Title: Lambda Inference Machine: accelerating computing by nonlinear evolution of spectrally modulated data

Abstract: We show that nonlinear spectral dynamics that underpin fascinating phenomena in nature can be leveraged to accelerate machine learning by transforming a difficult nonlinear classification problem into a simple linear separation task. This is done by seeding the nonlinear dynamics with the data that is to be classified and passing the output to a digital linear classifier. Data is first modulated onto the spectrum of a femtosecond pulse that is allowed to evolve through a nonlinear optical medium followed by a realtime spectrometer and the data acquisition. This unsupervised process projects the linearly non-separable data to a space where data become linearly separable. We validate this hypothesis by performing a nonlinear classification benchmark task followed by the classification of various datasets including brain intracranial pressure time series, time-stretch microscopy images, and spoken digit recognition. Since the data is first modulated onto the optical spectrum, this technique is fully compatible with the time stretch technique for ultrafast single-shot data acquisition and classification.