Neural network mechanisms of working memory interference

Abstract

[eng] Our ability to memorize is at the core of our cognitive abilities. How could we effectively make decisions without considering memories of previous experiences? Broadly, our memories can be divided in two categories: long-term and short-term memories. Sometimes, short-term memory is also called working memory and throughout this thesis I will use both terms interchangeably. As the names suggest, long-term memory is the memory you use when you remember concepts for a long time, such as your name or age, while short-term memory is the system you engage while choosing between different wines at the liquor store. As your attention jumps from one bottle to another, you need to hold in memory characteristics of previous ones to pick your favourite. By the time you pick your favourite bottle, you might remember the prices or grape types of the other bottles, but you are likely to forget all of those details an hour later at home, opening the wine in front of your guests.

The overall goal of this thesis is to study the neural mechanisms that underlie working memory interference, as reflected in quantitative, systematic behavioral biases. Ultimately, the goal of each chapter, even when focused exclusively on behavioral experiments, is to nail down plausible neural mechanisms that can produce specific behavioral and neurophysiological findings. To this end, we use the bump-attractor model as our working hypothesis, with which we often contrast the synaptic working memory model. The work performed during this thesis is described here in 3 main chapters, encapsulation 5 broad goals:

In Chapter 4.1, we aim at testing behavioral predictions of a bump-attractor (1) network when used to store multiple items. Moreover, we connected two of such networks aiming to model feature-binding through selectivity synchronization (2).

In Chapter 4.2, we aim to clarify the mechanisms of working memory interference from previous memories (3), the so-called serial biases. These biases provide an excellent opportunity to contrast activity-based and activity-silent mechanisms because both mechanisms have been proposed to be the underlying cause of those biases.

In Chapter 4.3, armed with the same techniques used to seek evidence for activity-silent mechanisms, we test a prediction of the bump-attractor model with short-term plasticity (4). Finally, in light of the results from aim 4 and simple computer simulations, we reinterpret previous studies claiming evidence for activity-silent mechanisms (5).