Background on Demis Hassabis:
Demis Hassabis is the co-founder and CEO of DeepMind, a neuroscience-inspired AI company acquired by Google in 2014. He is a former child chess and coding prodigy who founded the pioneering video games company Alexia Studios. After a successful career in technology startups, he returned to academia to complete a PhD in Cognitive Neuroscience at UCL, followed by postdocs at MIT and Harvard. His research on connecting memory with imagination was recognized as one of the top 10 scientific breakthroughs of 2007 by the journal Science. He is a five-time world games champion, recipient of the Royal Society’s Mullard Award, and a fellow of the RSA.

DeepMind as an Apollo Program for AI:
DeepMind was founded in 2010 and joined forces with Google in 2014. Hassabis describes DeepMind’s mission as an Apollo program effort for AI, aiming to accelerate progress in the field by bringing together top scientists and engineers in an ideal environment. The goal is to make rapid progress in AI research and development.

Focus on AI’s Potential to Solve Global Challenges:
Hassabis’s talk will focus on AI’s potential to address some of the world’s greatest global challenges, such as healthcare and climate change. He believes that AI can be harnessed to benefit all humanity, not just an elite.

Structure of the Presentation:
Hassabis will speak for 20-25 minutes, followed by a conversation with Matthew Taylor. The audience will then be invited to ask questions.

Company Overview:
DeepMind is a company focused on researching and developing artificial intelligence (AI). It employs over 350 people, including 250 research scientists, making it a significant force in AI research. The company aims to combine the best aspects of Silicon Valley startups and academic institutions to create a hybrid organizational structure that fosters innovation and creativity.

Mission and Goals:
DeepMind’s mission is to fundamentally solve intelligence and use that technology to solve practical problems and benefit society. This goal may seem ambitious, but the company believes it is plausible and aims to convince skeptics through its research and developments.

Dimensions of AI:
DeepMind considers four dimensions when thinking about AI: Learning systems vs. handcrafted heuristic systems: AI systems that learn from experience or raw data are preferred over those pre-programmed with specific solutions. Generality vs. special casing: The company focuses on developing general-purpose AI systems capable of performing a wide range of tasks, rather than narrow AI systems designed for specific purposes. Groundedness vs. logic-based: DeepMind believes AI systems should be grounded in sensory and motor experiences to interact effectively with the real world. Active learning vs. passive observation: Active learning systems that actively explore and direct their learning are preferred over passive systems that simply observe data.

DeepMind’s Approach to AI:
DeepMind is committed to the left-hand side of all four dimensions mentioned above. The company aims to build general-purpose learning systems that learn how to perform tasks rather than being given solutions by programmers. This approach is often referred to as artificial general intelligence (AGI) and is distinct from narrow AI systems used for specific tasks.

Deep Blue and Narrow AI:
IBM’s Deep Blue, which defeated Garry Kasparov in chess in the late 1990s, is an example of a narrow AI system. Deep Blue was specifically designed and programmed for chess and was not capable of performing other tasks. This limitation highlights the shortcomings of narrow AI systems and motivates DeepMind’s pursuit of general-purpose AI.

Deep Reinforcement Learning:
DeepMind pioneered deep reinforcement learning, a technique that combines deep learning and reinforcement learning. Deep learning is used to perceive the world around the AI system, while reinforcement learning is used to select actions that best achieve the system’s goals. DeepMind has used computer games as a training platform to develop and test its deep reinforcement learning algorithms.

Harnessing Virtual Environments for Efficient AI Development:
Virtual environments and games offer a more efficient testing ground for AI systems compared to real-world robotics. Faster iteration and quicker feedback enable rapid algorithm evaluation. Virtual environments allow for millions of simulations and training runs in the cloud.

Measuring Progress in Long-Term Missions:
Games provide a clear metric of progress through scores or proxies. This helps assess the effectiveness of algorithm changes in the short term. It’s crucial to know whether the AI is heading in the right direction early on.

