Fig. 1. Schematic sketch of the ONN system, which follows the photonic extreme learning machine architecture [21]. The input image is encoded by an SLM and passes through hidden layers made of a scattering medium. The ONN outputs are captured by six independent solar-cell pixels and are then multiplied with a layer of trained weights to predict classification results.

Fig. 2. Diagram illustrating the layered perovskite solar-cell design (left) and a 3D render of a six-pixel fabricated sample (right). Active regions indicate areas where the device composition as illustrated on the left is complete; therefore only light reaching the active regions will be converted into electric signals through photovoltaic effects. The top and bottom rows of gold contacts are shorted directly to the ITO conductor. They thus do not contain an active region and are used as ground electrodes.

Fig. 3. The graph shows the characterization results of the fabricated perovskite solar-cell samples, fitted with a logarithmic function.

Fig. 4. 3D render of the ONN optical setup, with a six-pixel solar-cell panel as its detector.

Fig. 5. Violin plot that compares the classification accuracy between the single-layer random ONN setup with solar cells or a CCD camera. The number of classes indicates the number of different MNIST digits the ONN aims to classify. Each violin illustrates the distribution of the accuracy evaluated over all of the possible combinations of a fixed number of classes. The number and the white line annotate the median accuracy of these combinations. The thicker and thinner lines indicate the interquartile range and the 1.5 times the interquartile range of the accuracy distribution, respectively. Each violin is independently normalized by its area.