Toward Bio-Inspired, Energy-Efficient Neural Computers: Integrating Spiking Neural Policies for Generalized I/O Alignment

TL;DR: What if we made Neural Computers more like brains by using spiking neural networks to handle input and output, making them more efficient and better at dealing with real-world noise? An initial experiment could compare spiking NCs and conventional NCs on multitask CLI/GUI video modeling benchmarks, measuring their energy use, robustness to noisy I/O, and ability to generalize interface primitives.

Research Question: Can spiking neural network architectures, inspired by biological computation, enhance the energy efficiency, robustness, and I/O alignment capabilities of Neural Computers, especially in the presence of noisy or unpredictable input streams?

Hypothesis: Incorporating event-driven spiking neural policies into NC architectures will enable better energy efficiency and improved resilience to input/output noise, while maintaining or exceeding the I/O alignment and control capabilities of conventional ANN-based NCs.

Experiment Plan: Develop a spiking NC architecture (e.g., using leaky integrate-and-fire neurons and surrogate gradient training, as in Mraidi et al., 2025). Benchmark this architecture against standard ANN-based NCs on the CLI/GUI video modeling task, but with injected real-world noise (e.g., sensor jitter, missing frames). Measure I/O alignment accuracy, energy consumption (using simulation or hardware estimation), and convergence rates. Assess performance in sim-to-real transfer settings, observing behavioral robustness under noisy, real-world-like conditions. Support/refute hypothesis by comparing the trade-offs in efficiency and stability.

References:

• Mraidi, A., Gyöngyössy, N. M., & Botzheim, J. (2025). Evaluating Spiking Neural Networks in Reinforcement Learning for Robotic Navigation. IEEE International Conference on Services Computing.