Speaker Introduces Demis Hassabis:
Demis Hassabis, co-founder of DeepMind and renowned AI expert, is the distinguished speaker for the evening. He is recognized worldwide as a brilliant thinker in the field of AI and 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.

Topic: Creativity and AI:
Hassabis’ lecture explores the implications of cutting-edge AI research for creativity and scientific discovery. He draws on his diverse experiences to discuss this topic.

Demis Hassabis Begins His Lecture:
Hassabis expresses his honor in delivering the inaugural Rothschild Lecture at the Royal Academy. He emphasizes the importance of fostering dialogues between the sciences and the arts. This becomes increasingly vital as we progress into the modern technological world.

Theme: Creativity Explored:
The central theme of Hassabis’ lecture is creativity, viewed through the lens of science and recent advances in artificial intelligence. He defines AI as the science of making machines smart.

DeepMind’s Mission:
To understand intelligence and recreate it artificially. Once step one is achieved, the potential to solve almost everything else naturally follows. Plans to accomplish this by building the world’s first general purpose learning system.

Expert Systems (Traditional AI):
Hand-coded knowledge in the form of rules. Can’t handle the unexpected and are brittle. Limited to solutions that programmers can express. Inspired by logic systems and logic theory.

Learning Systems (Modern AI):
Learn solutions from first principles and directly from data and experience. Can generalize to new tasks and potentially solve problems we don’t know how to. Inspired by neuroscience and how the brain works.

Deep Blue vs. Garry Kasparov:
Deep Blue, an expert system, defeated Garry Kasparov, the world chess champion, in the late 1990s. Kasparov’s creativity and ability to play other games and perform various tasks made him more impressive. Deep Blue’s hard-coded specialization limited it to chess and prevented it from playing simpler games.

Reinforcement Learning:
Reinforcement learning systems learn from first principles through trial and error. Agents in reinforcement learning systems interact with the environment through sensory apparatus and receive observations and rewards. Agents build statistical models of the environment to understand linkages and make decisions. Reinforcement learning focuses on learning general skills that can be applied to various tasks. DeepMind’s AlphaGo defeated Lee Sedol, the world’s top Go player, in 2016. AlphaGo was able to learn and adapt to new games with minimal data, demonstrating its generality.

The Agent System:
An agent system consists of two parts: a model of the world and an action selector. The model of the world continually updates based on new observations. The action selector chooses the best action to take to achieve the goal.

The Reinforcement Learning Cycle:
The agent runs out of thinking time and outputs the best action found so far. The action gets executed and may drive a new change in the environment. The agent updates its model of the world and selects a new action. This cycle continues until the agent reaches its goal.

Challenges of Reinforcement Learning:
There are many technical challenges that need to be solved to implement reinforcement learning effectively. The dopamine system in the primate and human brain implements a form of reinforcement learning.

Games as a Proving Ground for AI Algorithms:
Games are an ideal platform for developing and testing AI algorithms. They offer a diverse range of challenges and allow for rapid experimentation. Virtual simulations in games are more convenient than robotics for AI development.

Atari 2600 as a Testbed for Reinforcement Learning:
The Atari 2600 was the first console with a large diversity of games. The DQN agent system was tested on Atari games in 2013. The DQN agent only receives raw pixels as input and learns everything from scratch.

Generality of Reinforcement Learning:
The goal is to develop a single system that can play all different games out of the box. The DQN agent system successfully learned to play Breakout, a seminal game on the Atari system. The system improved over time as it gained more experience playing the game.

Optimal Breakout Strategy:
After 300 games of Breakout, the AI discovered a hidden optimal strategy: digging a tunnel on the left side and sending the ball behind the brick wall. This tactic is low-risk and highly rewarding, allowing the AI to hit multiple bricks with one shot. The discovery of this strategy was a significant milestone for DeepMind, as it showed that their AI could learn and develop strategies beyond what was explicitly programmed.

AlphaGo:
AlphaGo is a program that uses reinforcement learning to play the ancient game of Go. Go is a complex and strategic game played on a 19×19 grid, with players taking turns placing black and white stones on the board. Go has a long history and is considered a form of art in Asia, with Confucius including it among the four great arts for scholars.

Significance of AlphaGo:
AlphaGo’s victory over world champion Lee Sedol in 2016 was a major breakthrough in the field of artificial intelligence. It demonstrated that AI could master complex games that require strategic thinking, intuition, and creativity. AlphaGo’s success opened up new possibilities for AI research and applications in various domains.

Go as an Art Form:
Go is considered a profound art form, with 40 million active players and 2,000 professionals. The game’s simplicity, with only two rules, contrasts with its immense complexity, resulting in 10 to the power 170 possible ball positions. This complexity and esoteric nature make Go a challenging game for computers to play.

