I am mainly interested in innovation through machine learning techniques, with a particular focus on neural networks and deep learning approaches.
Neural Networks and Deep Learning. I built several kinds of neural networks (multi-layer perceptrons, convolutional and recurrent nets) for handwriting recognition. In my thesis, I study the impact of several modeling choices (depth, capacity, ...) on the results of the network alone, and in the complete text recognition pipeline. Cursive text recognition is interesting because of the sequential - and somewhat constrained - nature of the prediction on the one hand, and the two-dimensional - and ambiguous - nature of the input on the other hand, making the problem more challenging than it seems, compared for example to natural image classification. I took part in (and extended) the first experiments on applying the dropout technique LSTM networks. I also tried to get a better insight into the connectionist temporal classification training method and its intriguing "blank" label. On a more practical side, I participate in the development of the neural network library at A2iA, with a focus on the implementation and application of cutting-edge techniques to handwriting recognition.
Handwriting Recognition, Document Processing, Computer Vision. I applied deep learning methods in computer vision applications, namely the recognition of handwritten text for automatic document processing. I like this task because it is related to several topics in AI and machine learning, such as image understanding and language processing. I worked on complete pipelines, from image pre-processing and feature extraction, to neural network optical models, hidden Markov models and language models. I also evaluated system combination methods. In my PhD thesis, the results reported on three public benchmarks outperformed the best published at that time. With A2iA's research lab, we've won several international evaluations of handwriting recognition, such as the OpenHaRT challenge in 2013 for the recognition of Arabic texts, and the MAURDOR evaluation, for the recognition of handwritten and printed texts in English, French and Arabic. Moreover, I am interested in developing methods for the automatic mapping of transcript to images, at the level of text lines, from which we can train recognition systems. Those methods are particularly useful to leverage the huge amount of transcribed manuscripts available. They have been crucial in the Oriflamms project, which aims at using text recognition systems to extract characters for paleographical analysis, and for the creation of systems for the Bentham contests in 2014 and 2015. Finally, I enjoyed working for the Cognilego project, aiming at building new text recognition systems based on the recent findings in cognitive science about human's brain strategies.
Industrial Applications. I chose to do my PhD in a company because I am willing to tackle real-life challenges, and bring ideas and understanding that will have a direct impact on the new technology being developed and brought to the market. However, I am keen on publishing my research results and taking part in international conferences. I believe that this kind of sharing information and knowledge with the community is crucial for the rapid development of technologies.