Utilizing Atari 2600 Games as a Proving Ground:
Classic Atari 2600 games, such as Space Invaders, were chosen as initial testing grounds. The raw pixels on the screen serve as the input for the AI agents. The goal is to maximize the score, while everything else is learned from first principles.

Generalization across Different Games:
A single learning system aimed to master multiple games with diverse objectives and visuals. This demonstrates the generality and adaptability of the AI algorithms.

Breakout Example:
DQN, a reinforcement learning algorithm, was tested on the Breakout game. Over time, the AI agent improved its performance by learning to control the bat and prevent the ball from passing. After extensive training, the agent discovered an unexpected strategy of digging a tunnel and sending the ball around the back. This strategy, unknown to the programmers, showcases the AI’s ability to learn complex strategies on its own.

The Power of General Learning Systems:
General learning systems can master complex tasks that even the programmers may not know how to explicitly code. These systems have the potential to make discoveries and learn strategies that humans may not have considered.

Introduction to AlphaGo and Go:
AlphaGo is a program created by DeepMind to play the ancient game of Go. Go is a two-player game played on a 19×19 board, where players place stones on vertices to capture territory. The objective is to surround your opponent’s stones or empty spaces to claim them as territory.

History and Cultural Significance of Go:
Go has a long and storied history, originating in China over 3,000 years ago. In ancient China, Go was considered an art form and a way to embody the mysteries of the universe. The game has a strong cultural significance in East Asian countries, with professional Go schools and a large number of active players.

Complexity of Go and Challenges for Computers:
Go’s simple rules lead to profound complexity, with more possible board configurations than atoms in the universe. Brute force search methods used to master chess are not effective for Go due to the vast search space and difficulty in evaluating positions. Go has an order of magnitude larger branching factor and lacks a clear evaluation function like chess.

Two Main Challenges for Computers in Go:
The vast search space of possible moves in Go, with an average of 200 possible moves per position. The difficulty in writing an evaluation function to determine who is winning in a particular position.

Challenges of Mastering Go:
Vast Search Space: In Go, the number of possible moves is significantly larger compared to chess. Complex Evaluations: Unlike chess, Go doesn’t have the concept of materiality, making it difficult to evaluate positions accurately. Intuitive Nature: Go requires a high level of intuition and instinct, which is challenging for computers.

Solutions Implemented by AlphaGo:
Policy Network: Trained to predict the most probable next move based on data from 100,000 human games. Value Network: Developed through self-play against itself to learn to evaluate positions accurately.

Impact of AlphaGo’s Victory:
Historical Significance: AlphaGo’s victory over Lee Sedol in 2016 was a major milestone in artificial intelligence, occurring a decade earlier than expected. Cultural Impact: The match generated widespread interest and coverage, boosting the popularity of Go, especially in the West. Scientific Breakthroughs: AlphaGo’s success was based on significant advances in machine learning and neural networks, with a front cover article published in Nature detailing the scientific innovations.

AlphaGo’s Influence on Go Strategy:
Strategic Innovation: AlphaGo introduced novel strategies and moves, challenging traditional wisdom and leading to a shift in Go strategy. Historical Significance: Some of AlphaGo’s moves, such as move 37 in game two, are considered to be historically significant and are studied by Go enthusiasts. Influence vs. Territory Trade-Off: AlphaGo’s strategies highlighted the delicate balance between influence in the center of the board and territory on the sides, prompting discussions on the relative value of each.

Unthinkable Move:
AlphaGo made an unexpected move on the fifth line, which was considered unthinkable by human Go players. This move gave away territory but demonstrated the potential value of influence in the center of the board.

Long-Term Impact:
Fifty moves later, the stone played on the fifth line affected a fight in another part of the board. AlphaGo’s foresight and strategic planning were evident in this move.

Reactions from Professionals:
Michael Redmond, a 9 Dan professional Go player, was initially unsure about the move and thought it was a mistake. Lisa Dole, another expert Go player, was so shocked by the move that he left the match room for 15 minutes.

Influence on Human Players:
Human players have started incorporating similar strategies, including playing on the fifth line and using the motifs introduced by AlphaGo.

