Overview of AGI Approaches:
AGI, or Artificial General Intelligence, encompasses attempts to build systems that mimic human-level cognitive abilities. Approaches to AGI are categorized into two broad groups: biological and non-biological.

Non-Biological Approaches (Symbolic AI):
Symbolic AI includes formal logic systems, lambda calculus, and expert systems. Challenges faced by Symbolic AI include brittleness, difficulty handling ambiguity and uncertainty, time-consuming training, and poor generalization. Specific problems such as acquiring new symbols and symbol grounding hinder these systems. An example of a symbolic AI project is Doug Leonard’s Sight project, which aims to encapsulate common sense knowledge in a database. Challenges with the Sight project include time-consuming rule addition and resolving inconsistencies caused by new rules.

Biological Approaches:
Biological approaches draw inspiration from the brain as a blueprint for AGI design. This category encompasses a wide range of approaches. The choice between biological and non-biological approaches depends on underlying beliefs about the most effective path to AGI.

AGI Search Space:
Two possible regimes for AGI solutions: Regime 1: Small and dense search base, many potential solutions Regime 2: Large and sparse search base, fewer potential solutions

Evidence for Regime 2:
Evolution has only produced human-level intelligence once after hundreds of millions of years Serious AGI projects in the 80s and 90s largely failed to make progress

Biological Inspiration:
Using a biologically inspired approach may be the right way forward Spectrum of approaches: Loosely based on the brain (abstract) Closely based on the biology (biological)

Cognitive Science Architectures:
Examples: Soar, Actar, OpenCog Inspired by behavioral psychology Module circuit diagrams Problems: Unprincipled Hard to prove one architecture is better than another Proliferation of architectures

Whole Brain Emulation:
Examples: Blue Brain Project, MODIS IBM Synapse Program Technical feats pushing supercomputers to the limit Question: Is this level of implementation detail useful for building AGI? Serious advocates want one-to-one mapping between a living brain and an artificial substrate

Systems Neuroscience and Brain Algorithms:
Systems neuroscience focuses on understanding the brain’s algorithms, rather than just its structure or function. The goal is to extract the representations and algorithms the brain uses to solve problems, which can be applied to developing artificial general intelligence (AGI).

Maher’s Three Levels of Analyses:
Maher’s three levels of analyses are computational, algorithmic, and implementational. Computational refers to the goals of the system. Algorithmic refers to the representations and algorithms used to achieve those goals. Implementational refers to the physical realization of the system.

Different Approaches to AGI:
Whole brain emulation focuses on the implementation level, trying to build a physical replica of the brain. Cognitive science architecture focuses on the computational level, defining the goals of the system. Systems neuroscience focuses on the algorithmic level, extracting the representations and algorithms the brain uses to solve problems.

Recent Advances in Neuroscience:
In the last decade, there has been a revolution in cognitive neuroscience, driven by new technologies like fMRI and EEG. This has allowed researchers to start answering questions about the brain’s algorithms that are useful for AGI development.

New Experimental Techniques and Data Analysis Tools:
Recent years have witnessed a surge in novel experimental techniques and data analysis tools in neuroscience. Examples include advanced imaging methods, multi-cell recording, optogenetics, transmagnetic stimulation, photomicroscopy, and multivariate pattern classifiers. These advancements have enabled researchers to investigate the brain in unprecedented detail.

Exponential Growth of Neuroscience Publications:
There has been an exponential growth in the number of neuroscience papers published. This growth is evident in various techniques, including fMRI, PET, and MEG. The proportion of neuroscience papers using these techniques has also increased significantly over time.

Challenge of Keeping Up with Neuroscience Research:
The rapid pace of neuroscience research poses a challenge for researchers to keep up with the latest findings. It requires continuous learning and immersion in both neuroscience and AGI to remain updated. This can take several years of intensive effort for motivated and smart individuals.

Neuroscience for AGI:
Neuroscience offers valuable insights and guidance for AGI research. It provides inspiration for new algorithms and architectures in AGI. Neuroscience can also serve as a validation test for AGI systems, assessing their performance against real-world neural data.

