Descripción del proyecto
El estudio de los mecanismos de aprendizaje revela el poder del cerebro humano
La extraordinaria capacidad de aprendizaje del cerebro humano sigue superando la capacidad de cálculo del aprendizaje automático. No obstante, aún se desconocen algunos aspectos de las complejas operaciones de aprendizaje que acontecen en la maraña de circuitos neuronales del cerebro. En el proyecto DopamineLearnLoops, financiado por el Consejo Europeo de Investigación, se actualizará el modelo estándar que plantea que un único error en el sistema dopaminérgico es suficiente para «actualizar» toda la red neuronal. En la nueva teoría propuesta se postula que múltiples sistemas de aprendizaje dinámicos basados en la dopamina funcionan en circuitos paralelos. Su equipo empleará técnicas punteras para determinar los mecanismos subyacentes a los circuitos, cómo funcionan en un entorno dinámico y qué algoritmos utilizan. Comprender la forma en que el cerebro aborda problemas complejos contribuirá, sin duda, al desarrollo de nuevos algoritmos de aprendizaje profundo por refuerzo inspirados en el cerebro.
Objetivo
The brain’s ability to learn is arguably its most exceptional capacity. Learning in biological brains far surpasses machine learning and requires much less training. How does the brain accomplish this? Why is biological learning still better than the most advanced machine learning algorithms to date? According to the standard model of reward-based learning in the brain, a single error signal is broadcast from the dopamine system and used to update the entire network, implementing a simple form of reinforcement learning. However, the standard model fails to predict several recent experimental findings, leaving open the question of how learning is implemented in the brain. In this project, I propose a new theory of how the brain learns: learning is implemented by multiple dopamine-based learning systems working in parallel circuit loops. These loops relay partial error signals to specific processing areas and permit independent evaluation of the value of different features in the external environment as well as the internal state, enabling learning of complex tasks with multiple relevant features. The loops are engaged dynamically according to the demands of the task, enabling the system to be flexible for learning a wide variety of behaviours of varying complexity. The presence of multiple dynamic parallel learning loops might enable the ability to generalize learning, which is currently the hallmark of biological intelligence. We will use state-of-the art techniques under the framework of our theory to elucidate basic mechanisms underlying the functional circuitry of the learning system (Aim 1), how it operates under different behavioural dynamics (Aim 2), and what algorithm it implements (Aim 3). Success of this project will enable a novel understanding of how the brain learns complex tasks as well as pave the way for the development of new brain-inspired deep reinforcement-learning algorithms.
Ámbito científico
Programa(s)
- HORIZON.1.1 - European Research Council (ERC) Main Programme
Régimen de financiación
HORIZON-ERC - HORIZON ERC GrantsInstitución de acogida
32000 Haifa
Israel