TLDR: This research paper explores the use of memristor-based neural network accelerators for AI applications in space, addressing critical challenges like energy efficiency and radiation robustness. It introduces techniques such as bit-slicing, temporal averaging, and periodic activation functions (SIRENs) to mitigate performance degradation caused by memristor non-idealities like device variability and conductance drift. Through simulations on guidance and control networks (G&CNETs) and asteroid geodesy networks (GeodesyNets), the study demonstrates that these methods can restore competitive performance, bringing memristor-based AI closer to practical deployment for autonomous space missions.
Artificial intelligence (AI) is set to transform space missions by enabling spacecraft to operate with greater autonomy. However, deploying AI in space comes with significant challenges, including strict limits on energy consumption, extreme temperatures, and the need for radiation robustness. Traditional AI accelerators, often based on GPUs and CPUs, struggle to meet these demands due to issues like the Von Neumann bottleneck, which involves high energy costs for moving data between memory and processing units, and their vulnerability to radiation.
To overcome these hurdles, researchers are exploring alternative technologies. One promising area is neuromorphic hardware, which mimics the brain’s architecture to bypass the Von Neumann bottleneck. Among these, memristors stand out as an emerging technology for in-memory computing, where data storage and computation are integrated into a single element. Memristors offer high energy efficiency, non-volatility (they retain data without power), and inherent radiation resilience, making them ideal for space applications.
Despite their advantages, memristive devices are not without their flaws. They suffer from non-ideal behaviors such as device variability, where individual devices behave differently; conductance drifts, where their resistance changes over time; and device faults, which can severely degrade the performance of neural networks. This paper, titled “Memristor-Based Neural Network Accelerators for Space Applications: Enhancing Performance with Temporal Averaging and SIRENs,” explores how to mitigate these issues to make memristor-based AI viable for critical space tasks.
The research, conducted by Zacharia A. Rudge, Dominik Dold, Moritz Fieback, Dario Izzo, and Said Hamdioui, focuses on improving the reliability and precision of memristor-based neural networks through a simulation-based approach. They investigated several mitigation strategies: bit-slicing, temporal averaging of neural network layers, and the use of periodic activation functions, specifically SIRENs (Sinusoidal Representation Networks).
Bit-slicing involves using multiple memristor devices to represent a single weight in the neural network, effectively increasing precision. Temporal averaging, on the other hand, calculates a layer’s output multiple times and averages the results, reducing the impact of random noise and variability. SIRENs, with their periodic activation functions, have shown improved performance in certain tasks compared to traditional activation functions.
The study applied these techniques to two challenging on-board space applications: Guidance and Control Networks (G&CNETs) for spacecraft navigation and control, and GeodesyNets for estimating the shape and mass distribution of asteroids. Both tasks require high precision and are representative of real-world space mission needs.
For G&CNETs, the researchers found that using SIRENs significantly improved initial performance. However, these networks proved highly sensitive to device faults and conductance drift, meaning even small imperfections could lead to substantial performance degradation. Interestingly, bit-slicing and temporal averaging had only a minor impact on G&CNETs, suggesting that the choice of activation function and the network’s inherent sensitivity to weight perturbations (quantified by its Lipschitz constant) played a more critical role.
In contrast, GeodesyNets showed a different behavior. Without bit-slicing or temporal averaging, the network completely failed to learn the asteroid’s density field. However, by implementing these mitigation techniques, especially with a combination of 4 slices and 64 repeats for temporal averaging, the GeodesyNets achieved performance levels comparable to digital baselines. These networks also demonstrated better robustness to device faults and conductance drift, likely because the task involves integrating over many points, which naturally averages out local noise.
The findings highlight that while memristors hold immense potential for energy-efficient and radiation-robust AI in space, careful design and mitigation strategies are crucial. The effectiveness of these strategies can vary significantly depending on the specific neural network architecture and the application. The research successfully demonstrated that memristor-based neural networks can achieve competitive performance for on-board space tasks, bringing them closer to practical deployment.
Also Read:
- Action Chunking with Transformers: A Data-Efficient Approach for Spacecraft Guidance
- Training Biological Brains in Virtual Worlds: A New Approach to Organoid Intelligence
Future work will focus on energy estimation, exploring different network architectures, and ultimately, designing and fabricating actual hardware to bridge the gap between simulation and reality. This research provides a promising outlook for unlocking the full benefits of memristors for AI in space, enabling more autonomous and capable missions. You can read the full research paper here.
