Demis Hassabis (DeepMind Co-founder) – Systems Neuroscience and Artificial General Intelligence (Dec 2011)
Chapters
00:00:00 Biological Approaches to Artificial General Intelligence: Examining Systems Neuroscience
Approaches to AGI: Non-biological approaches (symbolic AI): mathematically/logically based, brittle, poor generalization, symbol grounding problem. Biologically inspired approaches: diverse range, from abstract to highly biological.
Search Space of AGI Solutions: Regime 1: small and dense, brain not special. Regime 2: large and sparse, brain may offer valuable insights. Evidence suggests we are in Regime 2.
Biological Approaches to AGI: Cognitive science architectures: abstract, box-and-wire diagrams, introspective design, unprincipled. Whole-brain emulation: detailed copying of the brain, wiring diagrams, imaging challenges, decades away from necessary technology.
Systems Neuroscience Approach: Sweet spot between abstract and biological approaches. Focus on algorithms the brain implements, not specific implementation details. Maps onto David Marr’s three levels of analysis: computational, algorithmic, and implementation levels.
00:11:20 Cognitive Neuroscience Revolution Drives AGI Research
Implementation Details and the Spectrum of AGI Research: Whole brain emulation focuses on the physical realization of intelligence, while cognitive scientist architecture approaches are more concerned with the computational level. This chapter focuses on the algorithmic level, which includes representations and algorithms for AGI.
The Revolution in Cognitive Neuroscience: In the last 15-10 years, cognitive neuroscience has seen a rapid revolution, with an exponential increase in research papers and the emergence of new experimental techniques. Newer technologies like fMRI, PET, and MEG are gaining prominence in cognitive neuroscience. Optogenetics, a recent technique, allows for precise control of neuronal activity using light, enabling causal studies of brain function.
Experimental Techniques and Their Contribution: fMRI measures brain activity by detecting changes in blood oxygen levels, providing insights into large-scale brain networks. PET scans track the distribution of radioactive tracers in the brain, allowing for the study of neurotransmitter systems and metabolism. MEG measures magnetic fields generated by neuronal activity, offering high temporal resolution for studying brain dynamics. Optogenetics enables the precise control of neuronal activity using light, allowing researchers to study causal relationships between specific neurons and behaviors.
The Potential of Cognitive Neuroscience for AGI Research: Cognitive neuroscience provides a wealth of data and insights into the algorithmic level of AGI, such as representations and algorithms used by the brain. By studying the brain, researchers can gain inspiration for designing AGI systems with similar capabilities and characteristics. The integration of cognitive neuroscience findings into AGI research can lead to more biologically plausible and efficient AGI systems.
00:13:30 Neuroscience and AGI: Collaborating for Progress
AGI and Neuroscience Connection: Demis Hassabis posits that neuroscience can play a significant role in developing AGI. Neuroscience offers two crucial aspects: inspiration for algorithms and architectures and validation testing by studying brain implementations.
Inspiration and Validation from Neuroscience: Neuroscience can provide inspiration for developing new algorithms and architectures, such as object recognition systems inspired by the primate visual system and navigation systems inspired by hippocampal place cells. Findings in neuroscience, like the ubiquitous implementation of reinforcement learning in the brain, serve as validation for the viability of algorithms in AGI systems.
Hybrid Approach: Hassabis advocates for a hybrid approach, combining the best of machine learning and neuroscience, rather than relying solely on neuroscience or pure machine learning. For known and useful components, state-of-the-art algorithms should be utilized, while for unknown components, both machine learning and neuroscience approaches should be pursued concurrently.
System Neuroscience Procedure: The process involves extracting principles behind algorithms used by the brain, creatively re-implementing them in computational models, and potentially resulting in state-of-the-art techniques and AGI components.
Interim Goals: Common interim goals include developing embodied physical robots and creating “toddler AGI” with cognitive behaviors similar to three-year-olds. Hassabis suggests that toddler AGI is an overly ambitious goal, requiring a vast range of capabilities, and proposes focusing on more attainable intermediate goals.
00:21:16 Conceptual Knowledge Acquisition and Representation
Basic Capabilities for AGI: Demis Hassabis proposes a set of core capabilities that are essential for human-level AGI, which he believes have been neglected in current and past AGI projects. These core capabilities include conceptual knowledge acquisition and representation, and the ability to plan and predict using that knowledge. Language, robotics, and logic systems are peripheral areas that have received more attention in AGI research but are not as fundamental as the core capabilities.
Conceptual Knowledge: Conceptual knowledge is abstract knowledge beyond perceptual information. The brain organizes knowledge into three levels: perceptual, conceptual, and symbolic. Machine learning algorithms can handle perceptual and symbolic information, but there is a gap in our understanding of how to bridge the gap to conceptual knowledge. This gap is known as the symbol grounding problem, which refers to the difficulty of connecting symbols to real-world concepts.
