Technical Summary: Brain-like Adaptive Dendritic Memristive Networks

Problem Statement

Current artificial intelligence systems rely on learning algorithms that adapt connection weights within fixed-topology hardware architectures. This approach decouples the learning process from the physical evolution of the computing substrate, preventing experience from directly reshaping the computational architecture. While memristive devices have been utilized as hardware accelerators for deep neural networks (e.g., via in-memory matrix-vector multiplication), they typically operate within rigid, pre-fabricated topologies. Even self-organizing memristive networks used in physical reservoir computing generally maintain fixed connectivity, with learning occurring externally through readout weights rather than through the physical adaptation of the network structure itself. There is a lack of hardware architectures that physically encode experience through the co-evolution of morphology and functional connectivity within the computing substrate.

Methodology

The authors present a self-organizing, morphologically adaptive network based on dynamically interacting electrochemical dendrites. The system utilizes an electrochemical cell composed of Gold (Au) electrodes acting as neuron terminals and a biocompatible Silver Nitrate (AgNO3) in Dimethyl Sulfoxide (DMSO) electrolyte.

Core Mechanisms:

1. Structural Plasticity (Morphogenesis): The system mimics synaptic morphogenesis and axon growth. Temporally correlated voltage pulses (bipolar pulses) applied between terminals drive competitive half-cell reactions: reduction ($Ag^+ + e^- \rightarrow Ag$) on the negatively biased electrode and oxidation ($Ag \rightarrow Ag^+ + e^-$) on the positively biased electrode. This results in the net growth of metallic dendritic branches that can structurally connect terminals, forming electrical junctions (synapses). The growth dynamics are tunable between isotropic (dendritic-like) and directional (axon-like) behaviors by adjusting stimulation parameters (pulse amplitude, inter-pulse interval) and electrolyte concentration.
2. Functional Plasticity (Memristive Dynamics): Once formed, dendritic connections exhibit memristive behavior. They display bipolar resistive switching (SET/RESET) typical of electrochemical metallization (ECM) cells, where conductive filaments thin/rupture or form/strengthen based on applied voltage. The system supports both volatile (short-term) and non-volatile (long-term) memory states, depending on the stimulation history and operating conditions.
3. Modeling: A two-dimensional physics-based stochastic model was developed to simulate the system. It integrates local electric potential at the metal-electrolyte interface with ion diffusion transport in the electrolyte, accounting for capacitive effects of the electrical double layer. The model uses a reaction-diffusion scheme with operator splitting to capture the interplay between diffusion-limited growth and potential-driven directional growth.

Experimental Setup:

• Fabrication: Two-terminal cells were fabricated on Si wafers with Au electrodes; multiterminal systems utilized commercial microelectrode arrays (MEAs) with patterned Au pins.
• Stimulation: Synchronous stimulation and acquisition were performed using an ArcTWO measurement board with 64 parallel source-measure units (SMUs).
• Readout: Conductance matrices were acquired by applying low-amplitude read voltages (50 mV) to terminal pairs while keeping others floating.

Key Contributions and Results

1. Co-evolution of Structure and Function

The study demonstrates that structural connectivity (the physical wiring diagram) and functional connectivity (synaptic weights) co-evolve in response to external stimulation. The system embeds spatiotemporal input patterns directly into the network's physical morphology.

• Result: Multiterminal networks structurally evolve to form connections between terminals with spatio-temporally correlated activity, effectively reducing the topological distance between co-activated nodes.

2. Implementation of Associative Learning Paradigms

The system implements biological associative learning principles directly in the material ($in\ materia$):

• Classical Conditioning (Pavlovian): The system was configured to mimic Pavlov's dog. A pre-programmed high-conductance connection represented the unconditioned stimulus (food) to response (salivation). By synchronizing stimulation of a neutral stimulus (bell) with the unconditioned stimulus, the network grew a new dendritic connection bridging the "bell" and "salivation" terminals. Post-conditioning, the "bell" stimulus alone triggered the response current, demonstrating the formation of a conditioned association.
• Operant Conditioning (Skinnerian): The system replicated Skinner's pigeon pecking and rat punishment tasks.
  • Pigeon Pecking: Rewarding inputs increased conductance (pecking rate). During extinction (no reward), conductance dropped, but the underlying dendritic topology remained largely intact ("savings"). Reconditioning rapidly restored and potentiated the conductance, faster than initial conditioning.
  • Rat Punishment: Punishment stimuli (shorter negative pulses) were delivered, eventually breaking the electrical connection. However, no significant variations in the dendritic topology were observed, maintaining the learned information within the dendritic structure. Reconditioning again showed rapid recovery and potentiation, confirming that the "memory" was not deleted but structurally preserved despite the functional break in the connection.

3. Solving Optimization Problems (Path Selection)

The authors demonstrated the system's capability to solve a route-planning optimization problem relevant to robotic navigation.

• Method: A terrain map with varying traversal costs (e.g., sand, rocks) was mapped to a spatiotemporal stimulation protocol. Electrodes represented terrain regions, and voltage amplitudes encoded traversal difficulty.
• Result: The dendritic growth process acted as a physical sparsifier, reinforcing low-cost conductive routes and suppressing high-cost regions. The resulting conductive subnetwork embodied the optimal path.
• Performance: Using Dijkstra's algorithm on the extracted conductance matrix, the system achieved an average success-weighted overlap of 85% with the ground-truth optimal path across 25 independent runs.

Significance and Claims

The paper claims to establish a route toward embodied intelligence at the matter level. By bridging information processing, learning, and memory encoding with adaptive materials and physical network dynamics, the work demonstrates that:

1. Learning is Physical: Learning is not merely the optimization of parameters in a fixed graph but the physical embodiment of experience through the co-evolution of structure and functional connectivity.
2. In-Materia Computation: Computation emerges intrinsically from material self-organization, where information actively shapes the computational structure through which it is processed.
3. Biological Fidelity: The system successfully reproduces complex biological learning phenomena, including classical and operant conditioning, savings, and rapid reconditioning, using electrochemical mechanisms that emulate synaptic morphogenesis and plasticity.
4. Optimization via Morphology: The hierarchical, fractal nature of the dendritic structures facilitates a balance between information segregation and integration, enabling the physical system to solve combinatorial optimization problems (like pathfinding) by transforming graph constraints into physical conductive pathways.

The authors position this work as a departure from top-down fabricated hardware, proposing a bottom-up approach where the computing architecture itself adapts during learning, offering a potential pathway for next-generation adaptive hardware that integrates processing, memory, and learning within a single physical system.