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

TL;DR

This work introduces SAPIN, a biologically inspired architecture that combines local Hebbian-like synaptic updates with global structural plasticity, enabling a 2D grid to both compute and reconfigure its topology in pursuit of homeostatic prediction. Grounded in the Free Energy Principle and predictive coding, SAPIN demonstrates that a dynamically reconfiguring substrate can learn control policies without external rewards, as shown by solving CartPole and achieving robust performance after parameter locking. Key contributions include a concrete implementation of simultaneous content and topology optimization, a locking mechanism to stabilize learned policies, and evidence that intrinsic prediction-error minimization can suffice for certain control tasks. The results highlight the potential of embodied, structurally adaptive networks as a proof-of-concept for active inference in physically dynamic substrates and motivate future work on longer-horizon planning and more complex environments.

Abstract

Traditional neural networks, while powerful, rely on biologically implausible learning mechanisms such as global backpropagation. This paper introduces the Structurally Adaptive Predictive Inference Network (SAPIN), a novel computational model inspired by the principles of active inference and the morphological plasticity observed in biological neural cultures. SAPIN operates on a 2D grid where processing units, or cells, learn by minimizing local prediction errors. The model features two primary, concurrent learning mechanisms: a local, Hebbian-like synaptic plasticity rule based on the temporal difference between a cell's actual activation and its learned expectation, and a structural plasticity mechanism where cells physically migrate across the grid to optimize their information-receptive fields. This dual approach allows the network to learn both how to process information (synaptic weights) and also where to position its computational resources (network topology). We validated the SAPIN model on the classic Cart Pole reinforcement learning benchmark. Our results demonstrate that the architecture can successfully solve the CartPole task, achieving robust performance. The network's intrinsic drive to minimize prediction error and maintain homeostasis was sufficient to discover a stable balancing policy. We also found that while continual learning led to instability, locking the network's parameters after achieving success resulted in a stable policy. When evaluated for 100 episodes post-locking (repeated over 100 successful agents), the locked networks maintained an average 82% success rate.

Figures (4)

• Figure 1: Grid Setup. The SAPIN 9x9 grid architecture. The 4 Input Cells (blue) are fixed, as are the 2 Output Cells (orange). The 30 Processing Cells (grey) are randomly initialized and migrate via structural plasticity. The top-right corner depicts the current state of the cart pole environment
• Figure 2: Signal Propagation (Update). Information propagates from the input cells (left) to the output cells (right). The next cell to propagate is the one with the highest absolute activation $|V|$. Nonlinearity is introduced via $\tanh(V)$ and the trigonometric angular weighting. In this visual, the cell that is currently propagating information is shown in pink. There are pink arrows pointing to the nearby cells which it is sending information to. The cells that have already been activated are colored yellow. The cells that have not yet been activated are colored green. The right column depicts the current state of the cart pole environment above a list of the value of each cell, with the value of the input cells in blue and that of the output cells in red. Below this can be viewed the expression for the current equation to propagate value from one cell to the nearby cells.
• Figure 3: Structural Plasticity (Movement). Cells with high long-term prediction error ($|\bar{V}_i - E_i|$) move to a new location. If over-activated, a cell moves away from its main signal source; if under-activated, it moves toward it. The blue arrows depict where a cell moves to. The red arrows depict where a cell was unable to move due to a collision. Order of movement is determined by highest desire. The right column depicts the current state of the cart pole environment above a list of the desire of each cell to move. A horizontal line cuts off the list of cells between those with a high enough desire to move and those with too low a desire to move.
• Figure 4: The Punishment Mechanism. Upon failure, random epicenters (at empty or non-empty cells) generate chaotic signals that propagate through the network, driving a synaptic update intended to associate the preceding states with surprise. In this visual, red arrows are shown from each of the random epicenters and connect to the nearby cells. Then a propagation wave happens like normal. The next cell to propagate is the one with the farthest value form 0, like normal. The first cell is shown in pink.