Periodic Reporting for period 1 - COVADIS (Deciphering the Code of Value Signals in the Human Brain)
Período documentado: 2021-01-01 hasta 2022-12-31
Humans routinely make complex decisions between alternative courses of action. Understanding how our brains enable us to do this is a fundamental question in neuroscience. A central hypothesis is that the brain represents the motivational relevance of the choice options through value signals. However, the neural code for value has remained elusive, and how the brain computes such values for multidimensional options remains unknown. This lack of knowledge has significant implications for society. It hinders our ability to adequately treat neuropsychiatric goal-directed disorders that affect millions worldwide and hampers the growing artificial-intelligence industry, which aims to develop agents capable of autonomous value-based decisions. This interdisciplinary project aims to take an important step towards deciphering the neural code for value through the combined use of cognitive neuroscience, computational modelling, and behavioral psychophysics methods. Specifically, the main objectives were:
Objective 1: Deciphering the neural code for value. We know very little about how value signals are coded by neurons. Single-unit recordings in the orbital frontal cortex (OFC) of non-human primates – an analogue of the human ventromedial prefrontal cortex (vmPFC) – have identified neurons that encode the subjective value of options in a linear manner, either increasing or decreasing their activity with value. However, non-linear population codes, where each neuron is tuned to a specific stimulus magnitude, could allow neuronal ensembles to represent both value magnitude and uncertainty simultaneously. Our aim was to investigate whether the human brain utilizes such a probabilistic neural code for value.
Objective 2: Establishing how the neural code for value shapes behavior. If the human brain employs a probabilistic code for value, representing both value and value uncertainty within the same neuronal populations, the properties of these value signals could explain why individuals vary in their preferences, make inconsistent choices, and how they form confidence judgments. Our aim was to establish a link between the neural code for value and how neural uncertainty about value shapes behavior.
Objective 3: Establishing how value is constructed from our senses. Rewards often consist of multiple dimensions, such as an apartment’s size, location, brightness, and price. How the brain integrates these attributes, inferred from our senses, into a global value remains elusive. Our aim was to investigate whether the brain calculates this global value by weighting the different attributes of a reward based on the precision of their perceptual representation.
Objective 1: Deciphering the neural code for value. We know very little about how value signals are coded by neurons. Single-unit recordings in the orbital frontal cortex (OFC) of non-human primates – an analogue of the human ventromedial prefrontal cortex (vmPFC) – have identified neurons that encode the subjective value of options in a linear manner, either increasing or decreasing their activity with value. However, non-linear population codes, where each neuron is tuned to a specific stimulus magnitude, could allow neuronal ensembles to represent both value magnitude and uncertainty simultaneously. Our aim was to investigate whether the human brain utilizes such a probabilistic neural code for value.
Objective 2: Establishing how the neural code for value shapes behavior. If the human brain employs a probabilistic code for value, representing both value and value uncertainty within the same neuronal populations, the properties of these value signals could explain why individuals vary in their preferences, make inconsistent choices, and how they form confidence judgments. Our aim was to establish a link between the neural code for value and how neural uncertainty about value shapes behavior.
Objective 3: Establishing how value is constructed from our senses. Rewards often consist of multiple dimensions, such as an apartment’s size, location, brightness, and price. How the brain integrates these attributes, inferred from our senses, into a global value remains elusive. Our aim was to investigate whether the brain calculates this global value by weighting the different attributes of a reward based on the precision of their perceptual representation.
We conducted a functional Magnetic Resonance Imaging (fMRI) study with 64 participants to decipher the neural code for value in the human brain. Participants performed a value and confidence rating task in the MRI scanner, where they rated the value of 64 food items and judged their confidence in these ratings. Outside the scanner, they chose between pairs of food items, picking their preferred one. We found that activity in the ventromedial prefrontal cortex (vmPFC) correlated with the subjective value participants assigned to the items, consistent with previous research. We extended these results using neuronal population decoding techniques to test whether the vmPFC employs a non-linear population code for value. We robustly decoded the subjective value expressed by participants from neural activity patterns in the vmPFC. This approach also enabled us to decode the precision of these neural value signals. We discovered that imprecise neural representations of value led to more variable preferences, inconsistent choices, and lower confidence. These findings provide a unified framework that links the neural code for value with fundamental aspects of behavior.
We then conducted an fMRI study with 40 participants to understand how the brain constructs the value of rewards with multiple attributes. Participants estimated the value of multisensory stimuli with two attributes: the orientation of moving dots and the auditory frequency of pure tone sequences. The more vertical the motion direction and the higher the tone frequency, the greater the value. We manipulated the precision of the visual attribute by changing motion coherence. Our finding showed that participants’ estimates of the multisensory stimulus value was influenced by both the levels of the auditory and visual attributes and the precision of the visual attribute. We decoded sensory imprecision from neural activity in the auditory and visual cortex and found that spontaneous fluctuations in sensory precision influenced value construction. These results provide strong evidence that the brain constructs the value of rewards with multiple attributes by weighting each attribute based on its perceptual precision.
These results have led to three manuscripts in preparation for submission to peer-reviewed neuroscience journals, and to presentations at four national and international conferences: the 2021 Zurich Neuroscience Center Annual Symposium in Zurich, Switzerland, the 2024 Orbitofrontal Cortex Meeting in Paris, France, the 2024 Annual Meeting of the Organization for Human Brain Mapping in Seoul, Korea, and the 2024 Annual Meeting of the Society for Neuroeconomics in Cascais, Portugal.
We then conducted an fMRI study with 40 participants to understand how the brain constructs the value of rewards with multiple attributes. Participants estimated the value of multisensory stimuli with two attributes: the orientation of moving dots and the auditory frequency of pure tone sequences. The more vertical the motion direction and the higher the tone frequency, the greater the value. We manipulated the precision of the visual attribute by changing motion coherence. Our finding showed that participants’ estimates of the multisensory stimulus value was influenced by both the levels of the auditory and visual attributes and the precision of the visual attribute. We decoded sensory imprecision from neural activity in the auditory and visual cortex and found that spontaneous fluctuations in sensory precision influenced value construction. These results provide strong evidence that the brain constructs the value of rewards with multiple attributes by weighting each attribute based on its perceptual precision.
These results have led to three manuscripts in preparation for submission to peer-reviewed neuroscience journals, and to presentations at four national and international conferences: the 2021 Zurich Neuroscience Center Annual Symposium in Zurich, Switzerland, the 2024 Orbitofrontal Cortex Meeting in Paris, France, the 2024 Annual Meeting of the Organization for Human Brain Mapping in Seoul, Korea, and the 2024 Annual Meeting of the Society for Neuroeconomics in Cascais, Portugal.
Our findings go beyond the classical view that value-related neurons employ a linear rate code, where firing rate correlates with subjective value. They demonstrate the existence of a non-linear population code for value in human vmPFC. Moreover, they establish a unified framework connecting this neural code with the probabilistic representation of value and shed light on how neural uncertainty about value influences preference variability, confidence and choice inconsistency – previously considered independent processes. Finally, they provide a mechanistic understanding of how the brain weights the multiple dimensions of a reward to construct its value: by weighting each attribute proportionally to its sensory precision. Our results may have a significant impact on the field of decision-making neurosciences by promoting the framework of probabilistic value signals and the use of population decoding techniques with fMRI. Furthermore, they may influence medical and translational research focused on understanding, diagnosing and treating behavioral disorders. Lastly, these findings could benefit the AI and robotics industries, helping develop new algorithms for choices and preferences in intelligent agents.