Predicting perceptual decisions using spectro-spatio-temporal EEG features

From :   to

Multi-stable perception: a continuous decision–making process
A visual multi-stable stimulus is a temporally stationary, but ambiguous stimulus, i.e., a stimulus that allows for dif-
ferent, mutually exclusive interpretations when visioned. When a subject is presented with a such stimulus, his or
her perception is spontaneously alternating over time between the different interpretations[Hupé and Rubin, 2003].
This perceptive multi-stability has been characterised in literature, mainly using descriptive statistics such as mean
residence time in a given interpretation, or the probability of changing from one defined percept to another defined
percept[Leopold and Logothetis, 1999, Parisot, 2020].
In our study, we investigate whether the underlying cognitive—but unconscious—decision process that makes
the subject change interpretation at some given time can be observed using electrophysiological and behavioural
measurements. Thereto, we are collecting electro-encephalographic (EEG), gaze, and behavioural measures in a
controlled experiment design (2x2 factorial design: ambiguous/non-ambiguous stimulus ; perceptual report by key-
press/no report). The stimulus that is used is composed of a moving plaid, superposed with a random dot kine-
matogram. Whereas the former is completely ambiguous, the latter allows to help enforce a given interpretation, as
such controlling what is perceived in half of the experimental trials. The collected data allows to have a proxy for the
perceptual decision (through a key-press; half of the trials) and comes accompanied by a continuous recording of eye
movements and positioning on the screen (gaze) as well as a cognitive window on the cerebral functioning (EEG).
The goal is to come up with a temporally predictive model of evidence accumulation resulting in perceptual decision making.


The candidate will start with a literature study, identifying potential EEG markers of decision making and evidence ac-
cumulation. Given the set of identified markers, the goal then consists in defining a predictive model for the decisions
reported through key-presses in the non-ambiguous stimulus and perceptual report by key-press condition.
The model will be probabilistic in nature, giving a decision probability at all time. The spectro-spatio-temporal
EEG features such as variations in instantaneous frequency, instantaneous power fluctuations, spatial power reallo-
cation, etc. will feed an evidence accumulation model: a drift-diffusion model under Gaussian additive noise, where
the drift and diffusion parameter as well as the value for the decision boundary will be a function of the observed fea-
ture values. Given a decision boundary and the actual evidence level (temporally integrated drift-diffusion process),
this will give us a probability density for the time-to-decision, or—in case of two alternative choices—the odds of one
choice that is preferred over another[Darling and Siegert, 1953][Ratcliff and McKoon, 2008].
This experiment is part of a larger project that has recently received a funding through ANR Grant Vision-3E
(started April 2022) which offers a fully funded PhD position in the continuity of the internship’s subject.

Desired profile

The candidate should have a solid basis in signal processing and time-frequency representations in particular. Scien-
tific programming skills should be at an autonomous level (preferably in python, at least matlab, octave, or similar).
Notions in modelling and stochastic signal processing will be appreciated

Host institution

The successful candidate
• is offered a stimulating working environment in which experimental setups are run on the IrmAGe platform
(Univ Hospital, UGA) and the Persee platform (GIPSA-lab) with a possibility to help with the data acquisition;
the programming environment is psychopy[Peirce et al., 2019].
• will develop competences in EEG (multi-channel, time-frequency, phase-synchronisation, etc.) processing,
more specifically using the MNE[Gramfort et al., 2013] python toolbox
• will develop competences in mathematical modelling, more specifically the drift-diffusion model applied to
visual decision–making processes
• will develop competences in predictive modelling (machine learning with temporal prediction)

Your opinion is important to us

Description de la soumission d'un avis

Your vote :
Your opinion is important to us