Neuromathématiques : année 2015/2016
-
Mardi 3 mai 2016, 14h30-16h30, salle de conférence de l’European Institute of Theoretical Neuroscience, 74, rue du Faubourg Saint-Antoine 75012, Paris.
Florent Meyniel
CEA, Neurospin Center, France
A normative account of the sense of confidence during probabilistic learning
The sense of confidence has been studied by psychologists over the past century. It has been under scrutiny only recently in the field of neuroscience. I will briefly review the topic and present the idea that the viewpoints of psychology and neuroscience on confidence can be unified by a definition of confidence as Bayesian probability. After this general introduction, I will present a focused investigation of the sense of confidence in a learning context. Learning in an environment that is both stochastic and changing consists of estimating a model from a limited amount of noisy data. Learning is therefore inherently uncertain, and at least in humans, the learning process is accompanied by a “feeling of knowing” or confidence in what has been learned. The talk will address the characteristics, the origin and the functional role of subjective confidence during learning using behavioral and functional MRI data in humans.
To this end, we developed a probabilistic learning task in which human subjects estimated the transition probabilities between two stimuli in a sequence of observations. The true probabilities changed unexpectedly over time and from time to time, subjects reported their probability estimates as well as their confidence in those estimates. We computed the optimal solution for this learning problem and we used it to analyze subjects’ data from a normative viewpoint. Behavioral data showed that humans not only infer a model of their environment, but they also accurately track the likelihood that their inferences are correct. Several characteristics of these confidence reports support that learning and estimating confidence in what has been learned may arise from the same, close-to-optimal probabilistic inference. Functional MRI data showed that the brain may resort to a hierarchical inference to solve this learning problem, and that confidence may be used in the learning algorithm to weight optimally the previously acquired knowledge against and the new incoming evidence.
Vous pouvez retrouver l’affiche de ce séminaire ici
-
Mardi 5 avril 2016, 14h30-16h30, salle de conférence de l’European Institute of Theoretical Neuroscience, 74, rue du Faubourg Saint-Antoine 75012, Paris.
Bertrand Thirion
INRIA Saclay-Ile-de-France
Seeing it all: Convolutional network layers map the function of the human visual system
How to demonstrate and analyze the complexity of visual experiences in a brain mapping framework? The key to this seems to reside in using natural stimulation while increasing the capacity of the analysis system. In this presentation we discuss a predictive model of brain activity following visual stimulation using the layers of a contest-winning object recognition convolutional network. We find that it explains both high-level and low-level visual areas well and that it can serve as a reliable predictor of brain activity for previously unseen stimuli. We use it to synthesize classical contrast-driven fMRI experiments and analyze the synthetic activity in a conventional way, revealing that the synthesis model captures the known details of the visual system. It is possible to recover classical contrast maps from this model on unseen images. To expose the brain mapping implicit in the model, we assess how well each contributing layer of the convolutional net fits each voxel. Visualizing these predictive scores reveals a profound correspondence between convolutional net layer depth and known hierarchy of visual cortical regions.
Vous pouvez retrouver l’affiche de ce séminaire ici
-
Mardi 1er mars 2016, 14h30-16h30, salle de conférence de l’European Institute of Theoretical Neuroscience, 74, rue du Faubourg Saint-Antoine 75012, Paris.
Laurent Perrinet
Institut de Neuroscience de la Timone
Towards understanding the inferential processes underlying the representation of trajectories in the primary visual cortex
Neural computations in the early visual system are optimized by evolution to efficiently process the trajectory of visual objects in natural scenes and in particular to modulate local mechanisms by the surrounding visual context. In the primary visual cortex, these computations are often characterized by the so-called association field, that is, by the set of rules that combine neighboring visual contour elements to refine more global visual processes. I will first show a simple method to compute the statistics of neighboring contour elements in static images. Surprisingly, we will show that this statistics are sufficient to characterize the category an image belongs to (for instance if it contains an animal), a function usually attributed to higher visual areas. Extending this endeavor to the temporal trajectory of a moving oriented bar, I will present results of a maximum likelihood decoding strategy applied to extracellular activity recorded in the primary visual cortex of behaving macaque monkeys (V1). The orientation and direction decoded in neural activity exhibits the signature of an anticipatory inferential processes optimizing the representation of the bar’s trajectory in V1. I will discuss these results in light of a probabilistic model of V1 integrating an explicit knowledge of sensory delays and minimizing its free-energy. This will allow to discuss the implications of these neuronal solutions to the representation of time in the brain, which is essential for the proper fusion of information in the central nervous system.
Vous pouvez retrouver l’affiche de ce séminaire ici
-
Mardi 2 février 2016, 14h30-16h30, salle de conférence de l’European Institute of Theoretical Neuroscience, 74, rue du Faubourg Saint-Antoine 75012, Paris.
Jonathan Touboul
Collège de France & INRIA
The pinwheel-dipole structure of orientation and spatial frequency, and their common organizing principles
In the early visual cortex of higher mammals, information is processed within functional maps whose layout is thought to underlie visual perception. Here, I will present a few theoretical thoughts together with experimental data on the possible principles at the basis of their architecture, as well as their role in perception. Using new optical imaging data with high resolution, I will show that spatial frequency preference representation exhibits singularities, precisely co-located with pinwheels, and around which the spatial frequency map organizes as an electric dipole potential. This is particularly interesting theoretically: I will demonstrate that both pinwheel and dipoles are the unique topologies ensuring exhaustive representation of both attributes while being optimally parsimonious. Eventually, I will raise the question of the functional advantages and drawbacks of the topology. I will show that the pinwheel dipole topology leaves room for a balanced detection of both attributions. But simulations predict that selectivity shall be sharper near singularity to ensure balanced detection, which I will confirm on biological data. This is a joint work with J. Ribot, A. Romagnoni, C. Milleret and D. Bennequin.
Vous pouvez retrouver l’affiche de ce séminaire ici
-
Mardi 1er décembre 2015, 14h30-16h30, salle de conférence de l’European Institute of Theoretical Neuroscience, rue du Faubourg Saint-Antoine 75012, Paris.
Sophie Deneve
Laboratoire de Neurosciences Cognitives, ENS
Efficiency turns the table on neural encoding, decoding and noise
Sensory neurons are usually described with an encoding model, e.g. a function that predicts their response from the sensory stimulus, e.g. with receptive field (RF) or a tuning curve. However, central to theories of sensory processing is the notion of « efficient coding ». We argue here that efficient coding implies a completely different neural coding strategy. Instead of a fixed encoding model, neural populations would be described by a fixed decoding model (i.e. a model reconstructing the stimulus from the neural responses). Because the population solves a global optimization problem, individual neurons are variable, but not noisy, and have no truly invariant tuning curve or receptive field. We review recent experimental evidence and implications for neural noise correlations, robustness and adaptation.
Vous pouvez retrouver l’affiche de ce séminaire ici
