🧬 biology

Structural Plasticity as Active Inference: A Biologically-Inspired Architecture for Homeostatic Control

The paper introduces SAPIN, a biologically-inspired computational model that utilizes both synaptic and structural plasticity to minimize local prediction errors, demonstrating its ability to solve the CartPole task through homeostatic control.

Original authors: Brennen A. Hill

Published 2026-02-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Brennen A. Hill

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The "Smart Moving Lego" Brain: A Simple Guide to SAPIN

Imagine you are building a robot, but instead of using rigid metal parts and pre-programmed instructions, you build it out of magical, living Lego bricks.

These bricks don't just sit there; they can change their "strength" (how they talk to each other) and, most importantly, they can actually crawl around the floor to find better positions. This is the core idea behind a new research paper called SAPIN.

Here is the breakdown of how this "living" computer works:

1. The Two Ways the Bricks Learn

In a normal computer or AI (like ChatGPT), the "brain" is like a fixed circuit board. The wires are soldered in place. To learn, the computer just changes how much electricity flows through those fixed wires.

The SAPIN model does two things at once:

The "Volume Knob" (Synaptic Plasticity): Imagine two Lego bricks talking to each other. If one brick says something that surprises the other, they adjust their "volume." They learn to predict what the neighbor is going to say so that they aren't constantly startled. This is like learning to anticipate your friend’s jokes so you aren't always caught off guard.
The "Musical Chairs" (Structural Plasticity): This is the special sauce. If a brick is in a "bad neighborhood"—meaning it’s constantly being overwhelmed by noise or is totally bored and ignored—it doesn't just sit there. It physically moves to a new spot on the grid to find a better "view" of the information. It’s like a person moving to a different seat in a crowded cafe to find a spot where they can actually hear the conversation.

2. The Goal: Finding the "Quiet Zone" (Homeostasis)

Most AI is trained with a "carrot and a stick"—they get a "point" for doing something right (the carrot) and a "penalty" for doing something wrong (the stick).

The SAPIN model is different. It doesn't care about points or punishments. It only cares about predictability.

Think of it like a person trying to sleep. You don't need a reward to tell you to find a quiet room; you just naturally move away from the loud construction site and toward the quiet bedroom because your body wants to minimize "surprise" (noise).

In the experiment, the researchers gave the SAPIN model a "Cart Pole" task (balancing a stick on a moving cart). The model learned to balance the stick not because it wanted a high score, but because a balanced stick is predictable. A falling stick is chaotic and "loud" to the brain. The model moved its "bricks" and adjusted its "volume knobs" until it found the calm, predictable rhythm of a perfectly balanced pole.

3. The "Locking" Trick

The researchers found one funny problem: because the bricks are always moving and changing, the robot sometimes "forgets" how to be good. It’s like a student who learns math perfectly, but then keeps changing their study habits so much that they eventually forget how to add.

To fix this, they discovered that once the robot becomes an expert, you can "lock" the bricks in place. It’s like taking a photo of a perfect formation and freezing it. Once locked, the robot became incredibly stable and could balance the pole almost every single time.

Summary: Why does this matter?

Current AI is very powerful, but it’s "stiff." It lives in a digital world where nothing ever moves.

SAPIN is a step toward "Embodied Intelligence." It suggests that true intelligence might not just come from being smart, but from having a body (or a structure) that can physically reshape itself to better understand the world. It’s a move away from "calculating" answers and toward "growing" solutions.

Technical Summary: Structural Plasticity as Active Inference

Title: Structural Plasticity as Active Inference: A Biologically-Inspired Architecture for Homeostatic Control
Author: Brennen Hill (University of Wisconsin-Madison)

1. Problem Statement

Traditional artificial neural networks (ANNs) rely heavily on global backpropagation, a learning mechanism widely considered biologically implausible due to its violation of the Local Learning Principle (the requirement that synaptic updates depend only on locally available information). Furthermore, most current AI models operate on static architectures where the "wiring" is fixed after training. In contrast, biological systems exhibit both synaptic plasticity (changing connection strengths) and structural plasticity (physical reorganization of cells/morphogenesis) to maintain homeostasis and optimize information processing. The paper seeks to bridge this gap by creating a model that learns both how to process information and where to position its computational resources.

2. Methodology: The SAPIN Architecture

The paper introduces the Structurally Adaptive Predictive Inference Network (SAPIN), a computational model inspired by the Free Energy Principle (FEP) and Active Inference.

Spatial Configuration: The model is instantiated on a $9 \times 9$ 2D grid. It consists of fixed Input Cells (4), fixed Output Cells (2), and a dynamic population of 30 Processing Cells that can migrate across the grid.
Core Principle (Active Inference): The network operates by minimizing local prediction error. Instead of maximizing an external reward, the system seeks to make its sensory inputs conform to its internal expectations (homeostasis).
Dual Learning Mechanisms:
1. Synaptic Plasticity (LTM Update): A local, Hebbian-like rule. Each cell maintains a "learned expectation" ( $E$ ) and "directional strengths" ( $s$ ). The error is the difference between the cell's actual activation ( $V$ ) and its expectation ( $E$ ). This error drives updates to both the expectation and the connection weights.
2. Structural Plasticity (Movement): Driven by "desire," defined as the long-term average prediction error. If a cell is chronically over- or under-activated, it physically migrates to a new coordinate on the grid to find a more predictable input stream.
Signal Propagation: Information flows through the grid via a "winner-takes-all" propagation wave. The cell with the highest absolute activation propagates its signal to neighbors, with signal strength attenuated by Manhattan distance and weighted by the sender's directional biases.

3. Key Contributions

Integration of Morphogenesis and Inference: SAPIN is one of the few models to treat the physical topology of the network as a dynamic, learnable parameter driven by the same error-minimization objective as the synaptic weights.
Biologically Plausible Credit Assignment: By utilizing local prediction errors and spatial movement, the model avoids the "weight transport problem" and the need for global error signals inherent in backpropagation.
Self-Sufficient Control: The model demonstrates that complex control tasks can be solved using only intrinsic drives (minimizing surprise) without any explicit external reward or punishment signals.

4. Results

The model was validated using the CartPole-v1 reinforcement learning benchmark.

Task Success: SAPIN successfully learned to balance the pole, achieving the 500-step success threshold within the first 10 episodes.
The Stability-Plasticity Dilemma: The authors found that continuous learning led to instability (a successful agent might fail immediately after succeeding). However, by "locking" the network (disabling plasticity once a goal is met), the agents maintained a robust 82% success rate over 100 subsequent episodes.
The Role of Punishment: Interestingly, the researchers found that adding explicit "punishment" (chaotic signals upon failure) did not significantly improve performance. This suggests that the intrinsic drive to minimize prediction error is a sufficient objective function for homeostatic tasks, as the "success state" is naturally the most predictable state.

5. Significance

This research provides a computational proof-of-concept for embodied intelligence. It suggests that intelligence can emerge from the interaction between functional learning (synapses) and structural adaptation (morphology). By demonstrating that an agent can solve a control problem purely through the minimization of local surprise, the paper reinforces the Free Energy Principle as a viable framework for both biological and artificial intelligence. It moves the field toward "self-organizing" architectures that do not require pre-defined layers or global optimization algorithms.