The Role of the Hippocampus in Conceptual Knowledge: The hippocampus is a brain region involved in memory and learning. During sleep, the hippocampus replays memories at an accelerated rate, which may help higher-order brain regions learn abstract information. This replay system emphasizes important memories and may lead to abstraction and semantic knowledge.
Potential Milestones in AGI Research: Abstract classification: Developing systems that can recognize abstract concepts like whether a container is empty or full, which requires understanding the relationship between objects and their properties. Abstract reasoning: Creating systems that can reason about abstract concepts and make inferences based on them. Generalization and transfer learning: Enabling systems to generalize knowledge from one domain to another and apply it to new situations. Lifelong learning: Developing systems that can continuously learn and adapt to new information and changing environments.
00:32:10 Conceptual Knowledge and Higher-Order Structure in AI
Knowledge Building and Grounding Problem: Classification systems face challenges in classifying concepts like “empty” and “full” due to the lack of knowledge about underlying concepts like solids, liquids, and containers. The “simple grounding problem” arises when labeling data sets without specifying which pixels or parts of the image are being referred to.
Discovery of Higher-Order Structure: Statistical learning techniques may be insufficient for discovering higher-order structure in data, as demonstrated by the number sequence 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 0, 1, 1, 1, 2, 1, 3. The chance of one number following another in this sequence is uniform, making it unpredictable despite having infinite data.
Algorithms for Sophisticated Environmental Models: The development of algorithms that can build sophisticated models of their environment is a potential solution to these limitations. Such algorithms could learn from past experiences and make predictions about future events by combining symbolic and statistical knowledge.
AGI Goals and Challenges: Transfer learning is vital for intelligence, allowing individuals to apply learned responses to new contexts. Difficulty in measuring progress towards AGI, with ad hoc task suites and lack of a principled measure of progress. Algorithmic IQ as a potential solution for integrated progress measurement.
Impressive AI Achievements: MoGo’s triumph as the first professional human Go player using Monte Carlo techniques. IBM’s Watson’s victory over human champions in the Jeopardy quiz show.
Towards AGI: Systems neuroscience insights will inspire solutions for key components of AGI. Transfer learning and conceptual knowledge acquisition capabilities in the next five years. Measurement tools and hill climbing algorithms will accelerate research progress. Safety concerns can be addressed as interim goals are achieved.
AGI Timeline: Human-level AGI is likely 20+ years away, but interim milestones will demonstrate progress and have practical applications.
Abstract
Understanding the Path to Artificial General Intelligence: A Systems Neuroscience Approach
The Pivotal Role of Neuroscience in Advancing AGI
In the quest to achieve Artificial General Intelligence (AGI), Demis Hassabis presents a compelling argument for a systems neuroscience approach. AGI aims to create machines with the ability to understand, learn, and apply intelligence broadly and flexibly, akin to human intelligence. Despite significant advances in AI, reaching human-level AGI is a complex endeavor, predicted to take over two decades. Hassabis, drawing on his deep expertise, emphasizes the potential of systems neuroscience in unlocking the secrets of AGI.
The Limitations of Current AI Approaches and the Search for a New Path
Traditional non-biological AI approaches, such as symbolic AI, face inherent limitations like brittleness, trouble with ambiguity, and poor generalization. Mathematically or logically based, these approaches often encounter the symbol grounding problem, where symbols are not easily associated with real-world concepts. In contrast, biologically inspired approaches attempt to replicate the brain’s form of intelligence. These approaches range from abstract cognitive science architectures, lacking deep underlying principles, to whole-brain emulation projects, hindered by the sheer complexity of the brain’s wiring. The core capabilities for AGI include conceptual knowledge acquisition and representation and the ability to plan and predict using that knowledge. Conceptual knowledge is abstract knowledge beyond perceptual information, while language, robotics, and logic systems are peripheral areas that have received more attention in AGI research. The hippocampus, a brain region involved in memory and learning, is crucial in the formation of conceptual knowledge.
Systems Neuroscience: The Sweet Spot for AGI Development
Hassabis argues that the sweet spot for AGI lies in systems neuroscience, which focuses on understanding the brain’s algorithms and representations. This approach aligns with David Marr’s three levels of analysis: computational, algorithmic, and implementation. By leveraging insights from neuroscience, researchers can identify key brain regions and mechanisms responsible for cognitive functions and develop more efficient AI algorithms. Furthermore, systems neuroscience will contribute in the development of algorithms capable of building sophisticated models of their environment, understanding relationships between objects and their properties.
