Active Inference Demonstrated with Artificial Spin Ice

Robert L. Stamps,Rehana Begum Popy,Johan van Lierop

2024-02-27

Abstract:A numerical model of interacting nanomagnetic elements is used to demonstrate active inference with a three dimensional Artificial Spin Ice structure. It is shown that thermal fluctuations can drive this magnetic spin system to evolve under dynamic constraints imposed through interactions with an external environment as predicted by the neurological free energy principle and active inference. The structure is defined by two layers of magnetic nanoelements where one layer is a square Artificial Spin Ice geometry. The other magnetic layer functions as a sensory filter that mediates interaction between the external environment and the hidden Artificial Spin Ice layer. Spin dynamics displayed by the bilayer structure are shown to be well described using a continuous form of a neurological free energy principle that has been previously proposed as a high level description of certain biological neural processes. Numerical simulations demonstrate that this proposed bilayer geometry is able to reproduce theoretical results derived previously for examples of active inference in neurological contexts.

Mesoscale and Nanoscale Physics,Disordered Systems and Neural Networks,Biological Physics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **How to use the Artificial Spin Ice (ASI) structure to achieve active inference, and experimentally study the dynamic behavior of this non - biological system**. Specifically, the author hopes to build a double - layer nano - magnetic component model to show how thermal fluctuations drive the evolution of this magnetic spin system under the dynamic constraints of the external environment, thereby verifying the free - energy principle and the active inference theory.

### Main problems and goals

1. **Verify the applicability of the active inference theory in non - biological systems**:
- The author hopes to prove that the active inference theory, which was originally used to describe the information processing and response mechanisms of biological neural systems, can also explain and predict the dynamic behavior of non - biological systems (such as artificial spin ice).

2. **Explore new research methods for complex non - equilibrium dynamics**:
- By studying the active inference process in the artificial spin ice system, the author hopes to provide new experimental and theoretical tools for the study of complex non - equilibrium dynamics.

3. **Develop a new physical neuromorphic computing platform**:
- If active inference can be implemented in physical devices, this will make it possible to develop a new neuromorphic computing platform, especially having potential application value in machine learning and data processing.

### Research background and motivation

- **Active inference and free - energy principle**: In recent years, a theoretical framework called "active inference" and "free - energy principle" has been proposed in the field of neuroscience to describe how the brain processes and responds to information. This theory holds that the brain minimizes free energy by comparing predicted and perceived information and adjusting future perceptual inputs according to the differences.

- **Applications of artificial spin ice**: Artificial spin ice is a geometric structure composed of nano - magnets, which was initially used to study complex frustrated systems. Recent research shows that artificial spin ice can also be applied to neuromorphic computing and machine learning.

### Main contributions of the paper

- **Construct a double - layer artificial spin ice model**: The author proposes a double - layer structure, where one layer is a traditional square artificial spin ice geometric structure, and the other layer serves as a sensory filter, mediating the interaction between the external environment and the hidden artificial spin ice layer.

- **Simulate and verify theoretical predictions**: Through numerical simulation, the author shows that the spin dynamics of the double - layer structure can be accurately described by the continuous form of the free - energy principle and can reproduce the active inference results previously derived in the context of neuroscience.

- **Potential for experimental research**: The author points out that experimentally studying this double - layer artificial spin ice system can provide new research methods for complex non - equilibrium dynamics and the possibility of developing a new physical neuromorphic computing platform.

### Key formulas and concepts

- **Free - energy principle**:
\[
E(\tilde{\mu}, \tilde{\phi}) = -\frac{1}{\beta} \log P(\mu, \phi)
\]
where \(E(\tilde{\mu}, \tilde{\phi})\) is negative entropy, and \(P(\mu, \phi)\) is the joint probability distribution of hidden states and perceived states.

- **Variational Bayes relaxation equation**:
\[
\frac{d\mu_\alpha}{dt} - D\mu_\alpha = -\kappa \frac{\partial E(\tilde{\mu}, \tilde{\phi})}{\partial \mu_\alpha}
\]
This equation describes the time evolution of the hidden state \(\mu_\alpha\), making it tend to the direction that maximizes the joint probability distribution.

- **Magnetization response**:
\[
M_\alpha = \langle \chi(B_\alpha) \rangle_T
\]
where \(M_\alpha\) is the magnetization measured in a certain direction, and \(\chi(B_\alpha)\) is the magnetization response under the field \(B_\alpha\).

Through these formulas and...