Demis Hassabis (DeepMind Co-founder) – Computational Neuroscience | Singularity Summit (Feb 2012)

Chapters

Abstract

Exploring the Frontier of Artificial General Intelligence: A Comprehensive Review

Artificial General Intelligence (AGI) stands as one of the most ambitious and challenging frontiers in modern science and technology. This comprehensive review article delves into the myriad approaches and significant advancements in the quest for AGI, drawing on insights from both non-biological and biological methodologies. The intricate interplay of these diverse strategies highlights the complexity and potential of this field.

Non-biological Approach to AGI: Symbolic AI and Cyc Project

At the heart of the non-biological approach to AGI lies Symbolic AI, characterized by formal logic systems, logic networks, lambda calculus, and expert systems. Projects in the 80s and 90s using these systems faced challenges in dealing with ambiguity, uncertainty, and generalization. Symbolic AI struggles with acquiring new symbols and creating meaning when symbols are only described in terms of other symbols.

Doug Leonard’s Cyc Project epitomizes the grand ambition and the formidable challenges in this approach. Cyc aims to encapsulate the world’s common sense knowledge in a massive database of predicate rules. However, the project grapples with the time-consuming nature of rule input and the complexities introduced by new rules, leading to inconsistencies that are difficult to resolve. Time-consuming development and maintenance with an estimated 600 man years invested, adding new rules can cause inconsistencies throughout the database that can take weeks to resolve.

Biological Approaches to AGI: From Brain Blueprints to Emulation

Biological approaches to AGI take inspiration directly from the human brain. This strategy is bifurcated into abstract approaches, like cognitive science architectures, and more direct ones like whole brain emulation. Abstract approaches seek to distill brain-like modules and functions, yet they are often unprincipled and suffer from a proliferation of differing architectures. Conversely, whole brain emulation aims for a detailed replication of the brain’s structure and function but raises questions about the practicality of such detailed implementation.

The challenges here are significant, especially the gap between understanding the brain’s wiring and its functions, and the daunting task of replicating the brain’s complexity in an artificial substrate. Biologically inspired approaches to AGI use the brain as a blueprint for AGI design. These approaches cover a wide range of methods and strategies. Some fall into regime one: a small and dense search space with many possible AGI solutions. Others, however, are in regime two: a large and sparse search space with relatively few AGI solutions. Current evidence suggests we are in regime two, a large and sparse search space.

Maher’s Three Levels of Analyses: A Framework for Understanding AGI

Maher’s three levels of analysis – computational, algorithmic, and implementational – provide a useful framework for understanding the various approaches to AGI. Whole brain emulation concentrates on the implementational level, trying to replicate the brain’s physical structure. In contrast, cognitive science architecture emphasizes the computational level, defining the system’s goals.

Systems Neuroscience: A Pathway for AGI Development

Systems neuroscience advocates a focus on the algorithmic level of the brain, aiming to extract representations and algorithms used by the brain to solve problems relevant to AGI. This approach benefits significantly from advancements in neuroscience, such as new imaging techniques and the exponential growth of neuroscience research. Recent progress in cognitive neuroscience, driven by advanced imaging techniques, has enabled the study of brain functions and algorithms at the algorithmic level. These advancements have opened up possibilities for extracting valuable insights for AGI development.

The Significance of Neuroscience in AGI Development

Neuroscience findings provide inspiration for new algorithms and serve as a benchmark for AGI systems. For instance, the discovery of hexagonal firing patterns in neurons and the workings of the dopamine system offer rich insights for developing AGI components. Systems neuroscience offers a pragmatic approach to understanding the brain’s algorithms, which can be valuable for AGI development.

Neuroscience Techniques and the Exponential Growth:

Neuroscience has seen a surge in new experimental techniques, including imaging, multi-cell recording, optogenetics, transmagnetic stimulation, photomicroscopy, and various analytical tools. With this, there has been an exponential growth in neuroscience research, evident in the substantial increase of citations and publications. fMRI, in particular, has exhibited exponential growth and now dominates the proportion of neuroscience papers. However, the vast amount of neuroscience data presents a challenge in identifying relevant information for AGI. Staying up-to-date with the latest developments is crucial.

Hybrid Approach and Conceptual Knowledge Acquisition

A hybrid approach that combines machine learning and systems neuroscience is increasingly advocated. This approach emphasizes the importance of conceptual knowledge acquisition and representation, key to AGI progress. The Semantic Pointer Architecture (SPA) stands out as a promising approach in this domain.

Systems Neuroscience for AGI Validation and Inspiration:

Neuroscience provides direction and validation testing for AGI. It offers inspiration for new algorithms and architectures based on the principles of brain function. For instance, Hubel and Wiesel’s work on simple and complex cells in the visual cortex influenced computer vision algorithms, Tomasso Poggio’s HMAX model mimics the primate visual system for object recognition, and John O’Keefe’s discovery of place cells and the Mosers’ discovery of grid cells provide insights for navigation systems. Neuroscientific principles can be creatively implemented in computational models, leading to state-of-the-art techniques and potential AGI components.

Levels of Information Processing and Analogies in Machine Learning

AGI development must consider different levels of information processing, from perceptual to symbolic. Analogies in machine learning, such as sensory learning techniques and symbolic AI, play crucial roles in bridging the gap between perceptual and symbolic processing.

Validation Testing with Reinforcement Learning:

Reinforcement learning algorithms are evaluated for their suitability in AGI systems. Temporal difference learning (TD learning) minimizes prediction error in future rewards. Reinforcement learning is ubiquitously implemented in the brain, supporting its consideration for AGI. Dopamine neurons exhibit behavior similar to TD learning, signaling prediction errors. The dopamine system’s architecture aligns with TD learning algorithms.

The Hippocampal Neocortical Consolidation System and AGI

The hippocampal neocortical consolidation system, crucial for memory storage and replay, offers valuable insights for AGI. This system’s ability to bias learning towards important information and facilitate the formation of abstract concepts is particularly noteworthy.

The Role of the Hippocampus and Entorhinal Cortex in Conceptual Knowledge:

Concepts are deemed crucial for progress towards AGI. Knowledge in the brain is divided into perceptual, conceptual, and procedural levels. Conceptual knowledge involves understanding relationships between concepts. The hippocampus and entorhinal cortex play roles in conceptual knowledge. Hexagonal firing patterns in the entorhinal cortex are used for spatial mapping, providing an intrinsic measure for space, like graph paper for the mind.

Implications for AGI and the Pursuit of Understanding Intelligence

Integrating machine learning with systems neuroscience can lead to groundbreaking algorithms and deepen our understanding of intelligence. The process of building AGI systems may also shed light on the enigma of consciousness, echoing Feynman’s assertion that building is tantamount to understanding.

Combining Machine Learning and Systems Neuroscience:

A hybrid approach that combines machine learning and systems neuroscience is advocated for AGI development. Utilize state-of-the-art machine learning algorithms where applicable. Explore systems neuroscience for potential solutions in unknown areas. Pursue a parallel approach, leveraging both disciplines simultaneously. Numerous systems in the brain hold potential for AGI research. Mirror neurons, model-based learning, theory of mind, working memory, and others are candidates for investigation.

Conclusion

The pursuit of AGI is a multifaceted endeavor, intertwining non-biological and biological approaches, advanced neuroscience, and machine learning techniques. This review underscores the complexities and potential of AGI development, highlighting the importance of a hybrid approach that leverages the strengths of different methodologies. As the field advances, it promises not only technological marvels but also deeper insights into the nature of intelligence and consciousness.

Notes by: oganesson