Session 1: INTRODUCTION: PRINCIPLES OF BAYESIAN INFERENCE

Modelling Bayesian inference by cortical circuits in the predictive coding framework

Rafal BOGACZ

This talk will introduce the predictive coding framework for modelling the computations in cortical circuits. This model assumes that the sensory cortex infers the most likely values of attributes or features of sensory stimuli from the noisy inputs encoding the stimuli. Remarkably, the model describes how this inference could be implemented in a network of very simple computational elements, suggesting that this inference could be performed by biological networks of neurons. Furthermore, learning about the parameters describing the features and their uncertainty is implemented in these networks by simple rules of synaptic plasticity based on Hebbian learning. The talk will also discuss how this model could be used to capture behavioural data.

Principles and psychophysics of Active Inference in anticipating a dynamic probabilistic bias

Laurent PERRINET 

Session 2: PERCEPTION IN THE FACE OF UNCERTAINTY

Measuring the sensitivity of visual confidence

Pascal MAMASSIAN

Visual confidence refers to our ability to predict the correctness of our perceptual decisions. Knowing the limits of this ability, both in terms of biases (e.g. overconfidence) and sensitivity (e.g. blindsight),

is clearly important to approach a full picture of perceptual decision making. The measurement of visual confidence with the classical method of confidence ratings presents both advantages and disadvantages. In recent years, we have explored an alternative paradigm based on confidence forced-choice. In this paradigm, observers have to choose which of two perceptual decisions is more likely to be correct. I will review some behavioural results obtained with the confidence forced-choice paradigm. I will also present two ideal observers based on signal detection theory, one that uses the same information for perceptual and confidence decisions, and another one that has access to additional information for confidence. These ideal observers help us quantify the limitations of human confidence estimation.

Using Bayesian models to investigate attentional mechanisms in the human brain

Simone VOSSEL

The deployment of attention rests on predictions about the likelihood of events and there is now accumulating evidence that the generation of such predictions can plausibly be described by Bayesian computational models. These models can be regarded as variants of predictive coding and provide a principled prescription of how observers update their predictions after new observations. In my talk I will present a series of studies employing a novel variant of Posner’s location-cueing paradigm in which the proportion of valid trials (i.e., cue predictability) was manipulated over the course of the experiment to create volatile contingencies. Behavioural computational modelling results as well as data from neuroimaging and neuromodulation experiments will be presented to elucidate the brain mechanisms underlying the flexible control of attention by inferred predictability. Additionally, I will discuss the advantages and the limitations of the modelling approach applied in this work.

Optimizing scene decoding with "three-party" generative modelS

Emmanuel DAUCE (Institut des Systèmes):

The active inference framework (Friston, 2010; Friston et al, 2012) models sequential scene uncovered through movement. Stemming from the auto-encoding theory (Hinton, 1994), it introduces a new perspective for it formally links dictionary construction from data with optimal motor control. In particular, motor control is here considered as a particular implementation of a predictive process that actively participates in estimating a complex posterior distribution. A “three-party” generative framework based on three mutually independent domains (i.e. object-in-space, gaze orientation and visual field) is proposed here that allows developing a foveated scene-decoding algorithm. It is shown efficient on a digit recognition database, providing biologically-realistic saccades and state-of-the-art recognition rates through saving about 90% image processing cost.

Dynamical activity patterns in the macaque posterior parietal cortex during path integration

Dora ANGELAKI (Baylor College of Medicine)

Neural circuits evolved to deal with the complex demands of a dynamic and uncertain world. To understand dynamic neural processing underlying natural behaviour, we use a continuous-time foraging task in which humans and macaques use a joystick to steer and catch flashing fireflies in a virtual environment. In order to solve the task, monkeys must dynamically update their position estimates by integrating optic flow generated by self-motion, a process known as ‘path integration’. We introduce a probabilistic framework to refute a popular account of path integration that attributes biases to forgetful integration. We instead find that such biases are explained naturally by an optimal strategy that maximizes rewards while accounting for prior expectations about our own movements. Interestingly, both humans and monkeys continue to track the target even after it was long gone, such that variability in subjects’ eye positions mirrors their behavioural variability. Our results suggest that the output of integration may be embedded in the brain’s oculomotor circuit, such that the eye position provides a dynamic readout of one’s distance to target during visual path integration. We use multi-electrode array and laminar probes to sample the activity of a large number of neurons in the posterior parietal cortex and find that different neurons are active during different epochs of integration. Neurons exhibit rich temporal diversity such that the integration dynamics appear embedded in the dynamical pattern of population activity. We are currently applying statistical techniques to characterise the precise dynamics of population activity to understand the associated neural computations.

