[Paper Review] Neural Circuit Architectural Priors for Embodied Control

Nikhil X. Bhattasali, Anthony M. Zador+1 more|arXiv (Cornell University)|Jan 13, 2022

Genetics, Aging, and Longevity in Model Organisms4 citations

TL;DR

This paper introduces a biologically inspired artificial neural network architecture—Neural Circuit Architectural Priors (NCAP)—inspired by *C. elegans* locomotion circuits to improve embodied control. By encoding constraints like excitation/inhibition balance, intrinsic oscillators, and sparse connectivity, the NCAP achieves comparable performance to fully connected MLPs with orders-of-magnitude fewer parameters, significantly improved data efficiency, and strong transferability to new body designs.

ABSTRACT

Artificial neural networks for motor control usually adopt generic architectures like fully connected MLPs. While general, these tabula rasa architectures rely on large amounts of experience to learn, are not easily transferable to new bodies, and have internal dynamics that are difficult to interpret. In nature, animals are born with highly structured connectivity in their nervous systems shaped by evolution; this innate circuitry acts synergistically with learning mechanisms to provide inductive biases that enable most animals to function well soon after birth and learn efficiently. Convolutional networks inspired by visual circuitry have encoded useful biases for vision. However, it is unknown the extent to which ANN architectures inspired by neural circuitry can yield useful biases for other AI domains. In this work, we ask what advantages biologically inspired ANN architecture can provide in the domain of motor control. Specifically, we translate C. elegans locomotion circuits into an ANN model controlling a simulated Swimmer agent. On a locomotion task, our architecture achieves good initial performance and asymptotic performance comparable with MLPs, while dramatically improving data efficiency and requiring orders of magnitude fewer parameters. Our architecture is interpretable and transfers to new body designs. An ablation analysis shows that constrained excitation/inhibition is crucial for learning, while weight initialization contributes to good initial performance. Our work demonstrates several advantages of biologically inspired ANN architecture and encourages future work in more complex embodied control.

Motivation & Objective

• To investigate whether biologically inspired neural network architectures can provide useful inductive biases for motor control, improving data and parameter efficiency.
• To develop a Neural Circuit Architectural Prior (NCAP) based on the well-mapped locomotion circuitry of *C. elegans*.
• To evaluate NCAP’s performance in terms of initial performance, asymptotic learning, data efficiency, parameter efficiency, interpretability, and transferability compared to standard MLPs.
• To identify which architectural components—such as sign constraints, weight initialization, or sparse connectivity—are essential for learning and performance in small, structured networks.

Proposed method

• Translating the *C. elegans* locomotion neural circuit into a feedforward artificial neural network with discrete-time dynamics and ReLU nonlinearities.
• Incorporating biological features: intrinsic oscillators (central pattern generators), sign-constrained synapses (excitatory vs. inhibitory), and sparse, modular connectivity.
• Using a simulated Swimmer agent from the DeepMind control suite as the embodiment platform for training and evaluation.
• Training the NCAP using both reinforcement learning (PPO) and evolution strategies (ES) to assess robustness across optimization paradigms.
• Implementing ablation studies to isolate the contributions of weight sharing, sign constraints, initialization, and connectivity sparsity.
• Benchmarking performance against a standard 3-layer MLP with identical architecture but unstructured, dense connectivity and random pruning.

Experimental results

Research questions

• RQ1 Can a neural network architecture inspired by biological neural circuits improve data efficiency and parameter efficiency in embodied motor control?
• RQ2 To what extent do biologically inspired architectural priors like sign-constrained synapses and intrinsic oscillators enhance learning in small, sparse networks?
• RQ3 How does the NCAP architecture compare to standard MLPs in terms of initial performance, asymptotic performance, and transferability to new body designs?
• RQ4 Which architectural components—such as weight initialization, sign constraints, or sparse connectivity—are most critical for successful learning in the NCAP framework?

Key findings

• The NCAP architecture achieved asymptotic performance comparable to that of a standard MLP while using orders of magnitude fewer parameters.
• The NCAP demonstrated dramatically improved data efficiency, learning the locomotion task with significantly fewer environment interactions than the MLP baseline.
• Constrained excitation and inhibition were essential for learning; without proper sign constraints, the NCAP failed to learn under the same training regime.
• Weight initialization played a key role in achieving good initial performance, suggesting that biologically plausible initialization supports faster convergence.
• Weight sharing contributed modestly to data efficiency, while sparse connectivity alone did not yield performance gains unless combined with sign constraints.
• The NCAP transferred effectively to new body designs, demonstrating modularity and generalization beyond the original agent configuration.