PhD offer in Grenoble InvenSense & Inria

Title

Constrained machine learning using deep neural networks for embedded sensor systems

CIFRE position, joint between InvenSense Grenoble & Inria Grenoble Rhône-Alpes.

Context

Deep neural network techniques show significant breakthrough for many data processing applications, including in the industrial domain such as speech recognition, face recognition, predictive maintenance, autonomous cars. In general, these networks require quite a large amount of data in the learning phase and a large amount of processing power, sometimes supported by distant machines equipped with powerful GPUs.

InvenSense's fundamental signal processing applications are most of the time designed under tight memory and latency constraints, as running on computing machines like e.g. chips, microcontrollers, or mobile phone host processor. Taking into account these constraints for its motion, pressure and ultrasonic sensing technologies, the design of detection, estimation or classification functions add feature differentiation to the competition. Most of these features are always on and require a particular attention to power.

The promises of deep neural network techniques such as the capability to learn from data, the capability to extract, incorporate and use any source of information present in the signal and deliver the best performance, under a unified and fast to market design process are attractive, but practically, these are still at a potential stage and only for some rare cases they could be demonstrated at a proof of concept level. The need for data for learning phase, in the good enough quantity, their associated labels, the footprint of the derived solution, the confidence in such a black box solution, and more in the design phase the absence of a framework and guidelines, keep these promising techniques far from our applications.

The opportunities are however strong, there are success stories for these techniques, there are industry applications, modern low power targets are evolving and will get ready to process efficiently the derived neural network solutions, the promise of gaining in development time cycle are very attractive, and so, despite the technical breakthrough needed, they remain attractive.

Scope and Goal

The PhD candidate will work closely with engineers to handle a few of company’s signal processing problems using deep neural networks. Applications will be particularly focused on 2D fingerprint ultrasonic MEMs sensors, with design of algorithms like e.g. fingerprint detection, fingerprint denoising, liveness detection or fingerprint matching. Opening to 1D motion or 1D pressure MEMs sensors domain is also likely.

Beyond bringing domain-specific machine learning knowledge - accurate network architecture/deepness, suitable loss function, suitable learning method for a given problem - the thesis will ultimately provide transversal tools, guidance and perspective regarding embeddable networks questions like e.g.:

• What are the impact of runtime power/latency/cost constraints on viable networks architectures and backends? Is it possible to design a meta learning algorithm that explore the possible architectures under given constraints?

• Which degree of confidence should be given to a network learned instance on unseen data? Can some low network footprint constraints intrinsically reduce overfitting probability?

• Should the network worth learn his own features or reuse the human extracted ones? How accurate is the constrained network solution against the baseline human-coded one?

• Do most constrained networks require fewer data or labels? How can data be labelled as sparsely as possible? Is it possible to design an iterative process on which labelling and network learning would be advanced conjointly? Is it possible to take advantage of some partial unsupervised mechanism in the learning process? Can we find rules for sufficient data amount and quality?

The candidate will also actively participate to the advancement of the software package necessary to deploy a learned neural network model to embedded targets.

Student profile

We are looking for outstanding candidates with the following profile:

* Master degree, preferably in Computer Science or Applied Mathematics

* Solid mathematical knowledge, especially statistics, linear algebra, and optimization

* Solid programming skills; the project involves programming in python and PyTorch, R, or similar deep learning framework

* Highly creative and motivated

* Fluent in English, both written and spoken

* Prior knowledge of / experience in deep learning

Contact

Julyan Arbel (firstname.lastname@inria.fr)

http://www.julyanarbel.com/

Location

Grenoble lies in the French alps and offers ideal conditions for skiing, hiking, climbing etc.

Paris can be reached in 3 hours by high-speed train connection.