Rules of Go:
Go is played by placing stones on a board one at a time until the board is filled. The goal is to wall off empty areas of territory and count the number of squares surrounded compared to the opponent. The person who surrounds the most squares wins the game.

Challenges for Computers in Go:
The enormous number of possible positions (10 to the power 170) makes Go a much harder game for computers compared to chess. Unlike chess, Go lacks a handcrafted evaluation function that allows computers to determine the winning side in the current position. This makes it difficult for computers to evaluate the strength of different moves and make optimal decisions during gameplay.

Go’s Complexity and Intuition:
Go’s complexity and esoteric nature make 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’s Learning Systems:
AlphaGo employs learning systems, including two neural networks, to address Go’s complexity and intuitive nature.

Policy Network:
The policy network analyzes the current board position and identifies the most likely moves. It reduces the vast search space by focusing on the most promising moves.

Value Network:
The value network evaluates the current board position and predicts the probability of each player winning. It provides a numerical assessment of the game’s state and helps AlphaGo make informed decisions.

Combining the Neural Networks:
AlphaGo combines the policy and value networks to solve the challenges presented by Go. The system leverages both intuition (learned from millions of games) and calculation (evaluating board positions) to make strategic moves.

AlphaGo’s Victory Over Lee Sedol:
AlphaGo, a computer program, defeated Lee Sedol, one of the greatest Go players in history, in a million-dollar challenge match in 2016. This was a significant achievement as no Go program had ever defeated a professional player before, let alone a world champion. The match’s outcome was proclaimed a decade ahead of its time and marked a turning point in AI development.

AlphaGo’s Unique Strategy:
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.

Move 37’s Significance:
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.

Gameplay and Creativity:
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.

Lee Sedol’s Move 78:
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 Documentary:
Greg Coase’s award-winning documentary provides insights into the human emotions and spirit of endeavor behind the AlphaGo-Lee Sedol match.

Lee Sedol’s Perspective on AlphaGo’s Creativity:
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’ Definition of Intuition:
Intuition is 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.

Interpolation:
Interpolation is a type of creativity where a new idea is created by averaging or finding a common ground among existing examples. Machine learning systems excel at interpolation due to their ability to analyze data and identify patterns.

Extrapolation:
Extrapolation is a higher level of creativity where new ideas are created by extending the boundaries of existing knowledge or examples. AlphaGo-like systems are capable of extrapolation, generating new strategies and moves beyond what human designers anticipated.

Invention or Innovation:
Invention or innovation is the highest level of creativity, where entirely new concepts or ideas are created, often outside the boundaries of existing knowledge. No AI systems have yet achieved true invention, such as creating a completely new game like Go or chess.

Machine Creativity:
Neural network systems are proficient at interpolation due to their statistical nature and ability to identify patterns in data. AlphaGo-like systems are advancing in extrapolation, demonstrating the ability to generate novel ideas beyond human knowledge. True invention remains a challenge for AI systems, as they cannot yet create entirely new concepts or ideas outside of existing knowledge.

Introduction:
Modern AI systems lack abstract thinking, memory systems, and imagination. Cutting-edge AI research focuses on these capabilities for out-of-the-box thinking.

Brain Inspiration:
Brain systems neuroscience offers insights into imagination’s mechanisms. Imagination is a constructive process that may share mechanisms with memory.

Memory as a Reconstructive Process:
Memory is not a perfect videotape but rather a reconstruction of components. Imagination, as a constructive process, may rely on similar brain mechanisms.

Hippocampus and Imagination:
Hippocampus is crucial for memory; its damage causes amnesia. Imagination tasks were tested on patients with damaged hippocampus.

Imagination Impoverishment:
Patients’ descriptions of imagined scenarios were significantly impoverished. Graph shows the difference in richness between patients and control subjects.

Conclusion:
Brain research, particularly on imagination, inspires AI advancements in creativity.

Spatial Coherence and the Hippocampus:
The hippocampus is crucial for imagination, as it binds together disparate elements of a scene into a coherent whole, enabling humans to predict the future and imagine counterfactual situations.

Five Brain Areas Involved in Imagination:
Brain scanning studies revealed five distinct brain areas heavily involved in different aspects of imagination.

Generative Query Network (GQN):
GQN is an AI system that can reconstruct a 3D model of a scene from a handful of 2D snapshots.

Inverse Graphics Problem:
GQN solves the inverse graphics problem by recovering the generative equations that create a 3D scene from 2D images.

Simple Scenes Reconstruction:
Currently, GQN can reconstruct simple scenes with a few geometric objects of different colors and textures.

Generating New Views:
GQN can generate new views of a scene from any arbitrary angle, matching the ground truth almost perfectly.

