can "gate" items in and out of memory, but without needing a gatekeeper. We think of this as a primitive but concrete implementation of a multiple demand network, holding pointers that bind features in sensory stores. Competition between plastic attractors replicates Hick's law of decisions, trial history effects, and some conflict effects previously explained by "event files". We were also able to apply plastic attractors to a simple situation where WM gets put to immediate use: Visual search. The template that is being searched for is encoded into WM, and the search display acts as a memory probe, triggering attention to select the matching item. The model also extends to continuous feature domains, where it generates predictions similar to the interference model. We hope this work encourages a view where WM encodes information in terms of active effects or responses, making representations inherently potent.