Biologically plausible gated recurrent neural networks for working memory and learning-to-learn.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Learning-to-learn, or meta-learning, accelerates knowledge acquisition but current artificial neural networks often use biologically implausible learning rules.
• Recurrent neural networks for meta-learning typically lack interpretability and brain-mappability, often relying on backpropagation-through-time, which is not biologically feasible.

Purpose of the Study:

• To propose a novel, biologically plausible neural network model for meta-learning that overcomes limitations of existing approaches.
• To develop a network capable of flexible information retention and forgetting using only local learning rules.

Main Methods:

• Introduction of RECOLLECT, a gated memory network with a single memory gate for flexible information management.
• Training RECOLLECT using biologically plausible trial-and-error learning, relying solely on local information.
• Evaluating RECOLLECT on a pro-/anti-saccade task and a reversal bandit task.

Main Results:

• RECOLLECT successfully learned to represent task-relevant information over extended memory delays in the pro-/anti-saccade task.
• The network demonstrated the ability to effectively flush its memory at the conclusion of trials.
• RECOLLECT exhibited meta-learning capabilities on the reversal bandit task, acquiring effective policies.

Conclusions:

• RECOLLECT provides a biologically plausible framework for meta-learning, utilizing local information and flexible memory.
• The network's learned solutions show resemblance to animal learning strategies, suggesting potential for modeling cognitive processes.
• This work advances the development of interpretable and brain-like artificial intelligence systems.