From neural coding to decision-making

Neurophysiological experiments on monkeys and rodents have highlighted the neural mechanisms of decision-making. Neural signals, that are correlated with specific elements of the decision-making process, reflect an accumulation of evidence until the decision is reached.In this thesis, I study a dynamic neural network model that uses a balance between excitation and inhibition to account for the mechanism of evidence accumulation. This model qualitatively accounts for neurophysiological data but has not been compared quantitatively with behavioral results from decision experiments with humans. During decision-making experiments, many behavioral effects can be observed, such as the slowing down of the decision after an error. Modeling these effects is critical to reproduce the decision-making process in animals and humans. I explore these different effects from the point of view of the neural attractor model. Despite the fact that this framework does not consist in the most common to study decision-making and its effects, it allows for detailed biophysical mechanisms.In this thesis, I show that this level of modeling does not just correspond to a refinement of the standard framework but is essential to reproduce some behavioral measures. Using a relaxation dynamics, the network accounts for many of the sequential effects such as history biases, post-error slowing and post-error improvement in accuracy. In a second step I have developed a psychophysics experiment in order to study confidence in decision-making. I have shown that an attractor network reproduces the sense of confidence of the participants, as well as the sequential effects due to confidence. These results show that the non-linear dynamics, characteristic of attractor neural networks, is essential to reproduce various aspects of decision-making.The last part of this manuscript consists in a study of the neural coding of information in a decision network. I focus on the learning process of a categorization task by the network. I show that a modulation of the reward signal by the confidence leads to a more efficient learning of the categorization task.