Demis Hassabis (DeepMind Co-founder) – Creativity and AI (Oct 2018)


Chapters

00:00:00 Creativity and AI: An Interdisciplinary Exploration
00:02:09 Machine Learning vs. Expert Systems: A Paradigm Shift in Artificial Intelligence
00:06:53 Learning and Generality: Key Concepts for Developing Intelligent AI Systems
00:10:48 How Reinforcement Learning Powers General Intelligence
00:15:39 Reinforcement Learning Systems for Game Mastery
00:17:50 Complexity and Beauty of the Ancient Game of Go
00:20:14 Neural Networks Enable AlphaGo to Conquer Go's Complexity
00:23:02 AlphaGo's Unconventional Strategy in the Game of Go
00:26:21 Exploring the Essence of Intuition and Creativity in the Context of AI and Go
00:30:44 Levels of Creativity in AI Systems
00:34:30 Gaps in AI Systems and Imagination as a Key to Creativity
00:38:32 AI's Cutting Edge: Unraveling Imagination and Inverse Graphics
00:43:34 Practical Applications of Artificial Intelligence
00:49:43 AI, Art, and Human Creativity
00:54:23 Neuroscience-Inspired AI: Exploring the Limits of Computational Intelligence
01:00:09 Intelligence and Consciousness: Dissociable Traits
01:02:32 AI Research and Practices: Current State and Future Directions
01:06:44 Machine Interpretability and the Evolution of AI Systems

Abstract

The Intersection of Creativity and Artificial Intelligence: Insights from Demis Hassabis and the Role of Imagination in AI

In the inaugural Rothschild Lecture at the Royal Academy, DeepMind co-founder Demis Hassabis explores the profound implications of artificial intelligence (AI) in the field of creativity and scientific discovery. This article delves into Hassabis’ perspective on how AI, especially through advancements in learning systems and algorithms like AlphaGo, is redefining our understanding of creativity, intuition, and problem-solving. With a focus on the breakthroughs in games like Go and the application of AI across various fields, Hassabis sheds light on the current capabilities and limitations of AI, its potential to solve complex problems, and the ethical considerations in its deployment.

Demis Hassabis, co-founder of DeepMind and renowned AI expert, is the distinguished speaker for the evening. Recognized worldwide as a brilliant thinker in the field of AI, he was nicknamed the superhero of AI by the Guardian. His eclectic experiences as an AI researcher, neuroscientist, and video game designer provide a unique perspective on creativity and AI. Hassabis expresses his honor in delivering the inaugural Rothschild Lecture at the Royal Academy, emphasizing the importance of fostering dialogues between the sciences and the arts, which becomes increasingly vital as we progress into the modern technological world.

AI and Creativity: A DeepMind Perspective

DeepMind’s mission, as articulated by Hassabis, is to unravel the intricacies of intelligence and harness it to address complex challenges. This endeavor has led to significant advancements in AI, particularly in learning systems that adapt and generalize beyond hard-coded expertise. A prime example is AlphaGo’s victory over Go grandmaster Lee Sedol, where the AI demonstrated creativity and strategic depth. Hassabis emphasizes the role of neural networks in this achievement, with AlphaGo employing a policy network for predicting moves and a value network for assessing positions, thus navigating the complexities of Go.

Go’s Complexity and Intuition:

Go’s intricate nature makes it challenging to define a rule-based system for AI to play. Unlike chess, Go players often rely on intuition and feel, making it difficult to articulate their move choices. AlphaGo tackles this by utilizing learning systems, including two neural networks. The policy network analyzes the current board position and identifies the most promising moves, significantly reducing the vast search space. Meanwhile, the value network evaluates the current board position and predicts the probability of each player winning, providing a numerical assessment of the game’s state and helping AlphaGo make informed decisions. By combining these networks, AlphaGo solves the challenges presented by Go, leveraging both intuition (learned from millions of games) and calculation (evaluating board positions) to make strategic moves.

Interpolation, Extrapolation, and Invention in Machine Creativity:

Machine learning systems like AlphaGo excel at interpolation, finding commonalities among existing examples. They are also capable of extrapolation, generating strategies beyond human anticipation. However, true invention, creating entirely new concepts, remains a challenge for AI.