Session 3: DECISION MAKING BEYOND CLASSICAL REINFORCEMENT LEARNING

Adaptive coding in the dopaminergic system in health and disease

Kelly DIEDEREN (King's College London)

Recent theories have construed the brain as performing a specific form of hierarchical Bayesian inference, known as predictive coding. In these models, the brain predicts upcoming information by weighting violations in its expectations (prediction errors) relative to their precision (reliability); a process termed adaptive coding. Although dopamine is hypothesised to play a key role in the adaptive coding of cortical unsigned (absolute) prediction errors, no experimental data has addressed this hypothesis in humans.  We used dopaminergic pharmacological manipulations in conjunction with an associative learning fMRI task that required adaptive coding. A computational model that included precision-weighting of prediction errors, provided the best fit to participants behaviour. At the level of the brain, unsigned prediction errors were adaptively coded relative to their precision in the superior frontal cortex. Decreases in neural adaptation significantly correlated with decreased task performance. A dopamine antagonist significantly attenuated adaptive coding, in line with predictive coding hypotheses. These findings are likely to have important implications for understanding altered behaviour in individuals with dopamine-perturbed states such as psychosis. Indeed, in a separate dataset we observed that decreases in adaptive coding were associated with an increase in positive psychotic symptoms. 

Learning the payoffs and costs of actions

Rafal BOGACZ Bogacz (University of Oxford)

To select the most appropriate behaviour, the brain circuits need to learn about the consequences of different actions. Much evidence suggests that such learning takes place in a set of subcortical nuclei called the basal ganglia. The basal ganglia circuit is organized in two main pathways connected with initiation and inhibition of movements respectively. It has been proposed that the neurons in these two pathways separately learn about payoffs and costs of actions, which are then differentially weighted during decision making depending on the motivation state. However, it has not been shown what plasticity rules would allow the basal ganglia neurons to learn about payoffs and costs of actions. This talk will show that the learning rules, which have been previously proposed to learn reward uncertainty in addition to mean reward, also allow estimating payoffs and costs associated with different actions. The resulting model accounts for diverse experimental data ranging from properties of dopaminergic receptors to the effects of medications on behaviour.

Brain circuits of urgent decisions for action

David THURA (University of Montreal)

Animals, including humans, constantly interact with a dynamic and unpredictable environment through successions of decisions and actions. Where in the brain decisions between actions are determined? What is the computational mechanism that transforms relevant information into action? Does an appropriate regulation of decisions and actions (especially the speed-accuracy trade-off, SAT) involve shared neural substrates? In this presentation, I’ll be presenting a comparison of neural responses measured in both cortical (dorsolateral prefrontal, dorsal premotor and primary motor cortices) and subcortical (external and internal globus pallidus) areas during a probabilistic decision task (the tokens task) performed by monkeys in either slow or fast SAT regimes. Combined with an electrical manipulation approach, I’ll demonstrate that these single-unit activities support a mechanism in which structures share labor during dynamic decision-making between actions, especially regarding the deliberation process, the commitment itself and the SAT adjustments. Moreover, I’ll show that data are compatible with a simple computational model in which sensory evidence is quickly integrated and combined with an urgency signal that pushes the system until a commitment event. The urgency signal appropriately sets and adjusts the SAT during both action selection and execution, maximizing what matters the most for subjects, the rate of reward.

Session 4: CONTEXT –DEPENDENT MOTOR BEHAVIOR

Pathology as maladaptive optimality: A computational dissection of decision and action in OCD and Parkinson's disease

Lionel RIGOUX (Max Planck Institute)

Optimality models have been successfully applied over the last decades to describe the organization of actions and decisions in humans and other species. According to those theories, our behavior is guided by rationality principles that balance the costs and benefits of our actions and eventually maximize our survival. Such models, for instance, can explain the characteristic shape of our movements as the best trade-off between the energy expenditure and the expected outcome associated with every actions.

