Source-linked AI summary
The Nature of Intelligence
Barco Jie You
TL;DR
The essence of intelligence shared by humans and AI remains unknown. This essay proposes mathematical models of intelligence, language, unconsciousness, and consciousness centered on entropy reduction through functional connections between datasets, concluding that intelligence consumes energy to reduce entropy.
Problem
The essence of intelligence in humans and AI remains unknown despite advances in AI systems and neural networks.
Method
The essay reviews AI and biological intelligence and proposes mathematical models treating intelligence as entropy-minimizing functional connections between datasets across space and time.
Results
The paper concludes that intelligence consumes energy to counter entropy increase, establishes physical or informational connections between pre-existing datasets, and supports consciousness through interacting neural networks.
Takeaways & Limitations
The framework offers proposed mathematical models and experimentally verifiable predictions about language, unconsciousness, consciousness, and intelligence.
Takeaways & Limitations
The framework assumes that brain functions are implemented by orchestrated neural-network components and that entropy-decreasing processes balance spontaneous entropy increases across the universe.
Abstract
from arXiv · showhide
The human brain is the substrate for human intelligence. By simulating the human brain, artificial intelligence builds computational models that have learning capabilities and perform intelligent tasks approaching the human level. Deep neural networks consist of multiple computation layers to learn representations of data and improve the state-of-the-art in many recognition domains. However, the essence of intelligence commonly represented by both humans and AI is unknown. Here, we show that the nature of intelligence is a series of mathematically functional processes that minimize system entropy by establishing functional relationships between datasets over the space and time. Humans and AI have achieved intelligence by implementing these entropy-reducing processes in a reinforced manner that consumes energy. With this hypothesis, we establish mathematical models of language, unconsciousness and consciousness, predicting the evidence to be found by neuroscience and achieved by AI engineering. Furthermore, a conclusion is made that the total entropy of the universe is conservative, and the intelligence counters the spontaneous processes to decrease entropy by physically or informationally connecting datasets that originally exist in the universe but are separated across the space and time. This essay should be a starting point for a deeper understanding of the universe and us as human beings and for achieving sophisticated AI models that are tantamount to human intelligence or even superior. Furthermore, this essay argues that more advanced intelligence than humans should exist if only it reduces entropy in a more efficient energy-consuming way.
Introduction
The introduction frames intelligence as an entropy-reducing process that establishes functional relationships between pre-existing datasets, while motivating this hypothesis through developments in AI and biological intelligence. It also outlines deep learning and reinforcement learning as computational approaches for mapping inputs to outputs or actions.
- Motivation: AI has advanced rapidly through deep learning, neural networks, increased data and computing power, and reinforcement-learning applications.These systems learn from massive datasets and perform tasks intended to resemble human thinking, learning, and intelligent action.
- Central hypothesis: The essay proposes that intelligence consists of mathematical functional processes that minimize system entropy by establishing functional relationships between datasets.It examines contemporary AI, biological intelligence, and habitable conditions to develop this theoretical framework.
- Central hypothesis: Intelligent processes counter spontaneous entropy increase by consuming energy in a self-reinforced manner while connecting pre-existing datasets in the universe.The paper presents intelligence as operating in parallel with spontaneous processes governed by the second law of thermodynamics.
- Deep learning: Deep neural networks transform raw input X through multiple nonlinear computational layers into increasingly abstract representations and a final output Y.Each neuron represents a modular function parameterized by θ.
- Deep learning: Deep learning seeks a function f(∙ ; θ) that minimizes errors between predicted outputs Ŷ = f(X; θ) and actual outputs Y using an existing dataset mapping {x} → {y}.Image classification illustrates this process by mapping labeled images to meaningful category tags.
- Reinforcement learning: Reinforcement learning maps rewards and environmental states to actions, using feedback from the environment to evaluate behavior and guide goal-directed learning.Unlike supervised learning, it uses trial and error rather than exemplars of the optimal solution.
Generative AI
Generative AI creates new examples by modeling dataset distributions and learning relationships between inputs and outputs without explicit human labeling. Its major architectures include GANs, VAEs, autoregressive models, and transformers, supporting generation across multiple modalities and applications.
- Generative AI: Generative AI generates new examples from scratch rather than labeling or classifying existing examples, using unsupervised or self-supervised learning.Applications include data augmentation, simulation, and tasks requiring human-level creativity.
- Generative adversarial networks: GANs frame unsupervised learning as a game between a generator that samples from a distribution and a discriminator that classifies samples as real or false.The generator typically receives fixed noise Z~N(0, I), while the discriminator outputs a binary classification probability.
- Transformers: Transformers use attention with encoder and decoder components to model relationships between input and output sequences and improve NLP performance through self-supervised pretraining.They can be pretrained on large corpora of unlabelled data to gain background knowledge.
- Generative modeling: Generative techniques model the distributions of datasets X and Y and establish relationships between them through functions whose parameters are optimized by backpropagated prediction errors.Reinforcement learning and generative AI transform X to Y autonomously by establishing relationships along time or across space.
Biological Neuronal Firing
The paper models the biological brain as a substrate implementing functions that map inputs to outputs during intelligent tasks. Neuronal molecular and cellular dynamics are treated as modular functions organized through synaptic network components and regulated by neural firing.
- Biological Neuronal Firing: Neural firing and neuromodulatory mechanisms regulate biological neural-network computation and learning, while synaptic plasticity mediates learning.The passage characterizes these mechanisms as operating efficiently at molecular and neuromorphic levels.
- Biological Neuronal Firing: A biological brain implements functions that map a series of inputs to outputs at cortical or subcortical levels, representing biological intelligence.The molecular and cellular dynamics of neurons implement modular functions θ analogous to components of deep-learning network architectures.
