Periodic Reporting for period 3 - NEURAL-PROB (A Bayesian sense of probability in the human brain: Its characteristics, neural bases and functions)
Okres sprawozdawczy: 2024-02-01 do 2025-07-31
The human brain constantly estimates probabilities to characterize the world. An accurate estimation of these probabilities enable us to better perceive what is around us and to make better decisions. For instance, an accurate estimation of the probability of experiencing a delay in your commute is useful to decide when to leave home to attend an important meeting. The human brain derives its probability estimates by different means; a prominent one is learning from past observations (e.g. from past delays in your commute).
This project makes the hypothesis that the estimation of probabilities is accompanied by a sense of confidence about the accuracy of the estimate, and that the sense of confidence plays a key role in regulating the learning of probability. The project aims to:
(1) better understand the way the brain estimates probabilities and confidence from a sequence of observations using behavioral experiments and computational models
(2) characterize the way probability and confidence are represented in the brain in terms of neural activity using modeling and brain imaging
(3) identify the neural mechanisms by which the learning of probabilities is regulated by confidence, using brain imaging and causal experiments.
A better understanding of the computation, representation and function of probabilities and confidence in the brain is important for the society. Since our decisions are based on the probabilities that we estimate (consciously or not), the better the probability estimates, the better the decisions. Understanding probability estimation and representation will help in the future to design interventions to make people better at estimating probabilities, and better at using these probabilities. In addition, since confidence regulates the learning of probability, it controls the extent to which people adapt their learning over time. Accurate learning requires to strike the balance between flexibility (to update our estimates when the world changes) and stability (to maintain our estimates when the world fluctuates randomly).
This project makes the hypothesis that the estimation of probabilities is accompanied by a sense of confidence about the accuracy of the estimate, and that the sense of confidence plays a key role in regulating the learning of probability. The project aims to:
(1) better understand the way the brain estimates probabilities and confidence from a sequence of observations using behavioral experiments and computational models
(2) characterize the way probability and confidence are represented in the brain in terms of neural activity using modeling and brain imaging
(3) identify the neural mechanisms by which the learning of probabilities is regulated by confidence, using brain imaging and causal experiments.
A better understanding of the computation, representation and function of probabilities and confidence in the brain is important for the society. Since our decisions are based on the probabilities that we estimate (consciously or not), the better the probability estimates, the better the decisions. Understanding probability estimation and representation will help in the future to design interventions to make people better at estimating probabilities, and better at using these probabilities. In addition, since confidence regulates the learning of probability, it controls the extent to which people adapt their learning over time. Accurate learning requires to strike the balance between flexibility (to update our estimates when the world changes) and stability (to maintain our estimates when the world fluctuates randomly).
We have proposed a general framework to study the neural representation of uncertainty (and thus, of probability and confidence in particular). This framework distinguishes two methods that researchers currently use, one making strong assumptions about the neural code of our belief and the associated uncertainty, the other looking for correlations between uncertainty (reported by participants, or inferred based on some model) and neural activity. This work, published in Nature Neuroscience, stresses the pros and cons of each approach, and the benefit of combining them.
Regarding the role of confidence in the regulation of learning, we have demonstrated that human learning is mathematically sound. Notably, the learning dynamics depend on how informative each new observation is relatively to our current state of knowledge. In particular, learning increases when confidence about the current estimate is low. This effect was observed in the learning of magnitude (e.g. the size of an object) and even more of probability (e.g. of event occurring of not).
To explore the neural representation of probability and confidence, we have collected a high quality dataset in human participants using ultra high field functional magnetic resonance imaging (fMRI at 7 Tesla). With this dataset, we have identified the pieces of human cortex whose activity tracks confidence levels during learning. Using artificial neural networks, we have also demonstrated that networks that are trained to estimate probabilities accurately, automatically develop a representation of confidence that regulates their learning process. This result was observed even with very small networks, provided that they are equipped with specific elements of architecture that we identified. This result demonstrates that the estimation of confidence during learning is a biologically plausible process.
Last, in order to understand the biological mechanisms of the regulation of learning by confidence, we have started to test the possibility that noradrenaline, a neuromodulator of the brain, inversely follows confidence. Being released more when confidence is low, it would render neural networks more plastic and boost learning. We have devised special fMRI methods to measure the activity of the locus coeruleus, a small nucleus in the center of the brain that releases noradrenaline. We have validated our methods in a simple task, and collected data in a more complex learning task to test our hypothesis.
Regarding the role of confidence in the regulation of learning, we have demonstrated that human learning is mathematically sound. Notably, the learning dynamics depend on how informative each new observation is relatively to our current state of knowledge. In particular, learning increases when confidence about the current estimate is low. This effect was observed in the learning of magnitude (e.g. the size of an object) and even more of probability (e.g. of event occurring of not).
To explore the neural representation of probability and confidence, we have collected a high quality dataset in human participants using ultra high field functional magnetic resonance imaging (fMRI at 7 Tesla). With this dataset, we have identified the pieces of human cortex whose activity tracks confidence levels during learning. Using artificial neural networks, we have also demonstrated that networks that are trained to estimate probabilities accurately, automatically develop a representation of confidence that regulates their learning process. This result was observed even with very small networks, provided that they are equipped with specific elements of architecture that we identified. This result demonstrates that the estimation of confidence during learning is a biologically plausible process.
Last, in order to understand the biological mechanisms of the regulation of learning by confidence, we have started to test the possibility that noradrenaline, a neuromodulator of the brain, inversely follows confidence. Being released more when confidence is low, it would render neural networks more plastic and boost learning. We have devised special fMRI methods to measure the activity of the locus coeruleus, a small nucleus in the center of the brain that releases noradrenaline. We have validated our methods in a simple task, and collected data in a more complex learning task to test our hypothesis.
Until completion of the project, we aim to better characterize the computations of probability and confidence in the human brain. We plan to test the extent to which the neural representation of confidence is general across domains (e.g. when learning about the occurrence of affectively neutral events, or affective events such as the delivery of some reward, or perceptual objects, in the visual and auditory domains).
The neural representation of probabilities themselves is less well known thant the neural representation of confidence. Several previous attempts have failed to identify neural representations of probability. We will test new hypotheses about more complex forms of coding of probability, which have been neglected in the past. Last, we will test more directly the hypothesis that noradrenaline is involved in the regulation of learning by confidence, by tracking the activity of the locus coeruleus (which releases noradrenaline in the brain) along confidence during a learning task, and by using interventions to test causally its implication.
The neural representation of probabilities themselves is less well known thant the neural representation of confidence. Several previous attempts have failed to identify neural representations of probability. We will test new hypotheses about more complex forms of coding of probability, which have been neglected in the past. Last, we will test more directly the hypothesis that noradrenaline is involved in the regulation of learning by confidence, by tracking the activity of the locus coeruleus (which releases noradrenaline in the brain) along confidence during a learning task, and by using interventions to test causally its implication.