Metacognition is the cognitive process by which, instead of just learning to associate a
response or a behavior with a situation, animals (and mainly primates) monitor the functioning
(and particularly errors) of simple cognitive processes, learn to inhibit automatic responses
and promote instead contextually appropriate behavioral rules.
Better understanding and modeling this process is important for several reasons. In cognitive
neuroscience, it paves the way to exploring higher cognitive functions like reasoning,
imagination and other kinds of deliberation-based thoughts. In Artificial Intelligence, it stands
on the same grounds as Generative AI and proposes different processes and algorithms that
might remedy several weaknesses of GenAI and suggest innovative brain-inspired extensions.
The Mnemosyne Inria team in Bordeaux and the Machine Learning CWI group in Amsterdam
have a long experience in cognitive computational neuroscience and have launched a project,
MetaBrain, with the aim of specifying the computational mechanisms of metacognition, also
defining relevant tasks to assess the performance of metacognitive models. In this perspective,
the two teams have recently begun considering how predictive coding, a major biologically
founded model of cortical learning and functioning, could be used to design metacognitive
functions (Brucklacher et al., 2025; Dora et al., 2021; Fontaine et al., 2025). Relying on their
past experience, both teams consider applications in the visual domain.
Accordingly, the role of the postdoctoral fellow to be recruited is to participate to the
MetaBrain project, under the following axes:
Axis 1: Specification of Metacognition and its main computational mechanisms:
When several elementary competences have been learned (possibility to associate a response
with a situation), the role of Metacognition (or cognitive control) is to control the
performances of the competences, the selection or the coordination of the most appropriate
ones depending on the context, or the creation of a new one. This is generally described
through three main mechanisms: (i) the possibility to monitor cues indicating difficulties in the
process of problem solving (errors or conflicts between resources), in order to inhibit
elementary default responses, (ii) working memory to keep in sustained activity the different
aspects to be integrated (goals and subgoals, predictions, constraints) and (iii) cognitive
flexibility corresponding to new goals and contextual rules that can be learned and integrated
in the process of problem solving.
Our two teams have already proposed models that address several aspects of these
mechanisms (Dagar et al., 2021; Kruijne et al., 2020; Nallapu et al., 2019; Van den Berg, 2023)
and are also aware of several other models in the literature proposing other candidate
mechanisms (see for example Alexander and Brown, 2015; Botvinick et al., 2001, Collins et al.,
2012; Domenech et al., 2015; Miller et al, 2024; VanRullen et al., 2021). All these models
indicate possible correspondence with cerebral circuitries and adaptive operations.
Nevertheless, they are many and split these general mechanisms in different pieces which are
not always consistent and may differ under several aspects. It will be consequently important
be become familiar with these elements, and propose a synthesis associating both a precise
description of the mechanisms and a map of their functional dependencies.
Axis 2: Integrating the framework of predictive coding:
Most models of metacognition evoked above consider that the elementary competences are
learned and represented in the cortex. Generally, this is done in the models with a classical
associative learning, whereas an alternative to this type of learning, predictive coding, is
developed for some years, along with strong biological and theoretical justifications (Friston,
2003). This formalism has been also studied recently in our teams (Fontaine and Alexandre,
2025; Brucklacher et al., 2025; Dora et al., 2021). It is particularly interesting in the perspective
of metacognition, since it explicitly manipulates error signals and is a strong basis for
generative models. Manipulating precision, a major parameter in predictive coding, could be
an interesting way to address metacognition, generally described as a way to control hyperparameters
in learning models.
A major contribution of this work will be to study this formalism and propose solutions to
integrate it within a metacognitive perspective. A good way to assess such a dual model will
be to implement it within a task, as described below.
Axis 3: Definition of relevant tasks in the domain of visual reasoning:
Although many standard tasks have been defined and shared for simple sensorimotor control,
it is not yet the case for cognitive control, generally corresponding to much more complex
behaviors. A variety of tasks have been proposed in models mentioned above but they
differently integrate fundamental constituents such as hierarchical and temporal
dependencies. In a view of standardization, the goal will be consequently to enumerate
properties that have to be assessed when developing such metacognitive models and propose
or design corresponding tasks.
Subsequently, the postdoctoral fellow will work on integrating the insights from Axes 1 and 2
and task definitions in this Axis, with an architecture that integrates selected mechanisms from
the different frameworks, particularly under the perspective of extending and evaluating
models proposed in our teams with novel properties.
This postdoc position is proposed for 24 months, starting on November 1st, 2026 and will be
mainly located in the Mnemosyne team, in Bordeaux, France, with visits to the CWI partner in
Amsterdam. More specifically, in addition to regular visio-meetings, a 2-3-weeks stay in
Amsterdam each year will be organized.
