We use rodents as the model system for most of our work. Rodents have sophisticated decision-making abilities and are able to act rationally — under existing constraints — based on their confidence about their beliefs. This makes them an ideal model organism for for studying the computations and neural basis of decision making. We also sometimes study the behavior of human subjects on psychophysical tasks, both because we are interested in features of decision-making that are generic across species, and also occasionally to explore more demanding cognitive processes. We typically focus on decisions about simple auditory stimuli, which can be controlled precisely without excessive interference from active sampling strategies.
Our general approach for studying a particular computation is to design a reduced behavioral paradigm for which this computation is a salient ingredient, and to analyze the subject’s behavior carefully, to make sure that they are indeed using the strategy that we want to study in order to solve the task. We believe that a good, ideally quantitative, understanding of the strategy the subject is actually using is very helpful in order to understand the algorithms at play and their neural implementation.
We use a parallel theoretical-experimental approach to study brain function. We use theory to be precise in what we mean when we describe the world, and experiments to keep our theories grounded and relevant.
Psychophysics is the study of the quantitative functioning of perception and sensory-guided judgements. Over the last 200 years, a number of precise mathematical regularities (psychophysical “laws”) have been discovered. The investigation of psychophysical laws represents an attractive approach for neuroscience, as they constitute powerful constraints that can be used to study the neural basis of perception and decision making. We have recently made an important contribution by discovering a new psychophysical regularity — at the level of reaction time in sensory discrimination experiments in rats and humans — which subsumes Weber’s Law, the oldest and most general psychophysical regularity. We termed this regularity the Time Intensity Equivalence in Discrimination (TIED). The TIED implies that changing the overall magnitude of two sensory stimuli being discriminated (by a common multiplicative factor) is completely equivalent to a change in the units of time of the sensory discrimination process. Because the TIED is a very strict regularity, it allowed us to rule out many previously proposed explanations of Weber’s law, and to identify its necessary algorithmic ingredients. In addition, the simplest implementation of the algorithm provides a virtually complete description of the accuracy and reaction time of the subjects. We are currently using this knowledge to design perturbation and recording experiments aimed at discovering the neural basis of Weber’s law and, by extension, of simple discriminations of sensory magnitude.
Normative theories of decision making
Normative theories of decision making specify “optimal” decision strategies given a well defined notion of success (a cost function) and existing constraints. Normative approaches provide useful benchmarks to interpret observed behaviour, clarify its functional consequences, and might reveal relevant overlooked constraints. Using tools from reinforcement learning and optimal control theory, we have developed a framework for studying how the cost of engaging in a task and suppressing alternative action policies (the cost of control) shapes perceptual decision making. Unlike standard approaches, which specify optimal policies as a function of the problem being solved by the agent, our strategy posits that optimal policies should also depend on existing action plans (possibly maladaptive for a particular task) and on the ability of the agent to suppress them if necessary. This theory explains a range of observed, but previously unaccounted for, data from sensory discrimination experiments, and clarifies how to think about optimal decision strategies in realistic (i.e., control-limited) biological agents.
The global dynamics of the brain undergoes coordinated changes — referred to as “brain states” — across different behavioural contexts, which both shape how sensory input is processed as well as its impact towards action. Using rodents as a model system, we have, throughout the years, made different contributions to this topic, from the clarification of how desynchronised states are possible given the strong recurrent connectivity of local cortical circuits, to the way brain state shapes cortical auditory representations at the population level, to the effect of brain state to the accuracy of sensory judgements. Our results suggest that, whereas the effect of brain state on sensory representations is salient and robust, the impact of brain state on choice is dynamic and subtle. In particular, cortical desynchronization — generally thought to be associated to situations where cognition is oriented towards the external environment — appears not to be necessary for good performance in sensory discrimination tasks in general, but becomes relevant when subjects make a mistake. We are currently exploring the extent to which errors reconfigure the brain circuits supporting performance in psychophysical tasks.
Laminar distribution of choice and stimulus related signals in sensory cortical areas
The anatomical logic of the cerebral cortex relies on its laminar organization, which defines rules of connectivity vertically and horizontally across hierarchically connected layers. However, clear functional correlates of the different cortical layers during behavior are still rare. We have been studying the involvement of neurons in the deep and superficial layers of the mouse auditory cortex during a delayed frequency discrimination task. Our results reveal a functional dissociation between deep and mid-superficial neurons. Information about the identity of the sound (distributed across layers but larger in mid-superficial layers) is largest right after stimulus onset, but decays almost completely by the end of the delay period. In contrast, information about the upcoming choice is largely restricted to the deep layers and gradually builds up during the delay period. This temporal separation of stimulus- and choice-related signals, as well as analysis of noise correlations, suggests a top-down origin for choice-related signals in the deep layers of auditory cortex.
Neural correlates of instructed movements
Recent results suggest that movement is widely and largely indiscriminately represented across the cortex, including in areas whose contribution to the specification of those movements is not apparent. Because the outcome of task-related computations is produced through instructed movements, the ubiquity of movement-related neural signals makes it difficult to interpret and quantify the potential causal contribution of these signals towards a specific behavior. In addition to experiments which directly perturb neural activity– which are still somewhat coarse and difficult to interpret — one can gain insight into the functional role of a particular neural correlate by examining its temporal relationship with the corresponding instructed movement. We have developed a simple and robust regression-based algorithm that fits the moment-to-moment lag between ongoing neural population activity and ongoing behavioral signals. Using this method, we’ve provided evidence that neural activity in the mouse prefrontal cortex has a flexible temporal relationship with instructed movement — in this case speed in a treadmill — with reversals in the sign of the lag between neural activity and speed. Such reversals are consistent with prefrontal activity having a putative causal role (speed lags neurons) at moments where changes in speed are “freely” chosen by the mouse, and with speed having a putative causal role on neural activity (neurons lag speed) when changes in speed are externally instructed.