The Evolution of AI Learning Systems

Earlier AI systems like Deep Blue, which defeated chess champion Garry Kasparov, lacked learning capabilities and adaptability. In contrast, modern AI systems use reinforcement learning, a trial-and-error approach where agents interact with their environment and build knowledge from scratch. This approach, inspired by the human brain’s dopamine system, allows AI to develop novel strategies, as seen in the AI’s unexpected tactics in the Atari game Breakout.

An Agent System:

Reinforcement learning systems can be viewed as agents consisting of two parts: a model of the world and an action selector. The model updates continually based on new observations, while the action selector chooses the best action to achieve the goal. The reinforcement learning cycle involves the agent running out of thinking time, executing the best action found, updating its world model, and selecting a new action. This cycle continues until the goal is reached. The dopamine system in the primate and human brain exemplifies a biological implementation of reinforcement learning.

From Games to Real-World Applications

Games have served as crucial testing grounds for AI, allowing for the development of algorithms in controlled, virtual environments. This approach has led to AI systems like the DQN agent, which learns from raw pixel input, as demonstrated in Atari Breakout. Beyond games, these AI technologies find applications in various fields, including healthcare, education, and scientific research, addressing challenges like exoplanet discovery and medical diagnostics.

AI has a wide range of commercial applications, including healthcare, medical diagnostics, optimization, energy, education, and virtual assistants. It has also been used in art and design, such as architecture, car engine design, and art transfer. In the realm of scientific research, AI’s contributions include discovering exoplanets, controlling plasma in nuclear fusion reactors, designing chemical compounds, and detecting eye diseases. Games offer a diverse range of challenges and allow for rapid experimentation, making them ideal platforms for developing and testing AI algorithms. Virtual simulations in games are more convenient for AI development compared to robotics.

The DQN agent system, tested on Atari games in 2013 with raw pixels as its only input, successfully learned to play Breakout, improving over time as it gained experience. The goal of reinforcement learning is to develop a single system that can play all different games out of the box. The DQN agent system demonstrated this potential by learning to play Breakout, a seminal game on the Atari system.

Creativity and Intuition in AI

Hassabis discusses the essence of creativity and intuition in AI, drawing a distinction between rule-based creativity and genuine invention. While current AI systems excel at tasks involving statistical averaging and pattern recognition, they struggle with tasks requiring abstract thinking and imagination. Hassabis points to the hippocampus’s role in imagination, suggesting that understanding human brain mechanisms can inspire more creative AI systems.

In Go, originality and creativity are judged by the effectiveness of a move, which is determined by the outcome of the game. AlphaGo’s move 37 in its match with Lee Sedol was initially considered a mistake by expert commentators, but later analysis revealed it to be an effective and creative move. AlphaGo’s approach to the game was distinct from traditional Go strategies used by professionals. In game two, AlphaGo made a surprising move (move 37) by placing a stone on the fifth line early in the game, which is unconventional and considered suboptimal. This move, seemingly wasteful, proved to be decisive in AlphaGo’s victory. 100 moves later, two stones placed by AlphaGo in the bottom left corner interacted with the stone placed in move 37, leading to a strategic advantage. AlphaGo’s foresight and ability to plan moves 100 steps ahead were remarkable and contributed to its overall win.

Modern AI systems lack abstract thinking, memory systems, and imagination. Brain inspiration, especially from the hippocampus, is crucial for imagination research in AI. Memory is not a perfect videotape but rather a reconstruction of components. Imagination, as a constructive process, may rely on similar brain mechanisms. The hippocampus is crucial for memory; its damage causes amnesia. Imagination tasks were tested on patients with damaged hippocampus, and their descriptions of imagined scenarios were significantly impoverished. A graph shows the difference in richness between patients and control subjects.

Challenges in AI Development

Despite its advancements, AI faces several challenges. The need for explainability and interpretability in AI systems is critical for their safe and ethical application. Hassabis advocates for human involvement in AI-driven decision-making processes. Additionally, the speed of AI research has led to an over-promising hype cycle, necessitating a more measured approach to development and empirical testing.

Artificial Intelligence: A Meta Solution for the Modern World

AI, as Hassabis outlines, offers solutions to the complexities and information overload of modern systems. It transforms unstructured data into actionable knowledge, aiding in the understanding of intricate systems. However, this potential comes with responsibilities, emphasizing the need for safe AI use and ensuring its benefits are widely accessible.