Those optimal models can also be reversed: by observing the choices and actions, we can infer on the internal representations that were used to determine the (optimal) behavior. Further, those models can help us dissect intricate behaviors and pin point the hidden potential causes of apparent maladaptive actions in neuro-cognitive disorders.More particularly, I will discuss how optimality models of action selection and production can respectively shed some lights on the complex symptoms seen in Parkinson's disease and Obsessive Compulsive disorders. Bayesian theory here will play a dual role. On the one hand, it will serve as a canonical model of belief updating, thus implementing an optimal model of cognition. On the other hand, Bayesian theory will be used a statistical tool to invert the computational models and, as such, will be instrumental in the identification and quantification of the latent causes of behavioral deficits.

Rapid delay compensation and state estimation following disturbances to the limb

Fred CREVECOEUR (Université Catholique de Louvain)

Weighting priors with sensory evidence in an optimal (Bayesian) way has proven a powerful framework for understanding how humans and animal make efficient decisions in the face of uncertainty. In the context of movement control, the same operation is complicated by the presence of temporal delays affecting the transmission of neural signals along the nerves. In this context, optimal state estimation requires a compensation for the delay through sensory extrapolation, prior to combining the extrapolated sensory data with current beliefs about the state of the limb. Here we will present a series of experiment showing evidence that the nervous system performs similar transformations very quickly, such that long-latency responses to mechanical disturbances (latency ~50-100ms in the upper limb) adapt to the expected perturbation profile. Furthermore, this estimate is quickly available to the visual system, suggesting that multisensory integration accounts for the presence of multiple delays across sensory systems. These results suggest that state estimation is performed within rapid feedback pathways and support flexible control. We discuss how these findings provide insight into the neural basis of movement control in human.

Decomposing the motor system

Opher DONCHIN (Ben-Gurion University of the Negev)

Abstract: Different models of the motor system decompose its function in different ways. In the talk, I examine some of the historical ideas about how the motor system is put together: feedforward models and feedback control models. I discuss, further, how people have mapped these models onto the underlying anatomy and physiology of the system. I propose an alternative model -- hierarchically structured and physiologically motivated -- that tries to incorporate the best insights from each model. In this context, I review some recent results from my own lab and those of my collaborators that is at least partially consistent with this model. I discuss our fMRI studies of cortical representation and their implications for the function of the parietal cortex. In addition, I discuss behavioral studies of noise and learning, and how they relate to the role of noise in  different parts of the system. Finally, I will discuss the limits of our model, especially in the context of the complex structure of explicit motor knowledge.

Short presentations by young researchers

Evaluating temporal predictive abilities in children with cochlear implants

Jacques PESNOT (Aix-Marseille University)

Despite decades of research and great advances in speech rehabilitation techniques, the outcome of a cochlear implantation in deaf children remains unpredictable. Instead of relying solely on low-level feature encoding abilities, we propose here to evaluate higher cognitive processes. Temporal predictive abilities, and more generally statistical learning has been proposed to be one of the major impairment in auditory deprived children. We want to test this hypothesis by using a combination of behavior, electrophysiology (EEG) and modeling. Our results will lead to a better understanding of the causes of the language impairments observed in children with cochlear implants and hopefully a good predictive tool of the outcome of the implantation.

What’s the hazard rate?

