Demis Hassabis, Geoffrey Hinton, Ilya Sutskevar (MITCBMM Panel Chair) – CBMM10 Panel (Nov 2023)


Chapters

00:00:00 Workshop Questions on the Relationship Between AI and Neuroscience
00:04:20 Neuroscience's Influence on AI and AI's Insights into Neuroscience
00:08:42 Challenges and Opportunities in Developing Artificial Intelligence for the Real World
00:14:45 The Interplay of Neuroscience and AI in Understanding Intelligence
00:29:08 Theory, Neuroscience, and the Future of AI
00:34:26 AI theorization for future benchmarks
00:43:45 Complexity of Intelligence: From Animals to AI
00:47:23 Rethinking AI and Neuroscience Measurement
00:58:26 Developing AI from Models: Challenges, Opportunities, and Collaboration
01:10:42 Neuroscience and AI: Unraveling Human Intelligence and Its Implications
01:13:44 AI-Enabled Scientific Revolution: Exploring Creativity and Benchmarking Challenges
01:23:30 Future Breakthroughs in Machine Learning and Neuroscience

Abstract

The Intersection of Neuroscience and Artificial Intelligence: Exploring the Frontiers of Intelligence

Engaging the Mysteries of Intelligence: A Collaborative Journey Between AI and Neuroscience

In the field of modern scientific inquiry, few areas offer as rich a field of discovery as the intersection of neuroscience and artificial intelligence (AI). This fascinating nexus is not just a meeting point of two disciplines, but a vibrant ground for exploring the deepest questions about intelligence. This article delves into the insights and inquiries of notable scholars like Tommaso Poggio, Geoff Hinton, Pietro Perone, and others, who are leading the quest to unravel the mysteries of intelligence through the lenses of AI and neuroscience.

Tommaso Poggio’s Inquiry: Comparing AI with Human Intelligence

Tommaso Poggio, a prominent figure in this field, raises pivotal questions regarding the comparison of large language models and human intelligence. He examines the potential of mutual benefits between neuroscience and AI in understanding the puzzle of dimensionality in deep learning. His work is essential for comprehending why neural networks, unlike classical and quantum computers, are not hampered by the curse of dimensionality, offering insights into the common principles of intelligence.

Geoff Hinton credits neuroscience for its central role in AI’s development, noting how the field inspired concepts such as dropout and ReLU activations. He anticipates the integration of fast weights into AI, requiring specialized hardware. Hinton also suggests that digital intelligences, which utilize backpropagation and allow identical model copies, might be superior to current AI models. However, he is skeptical about AI’s ability to significantly contribute to our understanding of neuroscience. Hinton challenges the idea that the human brain is more statistically efficient than artificial neural networks, citing the efficiency of large language models in few-shot learning tasks.

Geoff Hinton and the Influence of AI on Neuroscience

Geoff Hinton presents a unique perspective on how AI influences neuroscience. He challenges traditional views about the brain’s efficiency and suggests that digital intelligences might surpass current AI models. His insights are critical in understanding how AI’s rapid developments might reshape our understanding of the brain.

Pietro Perone argues that embodiment is crucial for intelligence, and text-based AI is limited. He emphasizes the need for AI systems to understand causation and conduct experiments, highlighting AI’s advantages and limitations, such as its inability to generalize to new situations and lack of common sense.

Pietro Perone: The Embodiment of Intelligence

Pietro Perone emphasizes the significance of embodied intelligence and critiques the prevailing focus on internet data analysis. He advocates for AI systems that can interact with the real world, underscoring the importance of understanding causation and experimental design. His views highlight the limitations of current AI in comparison to amateur biologists, while acknowledging their strengths in information processing and communication.

Bridging Gaps in AI and Neuroscience

Understanding intelligence requires a collaborative effort between neuroscience and AI. The goal is to develop a theory of intelligence that explains brain function and goes beyond prediction. Neuroscience has significantly contributed to AI, especially in deep learning and reinforcement learning. The internet’s vast human-curated data serves as a valuable resource for AI. Current AI systems lack proficiency in planning, handling factuality, and episodic memory, with potential for further inspiration from neuroscience. There is a shift towards analysis techniques in understanding AI systems, where neuroscientists can contribute their skills.

The Elusive Nature of AI Theory and Neuroscience Contributions

The concept of theory in AI varies, and the complexity of neural networks makes it unlikely to have a precise theory at the level of physics. Nevertheless, theory remains useful in AI. Neuroscience has contributed significantly to AI, but determining which ideas from the brain are useful for AI inspiration remains challenging. Studying artificial neural networks may provide insights into how the human brain works, and this research avenue is promising.

Tommaso Poggio and Geoff Hinton: Theoretical Perspectives

The most successful theory for deep networks is dropout, according to Perone. He notes the challenges in theory development due to the rapidly changing AI landscape and suggests collaboration between theoreticians and experimentalists.

Neuroscience, AI, and Psychophysics

The brain’s complexity surpasses that of transformer architectures, offering opportunities for neuroscience to gain insights from simpler models. The study of language should not overshadow other forms of intelligence. Exploring simpler embodiments of intelligence in different species can provide fundamental insights. Incorporating psychoph ysics into the discussion can offer benchmarks and insights into animal behavior, reflecting responses to physical stimuli.

Challenges and Future Directions

The brain’s complexity might be a result of evolutionary optimization, and it’s unclear if such intricacy is essential for intelligence. AI models have incorporated elements of neural diversity, but the extent to which this diversity is necessary for intelligence remains uncertain. Investigating existing trained neural networks could provide insights into the relationship between neuronal diversity and intelligence. The brain likely does not use backpropagation through time, suggesting a different learning mechanism. The architecture succeeding transformers remains unknown. Understanding how the brain learns could impact machine learning, as the biological plausibility of backpropagation and gradient descent in the brain is questionable. Discovering the brain’s learning method could lead to novel AI algorithms.

In conclusion, the collaborative journey of neuroscience and AI reveals a rich tapestry of ideas and challenges. It underscores the necessity of a synergistic approach to unravel the complexities of intelligence, blending theoretical exploration with practical application. As the field progresses, it promises not just advancements in technology and understanding but also a deeper comprehension of our own place in the universe.


Notes by: oganesson