Examples of Neuroscience-Inspired Algorithms and Architectures:
Visual System: Hubel and Wiesel’s work on simple and complex cells in the visual cortex has inspired the development of hierarchical feature extraction algorithms in computer vision. Place Cells and Grid Cells: Research on place cells and grid cells in the rat hippocampus has led to the development of spatial navigation algorithms that are invariant to scale and translation.

Neuroscience Procedure or Approach:
The neuroscience procedure or approach involves extracting the principles behind an algorithm the brain uses and creatively implementing it in a computational model. This approach has led to state-of-the-art techniques and potentially useful AGI components.

Validation Testing:
Validation testing aims to determine if an algorithm, potentially created through machine learning or other methods, can serve as a viable component of an AGI system.

Reinforcement Learning:
Reinforcement learning algorithms are a classic example of validation testing. These algorithms learn from rewards gained from the environment without explicit instructions. Temporal Difference (TD) learning is a popular approach to reinforcement learning.

TD Learning in the Brain:
TD learning has been found to be implemented in the brain, particularly in the dopamine system. Dopamine neurons exhibit firing patterns consistent with prediction error, a key signal in TD learning. The dopamine system is widespread in the brain, suggesting a general role for TD learning in various cognitive processes.

Architecture Comparison:
The brain’s implementation of TD learning and engineering approaches to TD learning share similar architectures. This similarity suggests that incorporating reinforcement learning algorithms into AGI systems is a promising approach.

Other Systems:
In addition to dopamine, other neurotransmitter systems, such as serotonin, acetylcholine, and norepinephrine, play important roles in cognitive functions.

Hybrid Approach:
The speaker advocates for a hybrid approach that combines the best of machine learning and systems neuroscience. For components where we know how to build them, use state-of-the-art algorithms. For components where we don’t know how to build them, explore machine learning and look to systems neuroscience for ideas.

Conclusion:
The speaker emphasizes the importance of parallel exploration of machine learning and systems neuroscience in developing AGI systems.

Conceptual Knowledge as a Key to AGI:
Concepts are essential for AGI progress. Knowledge in the brain can be categorized into perceptual, conceptual, and symbolic levels. Perceptual knowledge involves sensory input, conceptual knowledge involves abstract concepts, and symbolic knowledge involves labeling those concepts.

Current AI Systems:
Symbolic AI attempts focus on the symbolic level of information processing. Recent advances in deep learning techniques have made progress in sensory learning and processing.

The Challenge of Conceptual Knowledge Acquisition:
The gap between perceptual and symbolic knowledge remains a challenge. Solving this problem would address the symbol grounding problem and enable a seamless transition from sensory knowledge to conceptual knowledge.

The Hippocampal Neocortical Consolidation System:
The hippocampus stores memories of recent experiences and replays them during slow-wave sleep at fast-speeded rates. This provides the neocortex with a large number of samples to learn from, even if an important event was experienced only once. Memories are selected stochastically for replay, with rewarded, emotional, and salient memories being replayed more often.

The Ability to Circumvent the Statistics of the External Environment:
The hippocampal neocortical consolidation system can circumvent the statistics of the external environment and bias the learning of important things for survival or progress. This system allows for abstraction and semantic knowledge.

Dreams and the Epiphenomenal Nature of Consciousness:
Dreams are a byproduct of the hippocampal neocortical consolidation system. Evidence suggests that dreams are epiphenomenal, meaning they do not have a specific purpose or function.

The Importance of Systems Neuroscience in Understanding Intelligence:
The brain is a useful proof of concept for artificial general intelligence (AGI). Systems neuroscience can inspire new algorithms and validate existing ones. Combining machine learning and systems neuroscience can lead to a deeper understanding of intelligence.

Building AGI Systems and Understanding Natural Intelligence:
Building AGI systems requires distilling intelligence into an algorithmic construct. Comparing this algorithmic construct to the mind can shed light on enigmatic concepts like consciousness.

A Cross-Disciplinary Research Program:
A research program combining machine learning and neuroscience can benefit both fields. This collaborative approach can lead to a better understanding of intelligence and consciousness.

Quote by Richard Feynman:

“What I cannot build, I cannot understand.”