Background:
Dr. Demis Savas is a remarkable figure known for his contributions to memory research, the video game industry, and artificial intelligence (AI). He co-founded DeepMind, acquired by Google, and leads their general AI efforts.

Achievements:
Savas’s self-taught AI software achieved human-level performance in half of 49 games and made headlines with AlphaGo, the first program to defeat a professional Go player. He is a former chess prodigy, completed A-levels early, and coded the multi-million-selling game Theme Park at age 17. Savas earned a double first in Computer Science from Cambridge, founded Elixir Studios, and completed a doctorate at University College London. His research on memory and imagination was listed among the top 10 scientific breakthroughs of 2007 by the journal Science. Savas is a five-time World Games champion, a recipient of the Royal Society’s Mullard Award, and the Academy’s silver medalist.

Introduction:
Savas is honored to present the autumn lecture on general artificial intelligence at the Royal Academy. He will discuss AI, particularly general AI, based on DeepMind’s work.

DeepMind’s Mission:
DeepMind aims to fundamentally solve intelligence by understanding how it works and recreating it artificially. By achieving this, they believe they can use the technology to solve various other problems and enhance human intelligence.

General Purpose Learning Algorithms:
DeepMind focuses on creating general purpose learning algorithms capable of learning from raw experience and operating across a wide range of tasks. This approach, known as Artificial General Intelligence (AGI), aims to develop flexible, adaptive, and inventive systems.

Comparison to Narrow AI:
Narrow AI systems are designed for specific tasks and are brittle when faced with unexpected situations. Examples of narrow AI include IBM’s Deep Blue, which defeated Garry Kasparov in chess but lacked the ability to generalize its knowledge to other games.

Reinforcement Learning Framework:
DeepMind employs reinforcement learning as a framework for understanding intelligence. Reinforcement learning involves an agent interacting with an environment, observing its actions’ consequences, and adjusting its behavior accordingly. This approach is inspired by biological learning systems, particularly the dopamine system in humans.

Grounded Cognition:
DeepMind believes that true thinking machines must be grounded in a sensory-motor reality. They use games as a platform for developing and testing AI algorithms due to their efficiency and ability to provide sensory data similar to real-world situations. Virtual environments allow for rapid algorithm development compared to using real robots.

Benefits of Virtual Environments for AI Training:
Unlimited training data due to prolonged virtual environment runtime. Elimination of testing bias. Parallel testing for faster progress evaluation. Measurable progress for long-term research goals.

End-to-End Learning Agent:
Perception to action learning using raw pixel inputs. Focus on the entire problem stack, from perception to action. Deep reinforcement learning combines deep learning and reinforcement learning.

Deep Reinforcement Learning:
Combines deep learning for perception processing and reinforcement learning for action selection. Enables reinforcement learning to operate at a large scale.

DQN: A General Reinforcement Learning System:
Plays 50 classic Atari games without any game-specific knowledge or parameter adjustments. Learns everything from raw pixel inputs and game experiences. Demonstrates mastery across games with diverse visuals and objectives.

AlphaGo: Mastering the Ancient Game of Go:
AlphaGo program developed to play the complex game of Go. Go is a 19 by 19 grid game with black and white stones played on vertices. Considered more than just a game in Asia, with a long history and cultural significance.

The Popularity of Go:
Go, also known as Weiqi or Baduk, is a strategy board game that originated in ancient China and is now popular in various parts of Asia, including Japan, Korea, and China. It has a vast player base of 40 million active players and over 2,000 professional players. Go schools exist in countries like Japan, Korea, and China, where talented young players are enrolled at a young age to receive specialized training and education in Go.

The Complexity of Go:
Go is considered a complex game due to the enormous number of possible board positions. There are approximately 10 to the 170 possible board positions, far exceeding the number of atoms in the observable universe (10 to the 80 atoms). This immense complexity makes it challenging for computers to master the game using brute force methods.

