Russian scientists have developed an approach that allows training hybrid neural networks using alternative computing systems — photonic, neuromorphic, and analog chips. This was reported by the press service of the Artificial Intelligence Institute AIRI.
Today, such chips solve only individual tasks and cannot fully take on the training of neural networks. The solution lies in hybrid systems: classic microchips perform the main part of operations, while the most frequent computations are handled by a specialized co-processor. Researchers from AIRI, Sberbank's Quantum Technologies Center, and Skoltech have created an algorithm that trains such networks even without access to the co-processor's internal gradients. The approach was tested on six digital models of photonic layers and three classes of tasks.
Calculations showed that precise knowledge of what happens inside each component is not required. Observations of the device's reaction to random changes, along with an approximate co-processor model, yield results close to fully digital baseline models. In the future, hybrid systems will find application in data centers for training large models and in devices with strict requirements for speed and energy consumption.