# EPSRC Proposal ## Reasoning and Formalising Large-Scale Cross-Platform Machine Learning ### Introduction * Machine learning is popular * Large-scale machine learning requires cross-platform * Inference is still single-platform * People try to fit software into a fixed hardware architecture (?) ### Background * Hardware Platform for Inference * Paul Chow's Alveo Cluster * EVEREST * RIKEN * Machine Learning Compiler * FPGAConvNet from ONNX * HeteroCL from TensorFlow * CIRCT from PyTorch * Cross-Platform Compiler * Paul Chow's tool flow * ... ### Research Objectives/Challenges We aim to answer the following questions from this project: * How to define both software program and hardware architecture in the same software model? * How to model multiple platform instances in a unified system? * Efforts made for formalisations of individual instances * But these formalisations cannot be directly adapted into a unified system * => Build a unified abstraction for cross-platform computing * How to formalise the problem of mapping ML models onto the system? * Hardware mapping has been formalised dedicated for single instances locally * No global performance model for across platform * => Build reasoning enginee for the hardware efficiency from the abstraction * How to reason the hardware efficiency based on the formulation? * Can we evaluate the hardware efficiency using the limits given by our model? * Can we identify the system bottleneck and optimise the solution with the help of our model? * => Demonstrate a few ML applications adapted by the model ### National Importantce * Providing theoretical limits for industrial engineers and guiding them to improve their implementation by automated reasoning ### Contribution to Knowledge * Help us to understand the theoretical limits of cross-platform ML and guides designers/tools to optimise the hardware efficiency. ### Possible Collaboration * We will focus on the formalisation of abstraction and reasoning * VAST group at UCLA will help on mapping a series of applications onto the hardware through the abstraction