Explainable neural networks that simulate reasoning.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Deep neural networks (DNNs) are limited by their "black box" nature, hindering understanding of neural information processing.
• DNNs lack crucial neurobiological features and cannot adequately simulate higher cognitive functions.
• Current DNNs are insufficient computational models for deciphering the neural basis of cognition.

Purpose of the Study:

• To propose a novel computational framework for modeling neural information processing.
• To demonstrate how neural circuits can directly encode cognitive processes using simple neurobiological principles.
• To develop explainable deep neural networks capable of simulating higher cognitive functions.

Main Methods:

• Implementation of a non-gradient-based machine learning algorithm to train Essence Neural Networks (ENNs).
• Utilizing neurobiological principles for direct encoding of cognitive processes within neural circuits.
• Testing ENNs on benchmark computer vision tasks and evaluating their ability to simulate cognitive functions.

Main Results:

• ENNs provide intrinsically explainable neural information processing, even on complex tasks.
• ENNs successfully simulate higher cognitive functions like deliberation, symbolic reasoning, and out-of-distribution generalization.
• ENNs exhibit brain-associated network properties, including modularity, distributed/localist firing, and adversarial robustness.

Conclusions:

• ENNs present a viable computational framework for deciphering the neural basis of cognition.
• This approach advances the pursuit of artificial general intelligence by creating more biologically plausible AI.
• ENNs bridge the gap between artificial intelligence and neuroscience by offering explainable and functionally rich models.