The Elegance of Go:
Despite its complexity, Go is often regarded as one of the most elegant games ever devised. Its elegance stems from the simplicity of its rules, with only two primary rules governing the game.

The Main Rule of Go:
The main rule of Go centers around the capturing of stones. When a player surrounds an opponent’s stone with their own stones, the surrounded stone is captured and removed from the board.

Scoring in Go:
In Go, players aim to surround their opponent’s pieces and wall off empty parts of the board. The empty areas of the board and the captured pieces contribute towards a player’s final score. The player with the higher score wins the game.

Challenges of Go for Computers:
Massive search base: 200 possible moves per position compared to 20 in chess. Lack of an evaluation function: Determining who’s winning in Go is complex.

Neural Networks to Overcome Complexity:
Policy Network: Predicts the most likely moves in a position, reducing search breadth. Value Network: Evaluates positions and estimates who’s winning. Trained by playing against itself millions of times using reinforcement learning.

AlphaGo’s Performance:
Defeated top computer programs 100% of the time. Won 4-1 against Lee Sedol, the greatest Go player of the decade. Achieved a milestone in AI research, coming a decade before expectations.

Cultural Impact:
280 million viewers overall, 60 million in China for the first game. 35,000 press articles in a week. Popularized Go in the West, boosting sales and profile.

Further Reading:
Nature article on AlphaGo’s technical details.

AlphaGo’s Creative Move in Game Two:
AlphaGo’s move on move 37 in game two against Lisa Doll astounded the Go world for its unconventional nature.

Traditional Go Strategy:
In Go, players typically focus on playing on either the third or fourth line to gain territory or influence.

AlphaGo’s Unorthodox Approach:
AlphaGo broke with tradition by playing on the fifth line, sacrificing territory in the fourth line.

Long-term Impact of the Move:
This seemingly disadvantageous move ultimately influenced the outcome of the game 50 moves later.

Analysis by Go Expert Michael Redmond:
Michael Redmond, a renowned Go player and commentator, expressed surprise and disbelief at AlphaGo’s move. He initially thought it was a mistake or a misclick by the computer operator.

Significance of the Move:
AlphaGo’s unconventional move highlighted its ability to think strategically and plan long-term, changing the course of the game.

AlphaGo’s Move 37 and Lee Sedol’s Reaction:
Move 37 of the AlphaGo vs. Lee Sedol match was particularly astonishing, leaving Sedol perplexed. Sedol left the game room for 15 minutes to clear his head after witnessing the move. AlphaGo’s move had a 1 in 10,000 chance of being played, according to its own probability calculations.

Lee Sedol’s Transformation:
After the match, Lee Sedol experienced a resurgence in his career, winning 90% of championship matches. Sedol attributed this improvement to AlphaGo, which stimulated his creativity and flexibility in thinking. Sedol shifted his approach from relying on intuition to making more calculated decisions.

Intuition and Creativity in Go:
In Go, intuition and feel play a significant role in decision-making, unlike in chess, where players rely on calculated plans. AlphaGo’s ability to make intuitive moves demonstrated its creativity and its departure from traditional, rule-based AI approaches.

Deep Blue vs. AlphaGo:
Deep Blue relied on handcrafted chess knowledge and exhaustive search to beat Garry Kasparov. AlphaGo, in contrast, utilized selective search guided by neural networks, requiring fewer calculations.

Definition of Intuition and Creativity:
Intuition: Implicit knowledge acquired through experience, not consciously accessible or expressible. Creativity: Ability to synthesize knowledge to produce novel or original ideas.

Systems Neuroscience Inspiration:
DeepMind’s research program incorporates systems neuroscience to understand the brain’s high-level algorithms and representations. The hippocampus, a brain region involved in memory, is being studied for its role in imagination and future planning. DeepMind aims to create an artificial hippocampus to complement its neural network systems.

