Machine Learning Street Talk (MLST)

Dr. MAXWELL RAMSTEAD - The Physics of Survival

Sun Jul 16 2023

Free Energy PrincipleUnified TheoryNeuroscienceArtificial IntelligencePhilosophyComplex SystemsMarkov BlanketsActive InferenceConsciousnessPhysics

Description

This episode explores the free energy principle, a unified theory encompassing mathematics, physics, philosophy, complex systems, and cognition. It explains how living things maintain their existence by minimizing free energy and has implications for neuroscience, artificial intelligence, and philosophy. The episode covers topics such as the relationship between science and philosophy, the conversion experience of a philosopher, scientific modeling using the free energy principle, connections to classical mechanics and maximum entropy, Markov blankets and self-organization, active inference and machine learning, responsible AI development, misconceptions about the free energy principle, critiques and refinements, extensions of the principle including maximum caliber, consciousness and nested systems, and the unification of physics, mind, and intelligence.

Insights

The Free Energy Principle

The free energy principle is a unified theory encompassing mathematics, physics, philosophy, complex systems, and cognition. It explains how living things maintain their existence by minimizing free energy and has implications for neuroscience, artificial intelligence, and philosophy.

Markov Blankets

Markov blankets define boundaries between what's inside and outside a system. They can be visualized as cell membranes separating internal and external states. Markov blankets are important in the context of the free energy principle as they play a role in modeling systems that track aspects of their environment.

Active Inference

Active inference is a new way of understanding complex systems and is expected to be the future of machine learning. It involves building AI systems based on explicit generative models that describe dependency relations within the system. Generative models allow for performing inference using variational free energy.

Consciousness and Nested Systems

Consciousness corresponds to an inner Markov blanket or inner screen within a system. It emerges when inference reaches sufficient temporal and counterfactual depth. Conscious experience may be weakly emergent rather than strongly emergent, and a theory of consciousness can be derived directly from the Free Energy Principle.

Physics, Mind, and Intelligence

The free energy principle unifies all of reality under the auspices of classical mechanics. There is no distinction between physics, biology, and the mind; it's all just physics in some sense. The free energy principle allows us to think about the expression of intelligence in physical systems.

Chapters

Introduction to the Free Energy Principle
Philosophy and Formal Training
Conversion Experience and Correspondence with Carl Friston
Scientific Principle and Modeling Systems
Mathematical Truths and Connections to Physics
Markov Blankets and Self-Organization
Active Inference and Machine Learning
Active Inference and Explainability
Maximum Entropy and Free Energy Principle
Spatial Web Foundation and Responsible AI Development
Openness, Transparency, and Ethical Design
Research Group and Misconceptions about the Free Energy Principle
Critiques and Refinements of the Free Energy Principle
Maximum Caliber and Extensions of the Free Energy Principle
Ensemble Behavior and Nested Systems
Markov Blankets and Consciousness
Physics, Mind, and Intelligence

Summary

Transcript

Introduction to the Free Energy Principle

00:00 - 08:35

Maxwell Ramsted is a philosopher, mathematician, and leading prognosticator on the free energy principle.
The free energy principle is a unified theory encompassing mathematics, physics, philosophy, complex systems, and cognition.
The principle posits that systems must minimize free energy to safeguard their existence.
Living beings are engaged in a relentless struggle to model, predict, and comprehend their surroundings as a survival imperative.
The free energy principle extends its intellectual dominion over life, cognition, and intelligence.
It explains how living things maintain their existence by minimizing free energy.
The principle has implications for neuroscience, artificial intelligence, and philosophy.