Adrian RADILLO (University of Houston) -https://www.math.uh.edu/~adrian/presentations.html

How do we learn the structure of the world? More specifically, how do we learn its hidden structures? I investigate whether, and how, humans and animals infer latent temporal features of their environment in restricted experimental settings. I explore two tasks that have recently been designed to study perceptual decision making in a changing environments. In the dots reversal task, derived from the random moving dots motion discrimination task, a human or monkey is presented with a cloud of moving dots on a screen, a fraction of which moves coherently to the right or to the left. The direction of motion of the coherently moving dots alternates stochastically between left and right during the trial at a fixed rate – the so-called hazard rate – and the observer must choose the last perceived motion direction at the end of the trial. In a similar fashion, the dynamic clicks task, derived from the Poisson clicks task, consists in two streams of auditory clicks being presented simultaneously to a rat’s ears (one stream per ear). Each click train is generated by an inhomogeneous Poisson process with an arrival rate that can take on two constant values (low vs. high). The two streams are coupled in the sense that they have same transition times (governed by the hazard rate) but distinct instantaneous click rates; both ears always receive clicks with opposite arrival rates. The rat must choose the stream with highest instantaneous rate at the end of the trial.

In both experiments, the true state of the environment is in one of two states and alternates during the trial in a history-independent way. The subject receives sequential noisy samples from the hidden state until interrogation time, at which point she is required to identify the state in which the trial ended. In this talk, I will present ideal-observer models for both tasks and leverage their mathematical formulation to gain insight on experimental results and suggest future experimental designs. Modelling the animal’s decision making process as a Bayes optimal inference algorithm is key in several regards. Firstly, the model may guide our intuition in explaining animal behaviour, which may in turn generate testable predictions to help future experiments design. Secondly, our models are capable of producing numerical benchmarks on task performance which can be compared to animal performance. Thirdly, since they capture a key computation that is relevant to both tasks – learning the hazard rate – our models can lead us to approximate models that are more likely to be implemented in the brain. After presenting our canonical evidence accumulation model in the discrete-time set- ting will explain the key differences that appear when taking the continuum limit in each task’s setting. While the continuous nature of the evidence stream in the dots-reversal task yields a system of Stochastic Differential Equations, the pulsatile nature of the clicks task yields a system of jump Ordinary Differential Equations. Next, I will show how robust each model is to perturbation in the hazard rate parameter, as a function of the stimulus Signal-to-Noise Ratio (SNR) (See preliminary figure 1). This will provide us an answer to the question: “In what parameter regimes does it matter to learn the stimulus hazard rate?”. Finally, I will explore how speed of learning is impacted by the stimulus SNR and the prior belief on hazard rate. This will allow me to distinguish which regions in task parameters space facilitate learning, and which hinder it. 

Fractionating reaction times to probe the validity of the Drift Diffusion Model parameters

Gabriel WEINDEL (Aix-Marseille University)

The sequential sampling framework is a cognitive-behavioral modeling approach that can account for a large variety of experimental data. These models have an important place in psychology as empirical tools to infer cognitive processes from data, making them particularly useful for fundamental research, but also clinical investigation. The framework assumes that choices between response alternatives are derived from a progressive, noisy accumulation of evidence toward a decision threshold. For example, the Drift Diffusion Model has been emphasized for perceptual decision-making, and describes this process by several key parameters: speed of accumulation (drift), decision threshold (threshold), and a time external to decision (which accounts for stimulus encoding and response execution, t0). Despite this model's popularity however, few tests have focused on the interpretative validity of the t0 parameter. We conducted empirical tests of the DDM's decomposition of response times into decision time (drift / threshold) versus non-decision time (t0) in human participants. We designed a two-alternative forced choice experiments in which the reaction times could be fractionated by trial into pre-motor and motor times (MT), based on the onset of muscular activity from the electromyographic (EMG) recordings, which should be contained in the t0 parameter. On this decomposition we observe that the MT is highly influenced by the decision context, contrary to the view of a fixed execution time following a “central” decision. When comparing to the DDM estimated parameters we observe no correlation between the effect observed on MT and the effect on the t0 parameter. In addition, estimating the DDM parameters only on the pre-motor times shows that removing the MT affects the estimation of the decision time under certain conditions. Our results therefore suggest that the interpretation made of the DDM parameters needs to be re-evaluated using external validation.