Significance:
AlphaGo’s move 37 challenged traditional Go wisdom and demonstrated the potential for new strategies. It has influenced the way human players approach the game, leading to new developments and a deeper understanding of Go.

Intuition in Go:
Intuition refers to implicit knowledge gained through experience that cannot be consciously expressed or accessed. Intuition can be tested behaviorally by probing, such as by evaluating the quality of moves made by Go players.

Lisa Doll’s Experience with AlphaGo:
Lisa Doll, a professional Go player, had several dinners with Demis Hassabis, the CEO of DeepMind, during and after the AlphaGo match. Doll told Hassabis that the AlphaGo match was one of the greatest experiences of his life and renewed his passion for the game. Doll felt that AlphaGo made him realize there were still unexplored areas in Go, inspiring him to continue exploring the game. After the match, Doll’s win record improved, and he won additional titles against top human opponents.

Defining Intuition and Creativity in Go:
Intuition in Go is the implicit knowledge acquired through experience that cannot be consciously expressed or accessed. Creativity in Go is the ability to generate novel and surprising moves that deviate from established patterns and strategies.

Synthesizing Knowledge and Creativity:
AlphaGo demonstrated the ability to synthesize knowledge and create original ideas within the constrained domain of the board game Go.

Systems Neuroscience Inspiration:
DeepMind draws inspiration from systems neuroscience to understand the algorithms, representations, and architectures the brain uses. The focus is on the capabilities the brain has rather than the low-level biological details.

Systems Neuroscience Inspired AI:
DeepMind is actively researching areas inspired by how the brain accomplishes various tasks, including memory, attention, concepts, planning, navigation, and imagination.

Hippocampus and Artificial Intelligence:
The hippocampus is a brain region involved in memory and other capabilities, and DeepMind is working on creating an artificial hippocampus. The goal is to link this artificial hippocampus with neural networks developed for tasks like Atari games.

Real-World Applications:
DeepMind’s research on games is a means to develop general-purpose learning algorithms. The ultimate goal is to transfer these algorithms to the real world and apply them to high-impact problems.

AI Beyond Entertainment:
Demis Hassabis, an expert in AI, envisions the use of AI not just for entertainment purposes but also for addressing societal issues.

AI for Energy Efficiency:
Google’s data centers consume a significant amount of power, leading to high energy costs and environmental concerns. DeepMind, a company co-founded by Hassabis, applied AI to optimize the cooling systems in Google’s data centers. The AI system achieved a 15% reduction in the data center’s power usage, saving millions of dollars and reducing the carbon footprint. Hassabis suggests that similar AI systems can be used to optimize energy grids at a national scale, potentially saving 10% of a country’s energy consumption.

AI for Healthcare:
Hassabis highlights the potential of AI in healthcare, particularly in collaboration with the NHS in the UK. AI can assist in diagnosing diseases like head and neck cancers through image recognition and retinal scans for macular degeneration.

AI for Text-to-Speech:
DeepMind’s WaveNet, a text-to-speech system, has achieved state-of-the-art performance. WaveNet generates raw audio waveforms directly, resulting in more natural and less robotic speech compared to traditional concatenative systems. User tests showed an 86% preference for WaveNet’s speech over concatenative systems.

AI for Music Composition:
WaveNet can also be used to generate music compositions. Early results show that WaveNet can create free-form piano music with musically interesting sounds. Future improvements aim to enable WaveNet to stick to a single style for longer durations and create more sophisticated compositions.

AI as a Meta Solution:
Hassabis believes that AI can serve as a meta solution to address various societal challenges. AI’s ability to handle vast amounts of data and extract meaningful insights can aid scientific research and healthcare. AI systems can assist top experts in these fields by surfacing important information and accelerating breakthroughs.

AI’s Role in Science and Healthcare:
Hassabis emphasizes the potential of AI to empower scientists and clinicians in their research and clinical practice. AI can process large datasets, identify patterns, and generate hypotheses that human experts may overlook. This collaboration between AI and human expertise can lead to faster and more effective scientific discoveries and improved healthcare outcomes.