The neural and cognitive architecture for learning from a small sample

Abstract

Artificial intelligence algorithms are capable of fantastic exploits, yet they are still grossly inefficient compared with the brain's ability to learn from few exemplars or solve problems that have not been explicitly defined. What is the secret that the evolution of human intelligence has unlocked? Generalization is one answer, but there is more to it. The brain does not directly solve difficult problems, it is able to recast them into new and more tractable problems. Here we propose a model whereby higher cognitive functions profoundly interact with reinforcement learning to drastically reduce the degrees of freedom of the search space, simplifying complex problems and fostering more efficient learning.

Metacognition facilitates the exploitation of unconscious brain states

“…Because the stream of incoming perceptual information and the representational space itself (in the brain) are both high dimensional, in order to learn quickly the brain has to operate not at the feature level, but at a rather more abstract level (31). Together with metacognition, other cognitive functions such as episodic memory or attention may participate in this process: select few, relevant features to allow faster RL processes (6,25,32,33).…”

“…Exploiting unconscious states and reducing the dimensionality of the search space should thus be intimately linked. In the brain, synchronization of neurons through electrical coupling or synchronization between brain areas via higher order cognitive functions have been proposed as neural mechanisms controlling degrees-of-freedom in learning (6,34,35). Metacognition and consciousness could thus have a clear computational role in adaptive behaviour and learning (31,36), a point that is particularly interesting given the current success in developing artificial agents.…”

“…At the onset of RL, cortico-basal ganglia loops are predicted to be uniformly activated in a parallel search for the relevant (unconscious) states (6,13), alongside the basal ganglia (14). Thus, we should observe RL (RPE) -related activity in many areas of the cerebral cortex.…”

“…Results have so far indicated that metacognition interacts with RL, and that PFC and basal ganglia could be a potential neural substrate for this interaction. The metacognitive process over unconscious representations could use RPE to evaluate how close an estimated RL state is to a real RL state (6). From this viewpoint, if learning progresses, neural representations of confidence, RL state and RPE should become more synchronized.…”

“…Given the unconscious nature of the critically relevant information, it may seem improbable that subjects can learn to perform advantageously in this task. However, previously we have proposed that such problems may be solved via the mechanism of metacognition (6,7). We hypothesized that by creating low-dimensional representations in the prefrontal cortex (similar to the 'chunking' phenomenon in working memory (8)), metacognition may accelerate RL because operations occur in a reduced state-space rather than the original, multidimensional one.…”