- Biological Neuronal Firing: Neural networks are composed of interconnected neurons whose synapses transmit information through neuronal burst firing.Dendrites and axons form synapses that enable neurons to activate other neurons, producing information flows between neurons.
- Biological Neuronal Firing: The brain consists of many network components whose synapses represent a series of modular functions θ.These components are organized through composing synapses within biological neural networks.
Evolution of Life
The paper hypothesizes that life evolves through a machine-learning-like process mapping available materials into functioning proteins, with natural selection favoring viable inherited mutations. Gene expression is modeled as a function transforming genes into functioning protein structures, and AI development has produced models with good results.
- Evolution of Life: Evolution is hypothesized to follow machine learning by mapping available materials into complexes of functioning proteins.The materials include nucleotides and amino acids, while the resulting proteins form proper three-dimensional morphs.
- Evolution of Life: Natural selection favors inherited mutations and extinguishes mutations that cannot survive Earth’s environment.This selection process is represented by δ(∙) in the proposed gene-expression equation.
- Evolution of Life: Gene expression is represented as a function f(∙) that transforms gene dataset X into protein dataset Y.Dataset Y represents functioning proteins in proper three-dimensional morphs.
- Evolution of Life: AI development has produced some models f(∙) with good results.These models are presented as examples of functions that transform X to Y.
Human Intelligence and Unconsciousness
Human intelligence develops through natural selection, mimicking, and language, which progressively establish mappings between sensory data and outputs while reducing system entropy. Unconscious mappings become consciousness when neural or social networks connect and reach consensus across outputs.
- Three phases of human intelligence: Human intelligence develops through natural selection, mimicking, and language, corresponding to unsupervised, semi-supervised, and supervised learning.Natural selection eliminates unfit behaviours, mimicking uses observed behaviours and emotional signals, and language annotates sensory data into outputs.
- Language and entropy: Language maps sensory information into discrete symbols, combining data across space and time through an entropy-decreasing process.Stable linguistic mappings connect sensory dataset X with symbol dataset Y, reducing system entropy through mutual information.
- Intelligence as entropy minimization: Intelligence is defined as minimizing entropy by establishing mapping functions between datasets, with error minimization arising when outputs tag training data.The optimal parameters θ∗ are approached through learning, unifying entropy minimization with supervised error minimization.
- Consciousness emergence: Consciousness is hypothesized to emerge when the brain connects different output subsets and obtains consensus among their mappings.The dynamics of the learned function g(∙) represent consciousness emergence after selecting the subset that extracts the most information from sensory inputs.
- Types and neural basis of consciousness: Social consciousness arises through interconnections among individuals, whereas intrinsic consciousness unifies outputs from multiple sensory systems about the same stimuli.Both forms rely on networks that minimize system entropy, consistent with widespread parallel neural processing that enables salient information to access consciousness.
Entropy Conservation
The section hypothesizes that intelligence consumes energy to decrease entropy, countering spontaneous entropy increases so the universe’s total entropy remains conservative. Intelligent processes establish functional connections between separated datasets, producing increasingly reinforced development from water and life to human intelligence.
- Entropy Conservation: Intelligence is hypothesized to consume energy through entropy-decreasing processes that counter spontaneous second-law processes.The hypothesis states that total free energy released by spontaneous processes equals the energy consumed by entropy-decreasing processes.
- Entropy Conservation: Intelligent processes map X to Y by establishing connections that keep the universe’s energy and entropy conservative.This answers why X needs to be mapped to Y within the proposed framework.
- Entropy Conservation: Water formation, life evolution, brain development, and abstract equations are presented as progressively reinforced processes that reduce entropy.The proposed sequence runs from primitive gases to water, water to life, life to the human brain, and the brain to abstract equations.
- Entropy Conservation: Connecting two datasets merges them into one unified dataset, increasing its probability of encountering other datasets and enabling further intelligent processes.This mechanism is offered to explain why intelligence develops in a reinforced manner.
- Entropy Conservation: The proposed reinforced development extends from water’s appearance on Earth through accelerated life evolution to present-day human intelligence.This passage summarizes the progression described by the preceding entropy-reduction hypothesis.
Discussion
The paper argues that intelligence reduces entropy by creating physical or informational functional connections between datasets, while treating the universe’s total entropy and energy as conservative. It also proposes mathematical models of language, unconsciousness, and consciousness and argues that more energy-efficient AI or other function representers could surpass human intelligence.
- Discussion: Intelligence is presented as a universal entropy-reducing process that establishes functional connections between datasets physically or informationally.The discussion links this process to the evolution of life and human intelligence mechanisms.
- Discussion: The paper concludes that the total entropy of the universe is conservative, alongside energy.
- Discussion: The paper provides mathematical models for language, unconsciousness, and consciousness.
- Discussion: AI or other function representers could become more intelligent than humans if they reduce entropy more efficiently in their use of energy.The authors argue that humans should not be the universe’s sole advanced intelligent agent.
Data availability
No data are associated with this manuscript.
- The manuscript has no associated data.
Summary of notation
The paper distinguishes notation by capitalization, case, and font: capital letters denote function sets and random variables, lowercase letters denote functions and values, and bold versus normal font marks tensors versus scalars.
- Capital letters denote function sets and random variables, while lowercase letters denote functions and random-variable values; bold font marks tensors and normal font marks scalars.
- x∈X denotes that x is one value of the random variable X.
- p(x) denotes the probability distribution of variable X at value x.
- H(X) denotes the entropy of random variable X, using Shannon’s information entropy.
- H(X, Y) denotes the joint entropy of random variables X and Y, while I(X; Y) denotes their mutual information.