Philosophy and Formal Training

08:18 - 16:54

Philosophy and formal training go hand in hand, but it's a strange historical circumstance that philosophy is not seen as a formal discipline anymore.
In the past, there was no separation between science and philosophy, but now there is a distinction.
Being a philosopher of science is difficult because you have to be familiar with the history and contemporary philosophy of science, as well as get proper training in the discipline you are reflecting upon.
To effectively exist at the intersections of different disciplines, one needs to be a polymath.
The intellectual community around the free energy principle (FEP) and active inference is multidisciplinary by nature.
The FEP is an explanatory principle that applies to every scale at which physical systems self-organize, making insights from various disciplines relevant.
The FEP acts as a meta-theoretical architecture to fit claims together in a tractable way from physics and mathematics perspectives.
The speaker has always been interested in both sciences/mathematics and big questions, leading them to specialize in philosophy while remaining passionate about physics and math.
They were exposed to the free energy principle and Carl Friston's ideas through reading Andy Clark's paper called 'Whatever Next', which combined everything they found interesting.
Meeting Carl Friston in person was a privilege for the speaker.

Conversion Experience and Correspondence with Carl Friston

16:27 - 24:10

The speaker had a conversion experience after learning about the free energy principle and met Carl in person.
They corresponded extensively and Carl became their PhD supervisor.
The podcast hosts also found the free energy principle fascinating and had Professor Fristen on the show multiple times.
The free energy principle is an inversion of the question of survival, focusing on what behaviors must exist for something to continue surviving.
It explains that if something exists as a separate but coupled system, it will appear as if it is tracking or representing features of its environment.
This tracking or representing relation is weak and not necessarily contentful representations like images in one's head.
Instead, it represents abstractions of the environment to maintain enough information for prediction and tracking behaviors.

Scientific Principle and Modeling Systems

23:44 - 31:35

The free energy principle is a scientific principle used to construct models of systems that track aspects of their environment.
Modeling involves creating a simplified representation of the environment, like a map that is simpler than the actual territory.
Errors generated by the model are actually relevant signals in active inference and the free energy principle approach.
The model needs to balance fidelity (accuracy) and adaptability (flexibility) to accurately represent the environment.
Entropy is a form of flexibility that needs to be maintained in the model.
The free energy principle minimizes complexity while maximizing predictive accuracy.
It penalizes every new degree of freedom introduced into the model to explain data.
The free energy principle combines statistical predictive accuracy with Occam's razor.
Markov blankets are important in the context of the free energy principle as they define boundaries between what's inside and outside a system.
Markov blankets can be visualized as cell membranes separating internal and external states.

Mathematical Truths and Connections to Physics

31:13 - 39:04

The principles of classical mechanics, free energy principle, and maximum entropy are mathematical truths that govern the behavior of physical systems.
The physical universe conforms to these mathematical regularities.
The free energy principle introduces a new family of mechanical theories that connect classical mechanics, quantum mechanics, and statistical mechanics.
Bayesian mechanics is a physics that connects the thermodynamic entropy of physical states to the information entropy of probabilistic beliefs.
The equations of motion derived from the free energy principle are constrained by both the physics and thermodynamics of the system as well as the representations and beliefs about the rest of the universe.
The free energy principle generalizes the second law of thermodynamics to open systems far from equilibrium.
The Markov blanket allows for movement from equilibrium to non-equilibrium regimes by specifying the interface through which a system couples to its environment or particles.
The Markov blanket provides stability to a system while allowing for flux and movement within it.
The self-identical pattern of the Markov blanket can be observed at all scales of self-organization.

Markov Blankets and Self-Organization

38:47 - 46:24

Self-identical patterns can be observed at all scales of self-organization, exhibiting a fractal quality.
The brain has a Markov blanket structure at multiple scales, from neurons to micro circuits to brain networks.
Markov blankets have properties and features that can be reliably re-identified.
Defining boundaries becomes difficult when zooming in on smaller scales due to vagueness and the presence of empty space.
The concept of a Markov blanket is an oversimplification but serves as a useful idealization.
There are ongoing efforts to study weak or fuzzy Markov blankets using mathematical approaches like the blanket index.
The blanket index quantifies how far from perfect blanketedness a system is and tends to zero as systems increase in size.
In our universe, larger systems tend to generate Markov blankets with probability one.
Most physical systems considered in physics are large enough to exhibit Markov blankets.