Future Applications:
GQN has the potential to reconstruct real-world scenes from 2D images, opening up new possibilities for computer graphics and AI.

Conclusion:
Despite the challenges that remain, AI technology, such as GQN, is already proving useful in various applications.

Applications of AI:
AI has a wide range of commercial applications, including healthcare, medical diagnostics, optimization, energy, education, and virtual assistants. AI has been used in art and design, such as architecture, car engine design, and art transfer. AI has also been successfully applied in scientific research, including discovering exoplanets, controlling plasma in nuclear fusion reactors, designing chemical compounds, and detecting eye diseases.

AI as a Solution to Information Overload and System Complexity:
AI can help address the challenges of information overload and system complexity in various domains. AI can process vast amounts of data and extract meaningful insights, aiding in decision-making.

AI as a Tool for Scientific Breakthroughs:
AI can be used as a powerful tool to accelerate scientific breakthroughs and assist experts in their work. AI can help scientists understand complex systems, discover new knowledge, and develop innovative solutions.

Responsible and Ethical Use of AI:
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.

Understanding the Human Mind through AI:
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.

Conclusion:
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.

Feynman’s Perspective on Beauty:
Feynman appreciated the beauty of a flower not only aesthetically but also scientifically, considering the intricate cellular processes and evolutionary aspects that contribute to its existence.

Competitiveness in AI and Human Creativity:
Competitiveness, when used positively, can drive progress in developing relationships and knowledge between AI and human creativity. AlphaGo, while an AI program, is a product of human endeavor, involving skilled programmers and researchers.

AlphaGo’s Impact on the Go World:
AlphaGo’s victory over Lee Sedol was initially seen as a symbolic loss for human intelligence. However, AlphaGo has also inspired Go players to unleash their creativity and explore new strategies, freeing them from traditional constraints.

The Role of Randomness and Chance in Art:
The art world often embraces randomness and chance as elements that can foster creativity and lead to unexpected outcomes. Beckett’s idea of “failing or failing better” resonates with many artists, who see the pursuit of perfection as an impossible but motivating goal.

The Relationship Between Art and Science:
Art and science are not mutually exclusive but can coexist and complement each other. Both disciplines involve creativity, exploration, and the pursuit of knowledge, albeit through different methods and perspectives.

Creativity:
Hassabis distinguishes between two types of creativity: Extrapolation: Being creative within existing rules and boundaries. Invention: Breaking the rules and coming up with something truly new. Current AI systems are capable of extrapolation but not invention.

Art Criticism:
AI systems could potentially learn to judge art based on technical aspects and training data. However, human art appreciation goes beyond technicalities and includes the imprint of the artist’s soul, which is difficult for AI to replicate.

Machine Creativity:
Machine creativity may be able to simulate aspects of human creativity in the coming decades. However, it is unclear if machines can fully replicate the human experience and creativity, which involves wrestling with the human condition.

Consciousness and Quantum Effects:
Hassabis is open-minded about the possibility that some aspects of human intelligence, such as consciousness, may not be computable. There is speculation that quantum effects in the brain may play a role in consciousness, but no evidence of this has been found yet.

Reinforcement Learning and General Intelligence:
Hassabis believes that reinforcement learning can lead to general intelligence. However, it is unclear if reinforcement learning alone is sufficient to achieve the highest levels of creativity and invention.

Substrate Independence of Intelligence:
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.

Double Dissociation of Intelligence and Consciousness:
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.

Consciousness as a Potential Bottleneck:
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.

Discussion and Questions:
The presentation concludes with a question-and-answer session. Hassabis engages with audience members to discuss various aspects of consciousness and intelligence.

AI Research and Ethical Practices:
The speed of AI research has generally had positive impacts, but hype has led to over-promising and some rushed or flawed science. 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.

The Role of AI in Art and Architecture:
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 and Technology:
Art has historically harnessed technology, as seen in the creation of paintings using algorithms.

Interpretability in AI Systems:
Importance of interpretability in AI systems for scientific advancement and safe deployment in safety-critical applications. The need for humans to have the final decision-making authority, with AI acting as a tool for information provision.

Challenges and Progress in Interpretability:
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. The comparison of AI analysis to neuroscience, studying the behavior and architecture of artificial brains.

Overcoming Black Box Systems:
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.

Art World’s Representation of AI:
Critique of the art world’s tendency to depict dystopian futures and villains when exploring AI. The need for more creative and diverse representations of AI’s potential in art.

Science Fiction’s Role in AI Inspiration:
The inspiration that science fiction provides to scientists, including Demis Hassabis. Recommendations for exploring a broader spectrum of possibilities with AI rather than narrow and crude depictions.

Demis Hassabis’s Dislike for Westworld:
Personal opinion that Westworld is boring and obvious in its exploration of AI.

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