Machine Learning at the ES Group

Introduction

Machine learning and in particular deep learning has dramatically improved the state-of-the-art in object detection, speech recognition, robotics, and many other domains. Whether it is superhuman performance in object recognition or beating human players in Go, the astonishing success of deep learning is achieved by deep neural networks trained with huge amounts of training examples and massive computing resources. These data and computing requirements combined with the excessive energy consumption of DL networks pose challenges for successful applications in embedded systems. The Electronic Systems group aims to improve the computational efficiency (in operations/Joule) of machine learning algorithms and the circuits used to implement these algorithms.

Approach

ES group focusses on computational efficiency of machine learning, considering all aspects which potential lead to lower energy and more performance:

  • Efficient inference/learning algorithms and inference/learning networks
  • Efficient implementation (mapping of application to processing platform), e.g. by
    • quantization of data and weigths
    • advanced loop transformations
    • use of Halide as high level specification and implementation language
  • Efficient inference/learning processing architectures
  • Efficient inference/learning circuits

Past and current projects

SPITS: Traffic sign and Face recognition with Deep Convolutional Neural Networks

CNNs: Efficient Architectures and Implementations for Convolutional Neural Networks

  • FPGA, GPU and multi-core implementations
  • Flexible Full custom accelerators
  • Compiler for translating CNN description into VLIW, software pipelined, code for the CNN network accelerators
  • Advanced tiling optimizations
  • CNN network optimizations

ZERO-P2: Autonomous Acoustic Systems

  • Smart acoustic monitoring of cities
  • Smart autonomous hearing aids
  • Low power processors for PGM: Probabilistic Graphical Models for Bayesian learning

ZERO-P3: Autonomous Roadside Monitoring

  • Interpretation of radar images with deep learning
  • Self-learning algorithms for object (visual) recognition
  • Data fusion of vision and radar
  • Design of heterogeneous architectures supporting vision, radar and learning

BrainWave

  • Machine learning for the classification of EEG signals, used for
    • Early seizure detection for Epilepsy patients
    • Helping Parkinson patients to avoid freezing of gait
  • Low power processing architectures, based on CGRAs (Coarse Grain Reconfigurable Arrays) for EEG signal classification
    • Including: Low power circuits and VLSI design for EEG processing

PAC: Platform-aware Compilation

  • Modeling the impact of loop transformations like tiling, loop interchange and fusion, in the context of Deep Learning
  • Using Halide for design space exploration of many optimizations for Deep Learning algorithms
  • Using Halide as a target language for DL program generators (with a DL Network description as input)

SIMD architectures for Deep Learning

  • Compiler techniques exploiting auto-vectorizations and other loop transformations

oCPS: Optimizing Cyberphysical Systems

  • Using approximate neural networks for very low power implementations of loop nests

Silence

  • HW support for gesture recognition using SVMs

VSM: Vital Signs Monitoring

  • Detecting living skin for vital signs monitoring using SVM classifiers

Future projects and activities

EDL Perspectief program (starts fall 2018)

  • Creating a huge Dutch ECO system around machine learning
  • Joining Dutch academic computing and the Dutch academic learning community, together with many industrial players.

Special sessions at conferences organized by ES

  • AMDL session at DSD 2018 on Applications, Architectures, Methods and Tools for Machine- and Deep Learning
  • Embedded Deep Learning topic at DATE 2019
  • Special day on model-based design of intelligent systems at DATE 2019

New projects, researching

  • Tuning our CGRA for Deep Learning applications
  • Highly quantized DL networks, up till single bit weights and inputs, for extreme efficiency

2019: New course on Efficient Machine Learning

  • Treating both algorithms for ML and DL, and very efficient implementations and realizations

Many Master and BSc assignments and possible internships in the machine learning area, see e.g. the student projects on parse.ele.tue.nl

Applications

The Electronic Systems group has active research collaborations with companies on machine learning algorithms and circuits for machine learning. This includes the following companies: