[Paper Review] A Theory of Intelligences

Michael Hochberg|arXiv (Cornell University)|Aug 23, 2023

Emotional Intelligence and Performance4 citations

TL;DR

This paper proposes a first-principles Theory of IntelligenceS (TIS) that frames intelligence as the resolution of uncertainty toward goal achievement, partitioning it into path efficiency (solving) and goal accuracy (understanding). The theory introduces quantifiable macro-scale metrics—difficulty, surprisal, and goal resolution accuracy—and predicts that intelligent paths not only achieve goals but also expand future goal-space access, explaining non-Darwinian endeavors like art and politics.

ABSTRACT

Intelligence is a human construct to represent the ability to achieve goals. Given this wide berth, intelligence has been defined countless times, studied in a variety of ways and represented using numerous measures. Understanding intelligence ultimately requires theory and quantification, both of which have proved elusive. I develop a framework -- the Theory of Intelligences (TIS) -- that applies across all systems from physics, to biology, humans and AI. TIS likens intelligence to a calculus, differentiating, correlating and integrating information. Intelligence operates at many levels and scales and TIS distils these into a parsimonious macroscopic framework centered on solving, planning and their optimization to accomplish goals. Notably, intelligence can be expressed in informational units or in units relative to goal difficulty, the latter defined as complexity relative to system (individual or benchmarked) ability. I present general equations for intelligence and its components, and a simple expression for the evolution of intelligence traits. The measures developed here could serve to gauge different facets of intelligence for any step-wise transformation of information. I argue that proxies such as environment, technology, society and collectives are essential to a general theory of intelligence and to possible evolutionary transitions in intelligence, particularly in humans. I conclude with testable predictions of TIS and offer several speculations.

Motivation & Objective

• To develop a unified, first-principles theory of intelligence applicable across biological, artificial, and physical systems.
• To address the lack of a coherent theoretical and quantitative framework for intelligence by identifying core features such as difficulty, surprisal, and goal accuracy.
• To resolve conceptual ambiguities in intelligence by partitioning it into uncertainty reduction (solving) and goal accuracy (understanding).
• To explain non-Darwinian intelligent behaviors—like art, politics, and leisure—through the predictive power of path exploration and future goal-space expansion.
• To establish a theoretical foundation for intelligence that accounts for temporal dynamics, environmental influences, and evolutionary trajectories.

Proposed method

• Proposes a macro-scale framework for intelligence based on three quantifiable system features: difficulty, surprisal, and goal resolution accuracy.
• Partitions intelligence into two components: path intelligence (efficiency in uncertainty reduction) and understanding (accuracy in goal achievement).
• Introduces a compact mathematical form for surprisal and difficulty, enabling theoretical modeling of intelligence dynamics.
• Uses temporal spaces to model intelligence across past sources, present proxies, environments, and future evolution, including evolutionary fitness and aging.
• Applies the theory to diverse systems, from individual cognition to AI and non-designed physical/chemical systems, via comparative analysis of goal achievement and path efficiency.
• Employs a conceptual framework rooted in information theory, first principles, and system dynamics to model intelligence as a process of active goal resolution.

Experimental results

Research questions

• RQ1 To what extent are microscopic processes necessary to understand macroscopic patterns in intelligences?
• RQ2 Can we identify a core set of phenotypic traits explaining the main effects and variances in intelligences?
• RQ3 Are there typical patterns in hierarchical influences, entropy changes, and accuracy shifts along paths to goal resolution?
• RQ4 How do intelligence and goal complexity coevolve during a lifetime and across populations?
• RQ5 Does the age-related decline in fluid intelligence parallel the age-dependent force of selection, and if not, why?

Key findings

• The theory partitions intelligence into two distinct components: uncertainty reduction (solving) and goal accuracy (understanding), with the former enabling broader future goal-space access.
• Intelligent paths are not only efficient in achieving goals but also function as exploratory mechanisms that increase the probability of attaining new, previously inaccessible goals.
• The framework explains non-Darwinian intelligent behaviors—such as art, politics, and leisure—by showing they emerge from path exploration that expands future goal potential.
• The theory predicts that intelligence traits impact reproductive fitness and are subject to assortative mating, with fluid intelligence declining with age, consistent with evolutionary theory.
• TIS provides a compact mathematical form for surprisal and difficulty, enabling theoretical modeling of intelligence across diverse systems, including life, collectives, and AI.
• The theory is falsifiable: it predicts that in cases of low difficulty and low surprisal, active intelligence (as defined by equation 8) is unnecessary, which aligns with empirical expectations.