Neural Mechanisms of Working Memory Limits: Abstracts Neural Mechanisms of Working Memory Limits: Abstracts
2013 Computational Neuroscience Meeting: Paris, France
Co-organizers: Albert Compte, Institut d'investigacions Biomediques August Pi i Sunyer
Zachary Kilpatrick, University of Houston
Workshops: July 17-18, 2013

A computational and experimental study of the relations between precision and capacity of visuo-spatial working-memory for several items
Rita Almeida, Karolinska Institute
Working-memory (WM) limitations have been mainly explained in terms of conceptual abstract models. In this work we used a neuronal circuit model of visuo-spatial WM (vsWM) to investigate WM limitations in terms of the number of items about which information is retained (capacity) and in terms of the precision of the information retained. The model assumes that there is a topographic organization of a storage circuit responsible for retention of spatial information and that the capacity of this storage circuit can be boosted by a non-specific top-down input from another area. These assumptions lead to specific predictions, which we tested and confirmed in two behavioural and one functional magnetic resonance imaging (fMRI) experiment. Our findings suggest that interference between similar memories underlies some WM limitations and that engagement of a top-down input can lead to an increase in capacity at a cost of a decrease in precision of WM.

Working memory capacity and allocation reflect noise in neural populations
Paul Bays, University College London
Errors in short-term memory increase with the quantity of information stored, limiting the complexity of cognition and behaviour. In visual memory, attempts to account for errors in terms of allocation of a limited pool of working memory "slots" or "resources" have met with some success, but the biological basis for this cognitive architecture is unclear. An alternative perspective attributes recall errors to noise in tuned populations of neurons that encode stimulus features in spiking activity. We show that errors associated with decreasing signal strength in probabilistically spiking neurons reproduce the pattern of failures in human recall under increasing memory load. Divisive normalization provides a plausible account of changes in neural activity with load, and a mechanism of weighting activity to prioritize fidelity of behaviourally-important stimuli. Analysis of human recall performance for high- and low-priority stimuli indicates weighting of neural activity is optimally tuned to meet performance goals.

Neural Dynamics of Working Memory Capacity Limits in Prefrontal and Parietal Cortex
Tim Buschman, Princeton University
Cognition has a severely limited capacity: adult humans can only retain about four items "in mind." This places a fundamental limitation on brain function as reflected in the high correlation between an individual's capacity and their fluid intelligence. Although human capacity limitations are well studied, their mechanisms have only recently been investigated at the single-neuron level. I will present data from simultaneous recordings in monkey parietal and frontal cortex while the animals performed a visual working memory task. Neural results revealed that visual capacity limitations occurred immediately upon stimulus encoding and in a bottom-up manner. In addition, we found evidence for a two-tiered model of working memory: the left and right halves of visual space had independent capacities (i.e. discrete resources) but within each hemifield, objects interfered within one another (i.e. shared resources). Together, our results suggest capacity limitations are primarily due to interference between objects during encoding.

Fundamental limits on persistent activity in stochastic attractor networks
Yoram Burak, Hebrew University of Jerusalem
How does intrinsic noise affect a neural network's ability to represent information about the past? This question has been studied in the context of highly idealized linear networks. Much less is known about nonlinear networks of spiking neurons. I will discuss how intrinsic noise limits short term memory of continuous variables in networks of Poisson spiking neurons, with a continuous manifold of attractor states (as proposed theoretically for head-direction cells in rodents or for the oculomotor integrator in goldfish). For such networks, it is possible to derive a precise analytical expression, describing how memory degrades through diffusion within the attractor. Furthermore, by combining statistical and dynamical approaches, it is possible to derive a fundamental limit on the network's ability to maintain a persistent state: the noise-induced drift of the memory state over time within the network is strictly lower-bounded by the accuracy of estimation of the network's instantaneous memory state by an ideal external observer. Thus, an intimate relationship exists between the neural network's ability to maintain a persistent state, and its coding properties - as characterized for example by the tuning curves of neurons in the network.

Bump attractor dynamics in prefrontal cortex underlie behavioral precision in spatial working memory
Albert Compte, Institut d'Investigacions Biomediques August Pi i Sunyer
Prefrontal persistent activity during the delay of spatial working memory tasks is thought to maintain spatial location in memory. A "bump attractor" computational model can account for this physiology and its relation to behavior. However, direct experimental evidence linking parameters of prefrontal firing to the memory report in individual trials is lacking, and to date no demonstration exists that bump attractor dynamics underlies spatial working memory. Here, I will demonstrate model-derived predictive relationships between the variability of prefrontal activity in the delay and the fine details of recalled spatial location, evident in trial-to-trial imprecise oculomotor responses of monkeys engaged in a spatial working memory task. The results support a bump attractor model for spatial working memory maintenance instantiated in persistent prefrontal activity. The findings reinforce persistent activity as a basis for spatial working memory, provide evidence for a continuous prefrontal representation of memorized space, and offer the first experimental support for bump attractor dynamics mediating cognitive tasks in the cortex.

A topological model of the hippocampal spatial map and spatial learning capacity
Yuri Dabaghian, Baylor College of Medicine
Since the discovery that certain hippocampal neurons fire in a location-specific way, we have known that these `place cells' serve a central role in forming this internal spatial map, but how they represent spatial information, in what capacity, and even what kind of information they encode, remains mysterious. Since the downstream brain regions must rely on place cell firing patterns alone, the temporal pattern of neuronal firing must be the key. Furthermore, because co-firing of two or more place cells implies spatial overlap of their respective place fields, a map encoded by co-firing should be based on connectivity and adjacency rather than distances and angles, i.e., it will be a topological map. Based on these considerations, we modeled hippocampal activity with a computational algorithm we designed using methods derived from Persistent Homology theory. We found not only that an ensemble of place cells can, in fact, learn the environment (form a topologically accurate map), but that it does so within parameters of place cell number, firing rate, and place field size that are uncannily close to the values observed in biological experiments beyond these parameters, this learning region, spatial map formation fails. Moreover, we find that the learning region enlarges as we make the computational model more realistic, e.g., by adding the parameter of theta precession. The structure and dynamics of learning region formation provide a coherent theoretical lens through which to view both normal spatial learning and conditions that impair it.

