Emergent learning: neuromorphic photonic computing with accelerated training

Emergent learning transforms a disordered optical medium into a photonic device capable of storage, recognition, and classification of arbitrary memory patterns. First, we show that the intensity at the output of a multiply scattering system can be described by a dyadic matrix, the optical-synaptic matrix, exhibiting the same form as a Hebbian synaptic matrix containing a single memory. Then, we employ emergent learning - an approach inspired by neuroscience - to exploit the vast dictionary of raw memories inherently available within a disordered optical structure, thereby engineering the optical-synaptic matrix to store a user-defined attractor, or tailored memory. Importantly these photonic structures also works as an optical comparators providing an intensity-based measure of the degree of similitude between a query pattern and the stored pattern, realizing an hardware co-localization between memory and optical operator. Our system has an almost infinite hardware capacity of tailored memories/ operators ($\mathcal{M} \sim 10^{60557}$), thus these tailored memories can be then employed as examples to build a classifier hardware based on intensity comparison without the need of additional digital transformation layers. Remarkably, this Photonic Emergent Learning platform is not only flexible and fabrication-free, but also relies primarily on analog processes, thus shifting the computational burden of training from the digital layers to the optical domain reducing the computational cost and enhancing performance.

💡 Research Summary

The paper introduces Photonic Emergent Learning (PhEL), a neuromorphic photonic computing platform that exploits the intrinsic high‑dimensionality of multiply‑scattering media to perform memory storage, retrieval, and classification entirely in the optical domain. By modeling the intensity at each output mode of a disordered medium as Iₒ(S)=S·Fₒ·S†, where Fₒ = fₒ⊗fₒ† is a dyadic matrix, the authors show that Fₒ has the same structure as a Hebbian synaptic matrix storing a single memory. Consequently, each output mode naturally encodes a “raw memory” – a random binary pattern derived from the real part of Fₒ.

The core innovation is the application of an “Emergent Learning” strategy, borrowed from neuroscience, to select a subset Σ of M output modes whose raw memories are most similar to a user‑defined target pattern S*. The selected modes are summed incoherently, yielding an aggregated intensity I_Σ = S·R_Σ·S† with R_Σ = Σₘ Rₘ, where each Rₘ is the binary synaptic matrix of a raw memory. This aggregation constructs a tailored optical‑synaptic matrix that stores the desired attractor d_Σ ≈ S*. When the input S equals d_Σ, the corresponding aggregated mode produces maximal intensity, thereby acting simultaneously as a memory readout and an analog comparator measuring similarity between any query and the stored pattern.

Experimentally, the authors use a spatial light modulator (SLM) and a digital micromirror device (DMD) to imprint binary (+/–1) patterns onto a coherent beam that propagates through a strongly scattering, non‑absorbing medium (e.g., a porous TiO₂ film). A CCD camera records the intensity of O≈65 000 output modes. Simulations and measurements confirm a strong anti‑correlation between the Hamming distance of a raw memory to the target and its intensity, validating intensity as a proxy for similarity. By increasing M, the Hamming distance between the aggregated memory and the target drops dramatically (e.g., below 0.02 for M≈1000), demonstrating effective “noise‑cancelling” through mode selection.

For classification, the authors store P examples per class as tailored memories. During inference, the maximal aggregated intensity across the examples of each class, I_c, is computed; the class with the highest I_c is selected. This scheme eliminates the need for a trained readout layer (ridge regression, back‑propagation) typical of reservoir computing. On benchmark datasets such as MNIST and Fashion‑MNIST, PhEL achieves near‑state‑of‑the‑art accuracies (>99%) while reducing training FLOPs by more than an order of magnitude compared to conventional photonic reservoir computers.

A combinatorial analysis shows that the number of distinct memory configurations scales as O!/(O‑M)!, which for O≈10⁵ and M≈O yields an astronomically large capacity (≈10⁶⁰⁵⁵⁷). This “almost infinite” capacity arises solely from the physical disorder of the scattering medium, requiring no lithographic fabrication or re‑configuration of the scattering matrix.

The paper acknowledges practical challenges: (i) exhaustive intensity scanning to select raw memories can be time‑consuming; (ii) binary modulation limits expressive power compared to continuous‑valued inputs; (iii) scaling to deeper, nonlinear architectures will require additional optical elements. Nonetheless, PhEL demonstrates that a disordered optical medium can be transformed into a programmable, fully analog neuromorphic processor that co‑locates memory and computation, offering dramatic gains in energy efficiency, speed (picosecond light propagation), and hardware scalability. The work opens a pathway toward large‑scale, fabrication‑free photonic neural networks that shift the computational burden of training from digital electronics to the physics of light itself.