Real-World Applications:
DeepMind’s algorithms are being applied to various real-world problems, including healthcare, medical diagnostics, robotics, and energy efficiency. In a data center project, AlphaGo-like techniques optimized the cooling systems, saving 15% of the power usage. Future plans include optimizing entire power grids.

Speech Synthesis with Concatenative Models:
Concatenative models are state-of-the-art systems used in text-to-speech synthesis. Actors record 30 hours of various dialogues, and their speech is chopped into individual syllables. When new text is given, the system stitches together syllables to create new words and sounds. This approach, while effective, results in a robotic sound.

WaveNet: Modeling Raw Speech Waveforms:
WaveNet is a novel neural network that models the raw speech waveform directly. It can generate new waveforms without stitching together syllables from a pre-defined data set. This approach enables the creation of new waveforms for text that the network has never seen before.

Visualizing Speech Waveforms:
Audio waveforms are visual representations of speech. The waveform of Neil Armstrong’s famous phrase, “That’s one small step for a man, one giant leap for mankind,” is shown as an example.

WaveNet’s Key Innovations:
Dilation: WaveNet’s neural networks take inputs from multiple positions in the waveform, allowing for more complex relationships to be learned. Residual connections: WaveNet uses residual connections, which help the network learn faster and reduce the risk of vanishing gradients.

Benefits of WaveNet:
Naturalness: WaveNet-generated speech is highly natural and human-like, surpassing the quality of concatenative models. Expressiveness: WaveNet can capture the nuances of emotion and intonation, making the generated speech more expressive. Versatility: WaveNet can be used for a wide range of applications, including text-to-speech, voice cloning, and speech enhancement.

WaveNet:
WaveNet is an innovative neural network architecture specifically designed for audio generation. It excels in generating coherent and natural-sounding speech. WaveNet operates by progressively increasing the dilation of input channels, allowing it to take more and more time into account as it moves up the network hierarchy, providing the necessary coherence for speech. It also employs auto-regressiveness, where the generated sound is fed back as input, allowing the network to condition its next generation on all previous outputs, resulting in coherent speech and words. WaveNet significantly improves the quality of generated speech compared to the state-of-the-art, achieving a 50% improvement in the gap between current technology and perfect human speech.

Humanoid Robots:
Despite advancements in physical capabilities, humanoid robots still lack the intelligence to match their Hollywood portrayals. The primary obstacle lies in the development of algorithms and intelligence rather than physical capabilities. While ex machina-like humanoid robots are still decades away, practical industrial applications of robotics are expected to emerge within the next five years.

AI in Humanitarian Issues:
AI has the potential to be applied to humanitarian issues such as conflict prediction, food security, and child protection. Challenges arise in codifying complex factors such as emotions and human motivations into data. NGOs are seen as a promising starting point for AI collaboration, given their keenness to engage with technology and address issues like disease control, poverty, and water security.

AI in Automotive Safety:
AI can play a significant role in preventing car crashes and automotive accidents. Autonomous vehicles have the potential to reduce accidents caused by human error and improve road safety. Researchers are exploring various approaches, including deep learning and reinforcement learning, to develop safer and more efficient autonomous vehicles.

AI and Autonomous Vehicles:
Most current autonomous vehicles use learning systems for aspects like pedestrian identification and vision systems, but planners are usually handcrafted and rule-based. The debate on whether handcrafted systems will be enough for autonomous vehicles depends on the level of autonomy desired. For highway driving, a rule-based system may suffice, but for all road conditions and unexpected events, learning may be necessary.

Defining Virtual Life and Emotional Attachment to AI:
A simple definition of virtual life could be feeling a pang of guilt when turning off an AI system. Current AI systems like Atari programs and AlphaGo do not elicit such emotions, but future systems with more complex behaviors might. The design choice of giving AI systems freedom to design their own goals will also influence our emotional attachment to them.