Active Inference and Machine Learning

46:15 - 54:12

Most relevant things in a large system will have Markov blankets
As the scale of the system gets larger, blanketness approaches zero
Self-organization and emergence are defined by sparse connections to the rest of the world
The whole is less than the sum of its parts
Engine functions as an organized whole due to constraining its parts' behavior
Active inference is a new way of understanding complex systems
Active inference will be the way of doing machine learning in the future
Artificial intelligence built on active inference starts from an explicit generative model
Generative model describes dependency relations within the system
Markov blanket is part of the generative model that contains dependency relations
Generative models allow for performing inference using variational free energy

Active Inference and Explainability

53:45 - 1:01:41

Active inference involves writing down generative models that allow for performing inference
Generative models provide explainability advantage and are auditable by human users
Active inference formalizes the thermodynamics of information writing onto the boundary
New scale-free extensions to the free energy principle have been developed using quantum information theory
Active inference generates models that are predictably accurate and energy-efficient
The free energy principle is equivalent to the maximum entropy principle, which promotes parsimony in explanations

Maximum Entropy and Free Energy Principle

1:01:17 - 1:09:06

They are effectively the same thing. The principle of maximum entropy is the principle of parsimony in explanation.
The free energy principle (FEP) and maximum entropy are the same principle.
The FEP is the optimal dynamic systems model for a given system, based on current knowledge.
Maximum entropy distributions capture what is known about a system while being as spread out as possible within constraints.
The FEP balances predictive accuracy and complexity.
The FEP can help ensure fairness in models by forcing out irrelevant correlations.
Situations can be defined in terms of their sparseness, which relates to the FEP.
At Versus, they propose a standards-based solution to where priors come from in generative models.
The brain has a hierarchical structure with regular patterns of connectivity, which relates to sparseness and the FEP.
Each layer in the brain provides priors and receives context from adjacent layers, creating a contextualized hierarchy.

Spatial Web Foundation and Responsible AI Development

1:08:54 - 1:17:45

The brain is not one monolithic system, but rather a set of layers that contextualize each other.
Each layer of the brain specializes in encoding specific features of the situation and provides context.
An infrastructure project called the Spatial Web Foundation is being developed to create shared knowledge graphs.
The goal is to build a spatial web or hyper spatial web that reflects the structure of various situations humans deal with.
The hyperspatial modeling language (HSML) and hyperspace transaction protocol (HSTP) are being used to build graphical models and inference methods.
The technology stack combines active inference based AI with explicit generative models and relies on graphical techniques.
A standards-based approach is proposed to consolidate the international community around responsible development of AI technologies.
Avoiding silos and coordinating global efforts can help address potential harms caused by these technologies.

Openness, Transparency, and Ethical Design

1:17:21 - 1:25:47

Openness and transparency are important in the development of AI technologies.
Transparency includes not only the decision-making process but also the transparency of the models used.
Training AI systems with extremely curated datasets may not easily generalize.
Equipping AI systems with the capacity to evaluate itself and identify biases can help mitigate discriminatory bias in data sets.
Active inference technologies can allow AI systems to access and report on their own inferences.
Active inference may become the preferred set of AI technologies if legislation goes through.
Neural nets as currently used are black boxes, while active inference allows for explicit labeling and auditability.
The ethical design of AI systems is crucial, and a responsible, scalable approach is desired.
Profitability can be achieved by providing open standards while offering advanced versions of the technology for commercial use.
The Spatial Web Foundation aims to be a custodian of open standards while being an expert in building using those tools.
The group has gathered core luminaries in active inference research and development, including Carl Friston as chief scientist.

Research Group and Misconceptions about the Free Energy Principle

1:25:20 - 1:33:03

The research group has gathered luminaries in the act of inference tradition, including Carl Friston.
The group is committed to contributing to the public domain and open scientific publication.
They have published several papers and maintain an open-source Python package for active inference technology.
There are misconceptions about the free energy principle and act of inference, such as it being trivial or metaphysical.
The free energy principle is a mathematical physics approach to understanding systems' dynamics.
It is not a theory of everything but provides canonical models for our current knowledge.
The free energy principle directly ties physics to belief updating and inference.

