DIPVIEW

The objective of the project is the creation of a start-up that aims to offer products and services to optimise and accelerate deep learning on parallel and heterogeneous architectures. This will be achieved by implementing a recommender that reduces the energy cost/carbon footprint of training deep learning models.

The Dipview project consists of two main components:

• The Auto-CNNp (Automatic CNN parallelization) framework based on software components that automate the process of training convolutional neural networks (CNNs).

• A multi-criteria recommender for the ideal configuration of resources for deep learning.

• Automation of the training parallelization of CNNs
• Significant acceleration of the training process
• Decrease in AI frugality

A proof of concept (POC) of the Auto-CNNp solution on medical applications (COVID,...) has been implemented, evaluated and validated over TensorFlow and Horovod.

• Medical for the diagnosis of various pathologies (brain tumours)
• Industrial Robotics
• Autonomous cars
• Training and retraining in deep learning