Russian Scientists Add “Noise” to Neural Networks
Researchers in Saratov have found that physical noise can make neural network training more robust and reliable.
Scientists at Saratov National Research University named after Chernyshevsky have studied the role of noise in neural networks. The focus was on thermal fluctuations in signals – disturbances that affect all physical electronic devices. The team found that such noise can actually make neural networks more reliable and stable, TASS reported, citing university researchers.
Most neural networks today operate digitally on computers or graphics processors, a computationally demanding approach. Researchers have been exploring hardware neural networks, where neurons and connections are implemented directly at the electronic level. These systems, however, are more susceptible to noise, raising concerns that effective training might be impossible.
Toward a New Generation of AI Hardware
In experiments, scientists introduced Gaussian noise into neurons while training AI systems to recognize images and complex signals. The results showed that exposure to noise during training increased the networks’ resilience to it.
The approach could pave the way for energy-efficient next-generation AI hardware, designed for image processing, signal analysis, and computation in resource-constrained environments.
