AALOK

Analog Adaptive Learning via Optical Kernels

A single free-space photonic extreme learning machine that learns images, audio, and tabular data in one fixed optical pipeline — light does the feature map, only the readout is trained.

96.56 95.67 100.0 0.0699 Abalone NRMSE

Read the paper Open the encoder View the code

embedding mask · Fourier carrier

The optical pipeline

One apparatus. Every modality.

Data becomes a phase pattern on a spatial light modulator, free-space propagation does the high-dimensional mixing, an iris keeps only the informative first diffracted order, and a camera reads out 4096 intensity features. Nothing in the optics is trained.

◉
Laser
532 nm
coherent CW

▦
Phase SLM
Holoeye · 1920×1200
8 µm · φ = x + W

◇
4f + iris
Fourier mixing
first order only

▣
CMOS camera
|u|² intensity
4096 features

Σ
Ridge readout
trained layer
the only weights

Measured performance

Four modalities, one fixed reservoir.

Test performance with the Fourier embedding after independent ridge-λ sweeps. To our knowledge this is the first free-space PELM spanning image, audio-derived, and tabular tasks in a single physical pipeline, and the first with spectrogram-based spoken-digit classification.

Modality Task Best λ Test

MNIST 10-class handwritten digits 7.28×10⁻⁴ 96.56
FSDD Spoken digits · log-Mel spectrogram 1.89×10⁻³ 95.67
Mushroom Binary tabular · edible vs poisonous 1.0×10⁻⁵ 100.00
Abalone Tabular regression · shell-ring age 4.89×10⁻³ 0.0704 NRMSE

Multimodal Optical Feature Extraction with a Free-Space Photonic Extreme Learning Machine.
Kumari, Khisti, Choube, Satra, Murali & Kumar. arXiv:2605.29043
Laboratory of Optics of Quantum Materials · Department of Physics · IIT Bombay
Code & hardware: github.com/loqm/aalok

AALOK आलोक · light, illumination, to see
A LOQM optical-computing platform

Modality	Task	Best λ	Test
MNIST	10-class handwritten digits	7.28×10⁻⁴	96.56
FSDD	Spoken digits · log-Mel spectrogram	1.89×10⁻³	95.67
Mushroom	Binary tabular · edible vs poisonous	1.0×10⁻⁵	100.00
Abalone	Tabular regression · shell-ring age	4.89×10⁻³	0.0704 NRMSE