Posters

Asymmetrical coupling between eye and hand when aiming toward common or separate targets
Anaide THIBAUT (Aix-Marseille University), James Mathew, Chloé Pasturel, Laurent Perrinet, and Frederic R. Danion

Eye-hand coordination is central in many everyday activities. Literature has shown that reaction times (RT) of eye and hand are tightly coupled. However to our knowledge this coupling has always been shown in situations where both effectors have a similar spatial goal. As result it is unclear whether this temporal coupling arises because both effectors share a common action plan or simply because they are simultaneously recruited. To address this issue we designed an experiment in which participants had to reach right/left targets that could impose similar or opposite actions for eye and hand. Furthermore literature has shown that RTs are largely influenced by the stimulus probability, with shorter/longer RT for high/low probabilities. However to our knowledge it has never been explored whether eye and hand priors are computed independently. We reasoned that if eye and hand priors are computed in a joint manner, their RTs should remain largely coupled even when changing unexpectedly the target side for one effector only. Our results showed first that, although RTs were longer for incongruent than congruent actions (+60ms), the temporal coupling between eye and hand RT persisted. Second we found that, although eye and hand RTs increased when switching unexpectedly target side for the eye, only hand RTs increased when switching unexpectedly target side for the hand. Overall we conclude that the temporal coupling between eye and hand does not stem from a similar action plan, and is rather asymmetrical considering that the eye is more influent on the hand, than the other way around.        

Role of the Corpus Callosum in mediating interlimb transfer of motor skills: Insights from neurological patients

Penelope TILSLEY (Aix-Marseille University), Patricia Romaiguère, Eve Tramoni, Olivier Felician, Fabrice Sarlegna

Learning a motor skill can generalize to another scenario involving, for example, a different motor task or a different limb. The generalization of motor learning across limbs, known as interlimb transfer, has been well demonstrated by research on short-term sensorimotor adaptation, yet underlying neural mechanisms remain unclear (Criscimagna-Hemminger et al. 2003; Perez et al. 2007). Amongst the various theoretical models, many of them highlighted the corpus callosum (CC) as a key brain structure mediating interlimb transfer (Taylor and Heilman 1980; Parlow 1989). However, Criscimagna-Hemminger et al. (2003) reported interlimb transfer of force-field adaptation in a split-brain patient with complete commissurectomy. Here we aimed to expand on this research by studying a range of CC pathologies to clarify the role of the CC in interlimb transfer. According to the callosal access model, we hypothesized interlimb transfer in CC patients as compared to healthy controls. After assessing baseline performance in a reaching task, we used a confirmed prismatic perturbation procedure to assess interlimb transfer: participants wore prisms (which deviated the visual field by 17°) while reaching for 100 trials with the dominant arm before subsequent testing of the unexposed non-dominant arm looked to examine interlimb transfer. Preliminary data indicate normal prismatic adaptation and significant interlimb transfer for one patient with CC agenesis as well as another patient with CC lesions following a ruptured brain aneurysm. Whilst more controls and patients with CC lesions must be examined, our preliminary findings fall in line with the research done by Criscimagna-Hemminger et al. (2003) as it suggests that adaptation with the dominant arm/hesmisphere can generalize to the non-dominant arm through ipsilateral corticospinal projections.

What becomes of the reachable space in the eye of the cyclone?

Nicolas Leclere (Aix-Marseille University)

"In daily life, how we judge if an object is reachable or not in order to produce the right movement is a compelling question. A potential contribution of sensorimotor system to perceive the reachable space is highlighted by recent studies and we thus hypothesized an influence of sensorimotor adaptation on this perception. The aim of our study was to assess the effect of adaptation to a new force field environment (using a rotating platform) on the perceptive categorization of reachable space (using psychophysics). In our main study, we asked young healthy adults (N = 14) to perform a reachability judgment task before and after exposure to the perturbation. We conducted another study (N = 10) to control a potential effect of the rotation. Our results showed that only the sensorimotor adaptation have an influence on the reachability perception. More precisely, all the participants perceived their reachable limit shorter after the sensorimotor adaptation than before. We discussed this result in the context of internal model theory and proposed that it could be interpreted as a compression of the reachable space or a general space representation modification. Our study thus highlights a significant contribution of sensorimotor system in order to categorize our reachability space."

Asymetrical transfer of adaptation between eye and hand tracking when adapting to a visuomotor rotation
James MATHEW (Aix-Marseille University), Pierre-Michel Bernier, Cedric Goulon, Frederic Danion