AI, the Brain, and Our Future | Dr.Beren Millidge | TEDxMiami

Key Concepts:

• Artificial General Intelligence (AGI)
• Unsupervised Predictive Learning
• Scaling Laws
• Reinforcement Learning (RL)
• Long-Term Memory
• Continual Learning
• Cortex and Subcortex (Brain Structure)
• Basal Ganglia and Hippocampus (Brain Regions)
• AI Alignment

1. Introduction and Overview

• Dr. Baron Millage discusses the potential for humanity to create artificial minds that surpass human intelligence in the coming decades.
• He frames AI as both a potentially beneficial and dangerous invention, capable of solving fundamental questions about intelligence and the human condition.
• Dr. Millage's background is in machine learning and neuroscience, focusing on understanding intelligence in both brains and machines. He is the co-founder and chief scientist of Zeifer, an AI startup.
• He quotes Richard Feynman: "What we cannot create, we do not understand," emphasizing the link between AI development and understanding the human mind.
• The presentation aims to connect neuroscience knowledge with recent AI advancements, focusing on underlying principles and future steps toward AGI.

2. Introductory Neuroscience: Brain Structure

• The brain is divided into two main parts: the cortex and the subcortex.
• Cortex: The outer layer, responsible for primary intelligence, divided into specialized areas:
  • Visual Cortex: Processes visual information.
  • Somatosensory and Motor Cortices: Handle touch, sensation, and movement.
  • Prefrontal Cortex: Handles executive functions, reasoning, logic, and planning.
• Subcortex: Located in the middle and bottom of the brain, more heterogeneous, comprised of specialized modules for specific tasks.
  • Includes functions like maintaining homeostasis and triggering fight-or-flight responses.
  • Basal Ganglia: Core of reward function, innate drives, and decision-making.
  • Hippocampus: Location of memory formation, consolidation, and recall.

3. Core Principles Driving AI Advances

• Unsupervised Predictive Learning:
  • The primary method for training AI systems.
  • Involves predicting the next element in a sequence of data based on previous elements.
  • Example: Predicting the next word in a sentence.
  • The power lies in the implicit information within the data, eliminating the need for explicit labeling.
  • The brain uses a similar strategy in sensory cortices, predicting the next sensation and minimizing prediction errors.
• Scaling:
  • "Bigger is better" in AI.
  • Scaling laws relate the size of a neural network (number of neurons) and the amount of training data to performance increases.
  • Similar scaling behavior is observed in the natural world, relating neuron count and density to intelligence across species.
  • This suggests a fundamental aspect of intelligence.
• AI's current success is driven by performing unsupervised predictive learning at a large scale, creating general correlation learning machines.
• From a brain perspective, AI has learned to build a sensory cortex.

4. Remaining Steps on the Road to AGI

• The remaining steps can be understood by examining the remaining areas of the brain and their functions.
• Reinforcement Learning (RL):
  • Learning by directly interacting with the world.
  • Current AI models are like "brains in a vat," receiving passive input without interaction.
  • RL allows AI to learn novel strategies beyond human-derived data.
  • Performing RL on top of powerful sensory representations developed through unsupervised learning improves stability and efficiency.
• Long-Term Memory:
  • Current AI has short-term memory (context window).
  • Strategies like summarizing context are used, but are fundamentally flawed.
  • AI needs an artificial analog of the hippocampus for memory formation, consolidation, and reasoning.
  • Damage to the hippocampus in humans results in amnesia, highlighting its importance.
• Continual Learning:
  • Current AI operates in distinct training and inference phases.
  • Humans learn online continuously.
  • Lack of continual learning causes AI to drift out of date and hinders long-term tasks.
  • The brain has mechanisms for integrating new information while avoiding forgetting old information, such as gating learning by salience and replaying experiences during sleep.

5. Implications and Concerns

• Combining these advances would result in an AI system that can learn from the environment, improve indefinitely, and possess high-capacity long-term memory.
• The prospect of creating systems smarter and more powerful than humans is both exciting and worrisome.
• Intelligence is what separates humans from other animals.
• Creating AI could lead to a new kind of being that competes with or outcompetes humans.
• Even with aligned AI, there is a risk of relinquishing agency and control.
• Dr. Millage believes that greater understanding and study of AI systems is beneficial.
• He advocates for creating AI that is not only smarter but also more moral, compassionate, and selfless than humans.
• He emphasizes the importance of sharing the gains from AI broadly.

6. Conclusion

• Dr. Millage expresses excitement about the progress in AI and the growing understanding of the brain.
• He anticipates that within the next few decades, we may resolve age-old questions about our minds, brains, and ourselves.