Balance between excitation and inhibition implies that neural variability is not neural noise
Sophie Deneve, Ecole Normale Superieure
Two observations about the cortex have puzzled and fascinated neuroscientists for a long time. First, neural responses are highly variable. Second, the level of excitation and inhibition received by each neuron is tightly balanced at all times. Here, we demonstrate that both properties are necessary consequences of neural networks representing information reliably and with a small number of spikes. Spikes of individual neurons communicate prediction errors about a common population-level signal, automatically resulting in balanced excitation and inhibition and highly variable neural responses. We illustrate our approach by focusing on the implementation of linear dynamical systems. Among other things, this allows us to learn a network of spiking neurons that can integrate sensory evidence and maintain it optimally, as required by decision making or working memory, without any added "neural noise", yet is extremely robust against perturbations. Most importantly, our approach shows that neural variability cannot be equated to noise.

Forgetting over seconds: attention and the role of the hippocampus
Masud Husain, University of Oxford
Mechanisms underlying rapid forgetting remain controversial. New methods to measure the precision of memory using a continuous, analogue response provide a sensitive means to probe this issue. Here we use such techniques to examine rapid forgetting in patients with focal medial temporal lobe lesions involving the hippocampus. Hippocampal patients exhibited larger errors over short delays specifically when several items had to be remembered, but not for a single item. Crucially, their errors were strongly associated with an increased tendency to report features belonging to other items in memory. Such binding errors support the view that the MTL is involved in linking together different types of information, regardless of memory duration. Intriguingly, even in healthy humans, introducing a visual search task during maintenance also disrupted recall, with increasing misbinding errors. One mechanism underlying rapid forgetting therefore appears to be corruption of feature bindings belonging to items held in memory.

Associative memory encoding in bump attractor networks: switching between dual functions on the same network
Vladimir Itskov, University of Nebraska
The hippocampus is implicated in learning and memory, but it is also critical for spatial navigation. The spatial navigation function has been successfully modeled using bump attractor networks -- i.e., networks of neurons where the synaptic efficacies vary according to the cells' relative positions in a "feature space" that reflects the coding properties of neurons. These networks are characterized by a highly structured pattern of synaptic connections, whereas functions such as associative memory encoding appear to require a different synaptic organization. How can the varied functions of the hippocampus be accomplished in the same network? Remarkably, we find that both functions can be realized on a network obtained as a perturbation of a bump attractor network. We suggest that sparse perturbations of bump attractor networks might be a generic mechanism that allows the same neuronal network to implement both associative memory and spatial navigation.

Optimizing working memory with heterogeneity of recurrent cortical excitation
Zachary Kilpatrick, University of Houston
A physiological correlate of working memory is a stimulus specific rise in neuron firing rate, which persists long after the stimulus is removed. Network models with short range excitation and long range inhibition show persistent neural activity, suggesting a clear link between network architecture and working memory. Cortical neurons receive noisy input fluctuations which cause a persistent state to diffusively wander about the network, degrading memory over time. It is unclear how the cortical architecture that supports working memory affects the diffusion of persistent neural activity. Using a combination of spiking network and simplified potential well models we show that spatially long range and heterogeneous excitatory coupling increases the stability of a discrete number of persistent states, which reduces the diffusion of persistent activity over the network. However, heterogeneous coupling also quantizes the space over which memories can be stored, limiting the capacity of working memory. The storage errors due to memory quantization and diffusion tradeoff with one another so that information transfer between the initial and recalled stimulus is optimized at a fixed network heterogeneity. When the retention time is sufficiently long, the optimal number of attractors is less than the number of possible stimuli, suggesting that memory networks can under-represent stimulus space to optimize performance. Our results give a clear framework to investigate how network architecture interacts with both a stimulus space and stochastic fluctuations to optimize memory storage.

Christian Machens, Champalimaud

Descriptive and normative models of working memory limitations
Ronald van den Berg, University of Cambridge
In the first part of this talk, I will argue that current psychophysical models of working memory limitations can be considered points in a three-dimensional model space that comprises 32 models in total. I will discuss results from a factorial model comparison, in which we tested all models on data from ten delayed estimation experiments from six labs. In the second part of this talk, I will argue that working memory can be conceptualized as an economic problem: the brain invests energy in the form of spiking activity to store items, and receives returns in the form of task performance. I will formalize this idea into a quantitative model and show that it accounts for delayed estimation data as well as the best descriptive model from the first part of my talk. Altogether, these results suggest that working memory limitations reflect a compromise between two conflicting ecological goals.

Prefrontal mechanisms contributing to working memory
Jonathan Wallis, University of California - Berkeley
Primate neurophysiologists have been studying how items are held in working memory since the 1970s, but there have been few studies using more than one item. Thus, it remains unclear what the neural processes are that cause the dramatic capacity limit of working memory. To explore this, we trained two animals to perform a primate analogue of a color change detection task that has been used to determine visual working memory capacity in humans and recorded neurons in the ventrolateral prefrontal cortex. Surprisingly, few neurons encoded the color of the items. Instead, the predominant encoding was of the spatial location of the items and a gain modulation of those signals that was consistent with attention. Our findings challenge the dominant view of the prefrontal cortex, which is that prefrontal cortex is responsible for representing task relevant information in working memory.