Critiques and Refinements of the Free Energy Principle

1:32:38 - 1:40:36

Some recent literature on the free energy principle has been criticized by early career researchers who may not have had a full understanding of the principle.
The free energy principle is often misunderstood, leading to false statements about it.
The free energy principle is a theory that applies to all things but does not explain everything about all things.
There have been criticisms of the free energy principle, with varying quality in the literature.
One important critique pointed out inconsistencies in the formalization of the principle in earlier years, but those have since been corrected.
The relationship between physics and mathematics can be strained, with physicists borrowing tools from mathematicians and then mathematicians needing to refine them.
Similar to the history of theoretical physics, there is an ongoing back-and-forth process of refining and re-deriving core principles of the free energy principle using more established mathematics.
The development of concepts like the Dirac delta function in physics led to significant work and research in mathematics, similar to how the free energy principle has opened up investigations into generalized coordinates and maximum caliber.

Maximum Caliber and Extensions of the Free Energy Principle

1:40:13 - 1:48:29

Maximum caliber is an extension of maximum entropy that considers the entropy of entire paths throughout a system.
The free energy principle has opened up investigations into generalized coordinates and maximum caliber.
The free energy principle evinces interesting dualities to the space described by Jane's principle of maximum entropy.
There are two main families of application of the free energy principle: density dynamics formulation and path-based formulation.
In the density dynamics formulation, surprise is about how implausible a configuration of states is in a system.
In the path-based formulation, surprise scores the deviation of trajectories over time in a system.
The free energy principle can be seen as a way of writing down the principle of maximum entropy.
Emergence can be tamed using the free energy principle's multi-scale approach and shared generative models.

Ensemble Behavior and Nested Systems

1:48:01 - 1:56:00

Ensemble behavior emerges from the shared generative model and coordinated patterns of behavior.
The Free Energy Principle (FEP) provides a formal approach to modeling nested systems.
The FEP operates on the same objective function at every scale, minimizing free energy.
Systems designed using the FEP can perform inference at different scales using the same objective function.
State inference, parameter learning, and model structure estimation can all be done using the FEP.
Good models persist and leave copies of themselves, while bad models dissipate.
Conscious experience may be weakly emergent rather than strongly emergent.
Qualitative sensations are further inferences according to Andy Clark's notion of bazing.
Consciousness emerges when inference reaches sufficient temporal and counterfactual depth.
A theory of consciousness can be derived directly from the Free Energy Principle.
Consciousness corresponds to an inner Markov blanket or inner screen within a system.

Markov Blankets and Consciousness

1:55:31 - 2:03:06

The Markov blanket can be thought of as a holographic screen that separates the inside of a system from its environment.
The active and sensory states of the Markov blanket correspond to reading and writing onto the screen in a quantum mechanics perspective.
All the classical information needed to describe the coupling between systems lives on the boundary of the Markov blanket.
Systems with consciousness may have an internal markup or screen where classical information is stored.
There is a nested hierarchy of screens in the brain that resonate with each other, leading to the emergence of consciousness.
At the top of this hierarchy, there is a layer that only observes lower layers and doesn't observe itself, resolving the homunculus paradox.
Consciousness is a strong candidate for strong emergence and may be explained by combining philosophical work on Bayesian qualia with computational architectures based on the free energy principle.
The free energy principle is a fundamental discovery that has implications for understanding the physics of mind, similar to Galileo's impact on philosophy and physics.

Physics, Mind, and Intelligence

2:02:45 - 2:05:50

The free energy principle unifies all of reality under the auspices of classical mechanics.
There is no distinction between physics, biology, and the mind; it's all just physics in some sense.
Psychology is slower physics, and culture is even slower physics.
The free energy principle allows us to think about the expression of intelligence in physical systems.
Versus has adopted the approach of building AI systems based on the physics of intelligence.
The whole is radically less than the sum of its parts.