Mitigating Biases in AI Systems:
Analysis tools are being developed to understand what AI systems are learning and why they make certain decisions. By identifying biases in AI systems, we can mitigate them by adjusting the system or compensating with other data. The Virtual Brain Analytics project is an example of efforts to address these biases.

AI for Monitoring Humanitarian Efforts:
AI can aid humanitarian efforts by monitoring large volumes of data from satellites, drones, and other sources. This can help NGOs and other organizations track progress, hold governments accountable, and make data-driven decisions.

Ethical Considerations for AI:
AI technologies are inherently neutral and their impact depends on their intended use. Ethical considerations should be addressed as AI systems become more powerful. Research is needed to develop methods for setting appropriate goals, securing those goals, and ensuring responsible AI development.

The Singularity Debate:
Demis Hassabis does not subscribe to the idea of a technological singularity, considering it a science fiction concept. Assumptions about AI’s self-development capabilities may not hold true as we engineer and better understand these systems. Instead of focusing on a singularity, attention should be given to the ethical and responsible use of AI.

Human-AI Collaboration:
AI can be a powerful tool to enhance human capabilities. An eventual merging of AI and human intelligence is a possibility. Demis Hassabis envisions AI as a means for humans to expand their abilities and achieve more.

AI’s Role in Human Ingenuity:
Demis Hassabis believes that AI can multiply human ingenuity by processing vast data and finding insights. Computers can assist humans in keeping up with the latest information and making sense of it.

Knowability and Explainability of AI:
A participant expressed concerns about our ability to understand how AI algorithms learn and make decisions, especially when they are trained on enormous amounts of data. Hassabis acknowledged this challenge and emphasized the need for better tools to interrogate and understand AI systems.

Inspiration from Neuroscience:
Hassabis draws inspiration from neuroscience, where visualization and statistical tools are used to analyze brain activity. He calls for the development of similar tools for artificial neural networks, which they term “virtual brain analytics.”

Addressing AI Hype:
In response to questions about AI hype, Hassabis clarified that current AI systems like Siri and driverless cars are not examples of general AI. He categorized AI into narrow AI, wide AI, and general AI, with general AI being the most advanced and human-like.

Turing Test and Hard vs. Soft AI:
Regarding the Turing test and the hard vs. soft AI debate, Hassabis did not explicitly state his position. However, he mentioned that human-like behavior does not necessarily indicate human-level intelligence.

Conclusion:
Hassabis emphasized the need for better tools to understand AI systems and addressed concerns about the knowability and explainability of AI. He also clarified different categories of AI and highlighted the challenges in achieving general AI.

The Turing Test:
Demis Hassabis believes the Turing Test, initially proposed by Alan Turing in the 40s and 50s, is not a rigorous test for assessing machine intelligence. Hassabis argues that the Turing Test can be easily gamed and fooled, particularly with short text-based dialogues. He emphasizes that humans can be misled by systems that fulfill certain expectations, even if they exhibit limited intelligence.

Consciousness and Intelligence:
Hassabis considers consciousness and intelligence as dissociable traits. He believes it is possible to create highly intelligent machines that lack consciousness, similar to how humans perceive each other. Hassabis acknowledges that consciousness is not a well-defined capability and that philosophers have yet to reach a consensus on its definition.

AI’s Potential Impact on Consciousness Research:
Hassabis suggests that the development of general AI, inspired by neuroscience, might shed light on the phenomenon of consciousness. He envisions the potential for future artificial systems that can be compared to the human mind, providing insights into consciousness, dreaming, creativity, and other complex mental processes.

Skills Essential in the Future:
Hassabis predicts that creative skills will become increasingly valuable in the future, as AI systems handle data-intensive tasks. He believes top creative individuals will be able to leverage AI tools to create unique and valuable solutions. Hassabis also suggests that self-learning systems may democratize AI and engineering, potentially reducing the need for specialized programming knowledge. Human skills like empathy and emotional intelligence may also become more sought after in the future.