AI can help address the challenges of information overload and system complexity in various domains. It can process vast amounts of data and extract meaningful insights, aiding in decision-making. AI can be used as a powerful tool to accelerate scientific breakthroughs and assist experts in their work. It can help scientists understand complex systems, discover new knowledge, and develop innovative solutions. AI is a neutral technology that can be used for good or bad, depending on how it is deployed. It is important to ensure that AI is used responsibly, safely, and for the benefit of everyone. Open dialogue and collaboration among scientists, technologists, artists, and social scientists are crucial to shape the ethical development and use of AI.

AI, Human Creativity, and the Future

In conclusion, Demis Hassabis’ lecture encapsulates the dynamic relationship between AI and human creativity. While AI has made leaps in areas like game-playing and problem-solving, it still faces limitations in truly mimicking human creativity and consciousness. The future of AI, as envisioned by Hassabis, is not just about technological advancement but also involves a deep understanding of human intelligence and creativity, ensuring that AI serves as a complement to human endeavor rather than a replacement.

Lee Sedol’s move 78 in game four against AlphaGo, known as the “wedge move,” was a brilliant move that allowed him to win the game. AlphaGo’s networks misjudged this move, leading to its defeat. Greg Coase’s award-winning documentary provides insights into the human emotions and spirit of endeavor behind the AlphaGo-Lee Sedol match. After the match, Lee Sedol acknowledged AlphaGo’s creativity, recognizing that move 37 was a beautiful and creative move, beyond mere probability calculation.

Demis Hassabis defines intuition as implicit knowledge acquired through experience that is not consciously accessible or expressible. The quality of intuition can be tested behaviorally, such as by evaluating the quality of a move in a game like Go. By studying AI, scientists hope to gain a better understanding of the human mind and brain. Comparing AI algorithms to the human brain can shed light on unique aspects of human intelligence, such as creativity and consciousness.

AI holds immense promise for solving societal challenges and accelerating scientific progress. The responsible and ethical use of AI is essential to ensure its benefits are accessible to all. Studying AI can help deepen our understanding of the human mind and consciousness.

Supplemental Update:

Demis Hassabis believes intelligence is substrate-independent and can be developed through reinforcement learning, not just neuroscience. He distinguishes his approach from “whole brain emulation,” which seeks to precisely reverse engineer the brain’s biological structure. Hassabis proposes that intelligence and consciousness are double dissociable, meaning they can exist independently of each other. He suggests intelligent systems can be developed without consciousness and that certain animals possess consciousness despite limited intelligence.

Hassabis acknowledges the possibility that consciousness may eventually become a limiting factor in developing more intelligent systems. He suggests that understanding consciousness might be necessary to overcome future barriers in AI development. The presentation concludes with a question-and-answer session, where Hassabis engages with audience members to discuss various aspects of consciousness and intelligence. The AI research community is largely open and collegiate, promoting the sharing of research and collaboration. To establish best practices for deploying AI systems, further development and empirical testing are necessary.

Generative models, like GQN, can fill in missing parts of pictures or even draw photos, though they are not yet photorealistic. These systems currently lack the understanding of scene semantics and physics, limiting their ability to model complex scenes. With advancements in abstraction and concepts, AI systems could eventually Parcellate the world into semantic meaning and structure, enabling them to model more complicated scenes. Art has historically harnessed technology, as seen in the creation of paintings using algorithms.

The importance of interpretability in AI systems for scientific advancement and safe deployment in safety-critical applications is crucial. The need for humans to have the final decision-making authority, with AI acting as a tool for information provision, is emphasized. The current phase of AI development is focused on functionality rather than interpretability. Ongoing research and development of analysis, visualization, and behavioral testing tools to enhance interpretability are underway. The comparison of AI analysis to neuroscience, studying the behavior and architecture of artificial brains, is ongoing.

With further advancements in interpretability tools, the current black box AI systems will become understandable and interpretable. The starting point of interpretability research and the need for patience during this developmental phase are acknowledged. The art world’s tendency to depict dystopian futures and villains when exploring AI is critiqued, with a call for more creative and diverse representations of AI’s potential in art. The inspiration that science fiction provides to scientists, including Demis Hassabis, is noted. Recommendations for exploring a broader spectrum of possibilities with AI rather than narrow and crude depictions are made. Demis Hassabis expresses his personal opinion that Westworld is boring and obvious in its exploration of AI.

The need for the art world to expand its networks and engage with scientists, including those at the forefront of AI research, is highlighted. The potential of the Royal Academy as a forum for debating human consciousness and AI-related topics is recognized.


Notes by: oganesson