Brain-informed Artificial Intelligence: Random Graphs, Dynamical Systems and Beyond
Abstract:
Artificial Intelligence (AI) has made significant progress in recent years, increasingly enabling machines to perform tasks that were once considered unique to human cognition. Despite significant achievements, AI systems have not yet matched the depth of biological intelligence, especially in learning, adaptability, and creativity. Existing neuro-inspired AI models often fall short of fully capturing the complexity and richness of biological intelligence due to oversimplified assumptions about neuro-cognitive processes. This thesis introduces a new AI framework, Brain-Informed AI (BAI), designed to overcome these limitations by embedding more detailed, "brain-informed" inductive bias within AI architectures.
BAI is founded on principled Bayesian deep learning models grounded in modern brain science. It comprises two components: a task-specific module for executing various AI tasks, and a brain-informed module for encoding specific neuro-cognitive processes. This latter module is crucial, employing robust mathematical models, developed using random graphs, dynamical systems, and other advanced methodologies, to represent complex neuro-cognitive processes. These models are carefully designed to balance the intricacies of the processes with the computational efficiency required for integration into existing AI architectures. Beyond surpassing current benchmarks in areas such as acoustic modeling and domain adaptation, BAI allows for tackling previously unaddressed problems such as relational reasoning and timing prediction directly from sensory data.
Moreover, BAI introduces a new approach to artificial intelligence that emphasizes ``biological plausibility'' and theoretical soundness. This approach merges brain science insights with theoretically grounded Bayesian machine learning methods, enhancing AI’s reasoning, adaptability, and generative capability. Our framework has been validated through theoretical and empirical analyses, demonstrating substantial advancements in AI systems' reasoning, generation, and adaptation capabilities.
This thesis makes a significant contribution to the field of AI by offering a new approach that not only unlocks new capabilities of AI systems but also enhances our understanding of how artificial and biological intelligence can be integrated. The framework's ability to address complex, real-world problems with a brain-informed approach represents a paradigm shift in AI research, potentially influencing future developments across various domains of AI application.