The last couple of years have seen a renewed interest in deep learning and deep neural networks (DNNs). Advances in hardware and in particular the use of high-end graphic processing units (GPUs) have enabled the realization of large and powerful networks [1], that are the first pattern recognizers capable of human-competitive performances on handwritten characters recognition, traffic signs recognition or face recognition [2]. More generally, DNNs today outperform all other machine learning techniques in the fields of image classification, scene labeling and speech recognition. They are already used by companies like Google or Facebook to classify the huge amount of data published on their social networks. Because DNNs are very similar to the organization of the visual cortex in the brain, there is today a convergence between computational neuroscience and DNNs and deep learning, which are seen as a step towards realizing strong AI [3]. DNNs are therefore a Key Enabling Technology for realizing intelligent, autonomous systems, like domestic and service robots and more generally, Internet of smart Things (IoT) that need to be able to sense and identify their visual and/or auditory environment.

Nowadays, the DNNs used in information processing are implemented in software, using digital coding of the data (every pixel of an image is coded with a decimal number for example). The data processing is done image by image and is therefore entirely synchronous and sequential. This is different from the biological visual system, where the visual data stream is asynchronous and continuous and is made of spikes transmitted by the optical nerve. Moreover, the learning in DNNs is generally done using back-propagation, which is a classical mathematical optimization algorithm, very different from the biological learning algorithms in the brain.

The objective of this thesis is to increase the convergence between the formal models of DNNs and the visual cortex models coming from the neurosciences. We propose to study the use of neuro-inspired event-based (or spike-based) coding [4][5] for DNNs, along with the use of neuro-inspired unsupervised learning methods, such as Spike-Timing-Dependent Plasticity (STDP) [6][7]. This thesis proposes an original approach and is firmly interdisciplinary. The Ph.D. student will be based in an electronics and signal processing lab, with an advisor from the field, and will be supervised by a neuroscientist specialized in the visual system in the brain.

This position is open until it is filled.

Department: Département Architectures Conception et Logiciels Embarqués (LIST-LETI)
Laboratory: Laboratoire De Fiabilisation des Systèmes Embarqués
Start Date: 01-09-2015
ECA Code: SL-DRT-15-0102
Contact: olivier.bichler<στο>cea.fr