[Paper Review] A Theory of Natural Intelligence

Christoph von der Malsburg, Thilo Stadelmann+1 more|arXiv (Cornell University)|Apr 22, 2022

Neural Networks and Applications4 citations

TL;DR

This paper proposes that natural intelligence arises from structural regularity in the brain, encoded as self-organized net fragments—stable, recurring network patterns formed through network self-organization. These fragments act as powerful inductive biases that enable rapid learning, generalization, and autonomous behavior by bridging abstract goals and concrete sensory-motor experiences through homeomorphic mappings.

ABSTRACT

Introduction: In contrast to current AI technology, natural intelligence -- the kind of autonomous intelligence that is realized in the brains of animals and humans to attain in their natural environment goals defined by a repertoire of innate behavioral schemata -- is far superior in terms of learning speed, generalization capabilities, autonomy and creativity. How are these strengths, by what means are ideas and imagination produced in natural neural networks? Methods: Reviewing the literature, we put forward the argument that both our natural environment and the brain are of low complexity, that is, require for their generation very little information and are consequently both highly structured. We further argue that the structures of brain and natural environment are closely related. Results: We propose that the structural regularity of the brain takes the form of net fragments (self-organized network patterns) and that these serve as the powerful inductive bias that enables the brain to learn quickly, generalize from few examples and bridge the gap between abstractly defined general goals and concrete situations. Conclusions: Our results have important bearings on open problems in artificial neural network research.

Motivation & Objective

• To explain the superior learning speed, generalization, and autonomy of natural intelligence compared to current AI systems.
• To identify the missing mechanism—structural regularity in neural networks—that enables efficient learning beyond what genes or experience alone can provide.
• To propose that net fragments, formed via network self-organization, serve as the core inductive bias for intelligent behavior.
• To bridge the abstraction gap between innate behavioral schemata and concrete perceptual experiences using homeomorphic mappings.
• To provide a unified framework for autonomous, goal-directed learning in artificial systems by emulating the brain's structural regularity.

Proposed method

• Proposes network self-organization as the Kolmogorov algorithm for brain connectivity, minimizing the information needed to generate complex neural structures.
• Introduces 'net fragments'—self-organized, recurrent network patterns that encode structural regularity and serve as inductive biases.
• Uses homeomorphic mapping to relate abstract behavioral schemata to concrete scene representations via shared structural patterns.
• Models scene representation as a composition of net fragments that capture essential relational and spatial structures of the environment.
• Applies the concept of low Kolmogorov complexity to argue that brain connectivity requires minimal algorithmic information, implying high structural regularity.
• Suggests that meta-learning identifies the appropriate abstraction level for fixing experiences into memory via net fragment composition.

Experimental results

Research questions

• RQ1 How can natural intelligence achieve fast, generalizable learning with minimal data, given the low information content of genes and experience?
• RQ2 What structural mechanism in the brain enables the transition from innate behavioral schemata to flexible, context-sensitive behavior?
• RQ3 How do ideas and imagination emerge from neural network dynamics without relying on external programming or quantum effects?
• RQ4 What role does structural regularity—encoded as net fragments—play in enabling inductive bias for efficient learning?
• RQ5 How can artificial systems achieve autonomy and common-sense reasoning by mimicking the brain’s structural self-organization?

Key findings

• The human brain’s connectivity, requiring a petabyte to describe, is generated by a Kolmogorov algorithm of far lower complexity, implying extreme structural regularity.
• Net fragments—self-organized, recurring network patterns—act as domain-specific inductive biases that enable rapid learning and generalization.
• The ontogenetic development of retinotopic maps exemplifies network self-organization as a mechanism that generates structural regularity with minimal information.
• Structural regularity in the brain allows it to learn from few examples and generalize beyond seen data, overcoming the limitations of current AI systems.
• Homeomorphic mappings between behavioral schemata and scene representations enable the brain to link abstract goals to concrete sensory-motor experiences.
• The superiority of human intelligence over animal intelligence is attributed to a rich, culturally acquired set of schemata grafted onto innate, net fragment-based neural structures.