The Project

Personalized medicine and cancer morphodynamics, all-optical computing, nanometre thick optical synthesizers, and strain gauge sensing are the applications where the novel extreme photonic learning machines developed in this PELM proposal will be disruptive. Indeed, the usual scheme of a serial or parallel computer based on a reconstituted software is not able to efficiently and effectively address these problems where learning and neuromorphic information processing are instrumental to the solution.

Machine learning (ML) has the main task to infer plausible models to describe the observed data and use the inferred models to make predictions. ML is especially useful for optimisation and performance prediction for systems that exhibit complex behaviours and where analytical models are hard to derive and numerical procedures time consuming. In this framework, photonics emerges as the new pathway for low power, efficient and fast ML implementation. A single hidden layer feedforward neural network with randomly generated hidden nodes and analytically computed output weights is called an extreme learning machine (ELM). A photonic extreme learning machine (PELM) processor is the instrument that, in this proposal, we will:

  1. model and simulate as a universal solver of complex problems,
  2. develop by its implementation on different photonic platforms, and
  3. actually realize by fabricating experimental PELM prototypes, each one specialized to a target application.

Based on the recent developments of new paradigms in ML, and on our breakthrough results – including NP complete optical spin-glasses, novel neuromorphic photonic chip fabrication, liquid resonators, metasurface and biophysics – we will show that this unique platform is suited for exploiting all the potentialities of ML implementation in photonics. The common basis is a theoretical framework where matrices of optical computational nodes constitute the neural network structure for single feed-forward layer or recurrent networks. The different platforms we are going to demonstrate are described by a unique computational algorithm implemented by different physical realizations of the computation nodes (e.g. micro-ring resonators, single nanowires or liquid microspheres). The various platforms will have also different output weights training. This training will be realized by thermal tuning of the micro-ring resonances in the silicon photonic PELM. On the contrary, it will be achieved via the controlled spatial light modulation of the multimode light scattered by the metasurfaces or by the whispering gallery mode spherical droplet resonators in the other PELM. Specifically, we will realize silicon photonics chips, two-dimensional semiconductor nanowire metasurfaces and 3D liquid and biological whispering gallery mode droplet resonators.

Modelling, design and on-demand fabrication will allow applying the PELM concept to these different and complementary technologies where we will demonstrate proof-of-concept applications, ranging from computing and optical function synthesis, to strain gauge sensor and cancer detection or monitoring. With respect to the electronic circuit based ML, PELM - operated with free space diffused light beams or with light signal propagating in massive integrated optical circuits - are faster, cheaper, less power consuming, energetically efficient, “green,” scalable and inherently parallel (either via wavelength multiplexing or via multimode interference).

In addition, light-matter interaction can be used to implement complex pattern recognition protocols that revolutionize the sensor concept by adding the learning by experience concept. As an example, we will demonstrate PELM to sense the mechanical vibrational modes in deformable structures or to monitor the chemical dynamics of cancers and to assess the chemotherapy efficacy while simply illuminating a liquid or a biological tissue.

Quantitative targets of our demonstrators are:

  1. a python/C++ software able to simulate a PELM with 1000 internal nodes and 128 inputs with a growing training set starting from 1000 replications;
  2. a silicon photonic PELM which successfully accomplishes vowel recognition with an accuracy comparable to a conventional 64-bit computer using a fully connected neural network algorithm and benchmarked with respect to the results of;
  3. a metasurface PELM made by nanowire scatterers able to sub-wavelength focusing a 532nm laser beam with a starting transverse size of 1mm;
  4. an optofluidic PELM which realizes a strain-gauge sensor for acoustic sensing at 1 micropascal level with ultrasonic bandwidth higher than 100kHz;
  5. a bio-PELM which discriminates between 3 invasiveness grades of cancer and monitors the chemotherapy efficacy of drugs (e.g. 5-Fluorouracil, Gemcitabine) or the sensitiveness to cell metabolism switch (2-deoxyglucose and metformin) during the nutrient depletion.