Skip to main content
Back to timeline
arXivSource publication:

Diffractive optical networks stop prescribing output patterns: weak supervision lets handwritten scripts self-organize, cutting training steps by about 82%

Synopsis

Diffractive optical neural networks usually assign each class to a fixed detector region and train the light field to reproduce that preset output. A team at the Singapore University of Technology and Design and the National University of Singapore instead used script-level coarse labels and a modified contrastive objective that constrains only relations among samples, letting optical representations of handwritten Devanagari, Arabic and English characters cluster into three groups on their own. At a shared training endpoint the three-layer network reached 93.4% mean test accuracy against 84.7% for supervised optical networks, and needed about 18% of their training steps to reach a stable 84.5% test accuracy.

Source-provided article image: Optical Representation Neural Networks Enabled by Weakly Supervised Learning
Figure 1 · arXiv · Page 6

Interpretation

The organization of the optical output is now set by relationships in the data rather than by a class-to-location mapping specified before training. Earlier diffractive optical networks specified the desired output intensity distribution before training, with each category tied to a predefined detector region or output location that the light field had to reproduce. With the same network architecture, training dataset, batch size, optimizer and learning rate, the three-layer optical representation network reached 93.4% mean test accuracy at a shared training endpoint, versus 84.7% for supervised optical networks with randomly assigned output locations for the three scripts, over 9 runs per method.

The training objective compares only similarity between samples, pulling characters of the same script together and pushing different scripts apart, without assigning output locations. The modified contrastive objective acts on feature vectors built by partitioning the output intensity into a spatial grid and averaging the intensity in each cell, so the loss constrains only relations among samples. Under a 5×5 partition, the three-layer network needed 467±85 training steps to reach a stable 84.5% test accuracy, versus 2534±309 steps for the supervised network, reported as mean±standard deviation across nine runs, with stability defined as five consecutive evaluations at or above the threshold.

Perspective

Experimental validation centers on the single-layer diffractive network with the 5×5 partition, while three-layer results come from numerical simulation; the 1200 measured intensity distributions give about 84% script-level classification accuracy, against about 91% for the corresponding single-layer numerical model. Experimental results under the 4×4 and 6×6 partitions are reported in the supplementary material, showing that script-dependent semantic organization can also be identified experimentally under different partition configurations.

The drop from numerical to experimental performance likely comes from phase discretization, fabrication deviations, optical misalignment, background scattering and detector noise; a single layer is easier to fabricate and align but offers fewer degrees of freedom for field transformation than three layers, reducing representation capacity and slowing convergence; the semantic-encryption reconstruction reaches pixel accuracies of about 72%, 68% and 87%, affected by variations in the learned representations and downstream decoding errors, and remains a proof of concept.