Current AI systems excel in a narrow set of well-defined tasks and are resource hungry.
However, biological counterparts embody adaptive learning, integrate contextual information, require low energy, and learn without extensive annotated data compared to current AI approaches.
In this thrust we hypothesize that biological neural systems can be used to improve the design of AI algorithms and to better implement its architectures. For this purpose, we will incorporate the findings from experimental and theoretical neuroscience to learn, draw inspiration, and also identify exemplars that can be applicable to designing high-performing AI systems. There is increasing evidence that shows that the solution space of neuro-inspired AI either converges with or is complementary to the conventional mathematical and logic-based approaches in AI. In this endeavor, we explore the following topics:
- Develop new generalizable theories for intelligence that push the boundaries on existing models based on cortical processing (eg: lifelong learning, credit-assignment)
- Implement new computing substrates and accelerators incorporating single neuronal models with low energy portfolio, that are hallmark features for next generation AI;
- Obtain ground-truth data for complex spatial navigation and behavioral tasks, which will be used to generate new AI systems;
- Create synergistic datasets that can be beneficial to both neuroscientists and AI researchers.
- Applications in image-guided therapeutics, physiology, and pathophysiology.
To pursue this research, we will use transdisciplinary tools and research approaches from neuromorphic computing, machine learning, statistics, imaging, and experimental neuroscience, that will allow us to build bi-directional knowledge bridges.