Figure 1 | Training process of the B-ONN framework. (IMAGE)

Light Publishing Center, Changchun Institute of Optics, Fine Mechanics And Physics, CAS

Figure 1 | Training process of the B-ONN framework.

Caption

Figure 1 | Training process of the B-ONN framework. B-ONN adopts a layer-wise decoupled synchronous training mechanism. Each layer maintains two sets of trainable parameters: a forward optical layer for feature transformation and an error layer for error back-construction. During training, after forward propagation, the error layer of each layer uses the local error signal from the subsequent layer to generate the target optical field for the current layer via target propagation mechanism. Subsequently, the local error between the actual output and the target optical field is computed synchronously within the layer via optical interference. This error directly and independently guides the update of both parameter sets in the layer, thereby achieving localized learning that is equivalent to the global backpropagation algorithm but physically more feasible to implement.

Credit

Min Gu et al.

Usage Restrictions

Credit must be given to the creator.

License

CC BY

Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.