Faites connaître cette offre !
Reference : UMR6072-FREJUR0-011
Workplace : CAEN
Date of publication : Thursday, September 05, 2019
Type of Contract : FTC Scientist
Contract Period : 12 months
Expected date of employment : 1 November 2019
Proportion of work : Full time
Remuneration : 30 to 35k € gross annual according to experience
Desired level of education : PhD
Experience required : Indifferent
Our main objective in this project is to develop methods to learn "lean" deep models that are able to maintain generalization performance, while significantly decreasing the computational resources and capable of dealing with few supervision by exploring unsupervised modeling. Massive unlabeled data that are available for free in many Vision applications are usually considered as a green context for learning in the sense that they do not require human annotations. The idea of using unsupervised methods that fits with supervision is not new (see eg. The work of Suddarth and Kergosien, 1990). The basic criterion for unsupervised learning is the reconstruction (or generation) of the input. For instance, with an auto-encoder for natural images, the decoder will try to reconstruct the original input (with all details) from the internal representation. This approach, however, might not be optimal for tasks such as image classification where pixel-level details are not relevant, since such models should actually be invariant for many low-level imaging conditions. The more recently proposed ladder network (see Valpola's book, 2015, chap 8) modifies the basic auto-encoder reconstruction scheme by not requiring the reconstruction to be performed from the deepest internal representation. Instead, the decoder also receives input from corresponding layers of the encoder by adding lateral connections, thus fine details do not need to be propagated to deep representational layers. Based on the same idea, we will explore other schemes to learn hierarchical representations suitable for abstract high-level tasks such as image classification, while also maintaining the capability to reconstruct fine details in the input to ensure compatibility with reconstruction-based loss functions. We will explore 3 different directions: models with scattering operators, mixed representations composed of both continuous and discrete values, deep generative models.
- algorithm development
- reading and writing articles
- experimental validations
- machine learning
- deep learning
- computer vision
The IMAGE team is one of the 7 research teams of the GREYC laboratory, located in Caen / France. Our research activities are focused on the development of new methods to process and analyze images and signals. We also apply the resulting algorithms on real cases thanks to our collaborations with biomedical research centess in Normandy (Cyceron, CHU de Caen, Centre F. Baclesse, Centre Hospitalier Public du Cotentin). The team has a strong expertise on these two themes:
– Variational methods, PDEs and statistics for image processing. The scientific base of this theme gathers works on inverse problems, optimisation, signals defined on graphs, 3D point clouds, high-dimensional statistics (estimation, decision, optimal transport), computational photography and multi-valued mathematical morphology.
– Pattern recognition and feature extraction from images and videos. This theme is organized around pattern recognition with methods based on graphs or neuronal networks (deep learning), metric learning, image descriptors and knowledge engineering for the conception of image processing applications.
The team members have different backgrounds (computer science, signal/image processing,applied mathematics, artificial intelligence). This variety of skills is one of our strengths as we can approach image processing and analysis from different scientific viewpoints and paradigms.
We talk about it on Twitter!