Attention-gated reinforcement learning of internal representations for classification

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Animal learning involves changes in neural connection efficacy, studied via neural networks.
• Supervised learning is efficient but biologically implausible; reinforcement learning is plausible but inefficient.
• A key challenge in reinforcement learning is the credit assignment problem: identifying crucial early processing units.

Purpose of the Study:

• To address the credit assignment problem in biologically plausible neural network learning.
• To introduce a novel learning scheme integrating attention and reinforcement signals.
• To demonstrate the efficiency and biological realism of the proposed model.

Main Methods:

• Development of attention-gated reinforcement learning (AGREL).
• AGREL utilizes a homogeneous reinforcement signal and an attentional feedback signal.
• The attentional feedback limits plasticity to critical early processing units.

Main Results:

• AGREL efficiently solves the credit assignment problem.
• The model achieves efficiency comparable to supervised learning in classification tasks.
• AGREL demonstrates biological realism by integrating key learning mechanisms.

Conclusions:

• Attention plays a crucial role in solving the credit assignment problem in neural learning.
• AGREL offers a biologically realistic and efficient framework for synaptic plasticity.
• The model integrates reinforcement learning, attention, and feedback mechanisms coherently.