Neuroscience’s Contributions to AI and AGI
Over the last two decades, cognitive neuroscience has witnessed rapid advancements, with technologies like fMRI, PET, and MEG becoming more prominent. Optogenetics, allowing precise control of neural activity, further enhances our understanding of the brain. The implementation details of AI systems fall into two broad regimes: small and dense (Regime 1) and large and sparse (Regime 2). While biologically inspired approaches are generally less principled and not as effective in Regime 1, evidence suggests that we are in Regime 2, where the brain may offer valuable insights. Classification systems face challenges in classifying concepts like “empty” and “full” due to the lack of knowledge about underlying concepts like solids, liquids, and containers. Statistical learning techniques may be insufficient for discovering higher-order structure in data. Whole-brain emulation approaches focus on the physical realization of intelligence, while cognitive scientist architecture approaches are more concerned with the computational level. This article focuses on the algorithmic level, which includes representations and algorithms for AGI.
Setting Realistic Interim Goals Towards AGI
Hassabis advocates for setting practical interim goals in the development of AGI. Rather than chasing full-scale embodied robots, which might distract from the primary goal of developing intelligence, more achievable milestones should be focused on. These goals should balance feasibility with their contribution to the gradual advancement towards human-level AGI. Potential milestones include abstract classification, abstract reasoning, generalization and transfer learning, and lifelong learning.
Core Capabilities for AGI: Conceptual Knowledge and Planning
Central to AGI is the ability to acquire and represent conceptual knowledge, coupled with planning and prediction capabilities. The brain’s hippocampus plays a crucial role in memory storage and replay, providing a teaching signal for the neocortex to learn abstract information. Understanding and replicating this process is vital for AGI development. Current AI systems struggle with abstract classification tasks, such as distinguishing empty and full containers, highlighting the need for advancements in conceptual knowledge.
The Challenge of Building Sophisticated AI Models
AI research aims to develop algorithms capable of building sophisticated models of their environment, understanding relationships between objects and their properties. However, current statistical learning techniques may fall short in discovering higher-order structures in data. This limitation underscores the need for AI systems that can abstract underlying rules from learned responses and apply them in different contexts.
Measuring AGI Progress with Algorithmic IQ
Evaluating the progress of AGI requires more than just ad hoc task-based assessments. Hassabis highlights the importance of a principled measure of progress, such as Algorithmic IQ, developed by Shane Legg and Joel Vaness. This approach evaluates the general intelligence of algorithms, providing a more comprehensive understanding of their capabilities.
The Future of AGI: Predictions and Insights
Looking ahead, Hassabis predicts significant progress in transfer learning and conceptual knowledge acquisition in AI systems within the next five years. He anticipates that improved measurement tools will open new research avenues and deepen our understanding of AGI and its associated safety issues. While human-level AGI may be more than 20 years away, tangible progress will be evident through interim milestones and practical applications.
The Interplay of Neuroscience and AI in AGI Development
Demis Hassabis’ perspective on AGI underscores the importance of a systems neuroscience approach, combining the strengths of both machine learning and neuroscience. By focusing on core capabilities, setting realistic interim goals, and employing principled measures of progress, the journey towards AGI appears both challenging and promising. Neuroscience, with its deep insights into brain functions, stands as a pivotal contributor in this endeavor, potentially accelerating the development of flexible, adaptable, and capable AGI systems.
Demis Hassabis' journey in AI spans from early fascination with chess and game programming to spearheading revolutionary achievements like AlphaFold and GATO, while also emphasizing the ethical development of AI and its potential to expand human knowledge and understanding. Hassabis envisions AI as a tool for humanity's advancement, scientific discovery,...
Artificial general intelligence (AGI) combines non-biological and biological approaches, with a focus on a hybrid method that leverages machine learning and systems neuroscience for conceptual knowledge acquisition and AGI validation. The pursuit of AGI aims to understand intelligence and consciousness through the development of advanced algorithms inspired by the brain's...
Demis Hassabis, a game-playing prodigy, left competitive gaming to pursue AI, combining neuroscience and computer science to create AI systems that learn like the human brain. Hassabis' motivation is not money, but achieving great things and leaving a lasting legacy....
Demis Hassabis' background in chess and gaming shaped DeepMind's culture and approach to AI, emphasizing long-term planning and resilience. AlphaGo's success was driven by Hassabis' vision, innovative strategies, and focus on transfer learning and intrinsic rewards....
Demis Hassabis' unique journey from chess prodigy to AI trailblazer involves his desire to use AI to solve humanity's grandest challenges in fields like scientific discovery, healthcare, and climate change. Hassabis envisions AI as a "meta-solution" to societal problems and emphasizes the need for responsible AI development and ethical considerations....
Neuroscience and AI converge in the pursuit of Artificial General Intelligence (AGI), with biological and non-biological approaches influencing the search space and development of AGI systems. Neuroscience techniques provide valuable insights into the brain's algorithms and representations, inspiring new AGI algorithms and validating existing ones....
DeepMind, led by Demis Hassabis, aims to create artificial general intelligence (AGI) by integrating neuroscience and AI, focusing on systems neuroscience and grounded cognition. Neuroscience inspires AI algorithms, architectures, and analysis techniques, leading to more capable and intelligent AI systems....