New Ph.D. offer: “Towards Large-Scale Evaluation Protocols for the Visually Impaired”

The Biovision Lab is seeking to recruit a highly-qualified Ph.D. candidate to join our dynamic, multi-disciplinary research team, which primary goal is to study vision impairment from theoretical and applied perspectives. You will be based in Sophia-Antipolis on the French Riviera, in the Inria Research center. In this Ph.D., you will be part of a pluridisciplinary academic consortium, with three supervisors bringing their complementary expertise (Pierre Kornprobst, Jean-Charles Régin and Aurélie Calabrèse from Inria/Université Côte d’Azur), and one international expert in low-vision (Gordon E. Legge; University of Minnesota). You will also benefit from the participation of a private partner interested in natural language production (MantuLab Amaris Research Unit in Sophia Antipolis). Overall, this project will foster promising perspectives for your career, notably in the fields of natural language processing and AI, which are highly demanded in academia and industry.


TITLE: Pushing the limits of reading performance screening with Artificial Intelligence: Towards large-scale evaluation protocols for the Visually Impaired

GENERAL GOAL: This Ph.D. project aims at developing novel solutions for diagnosing visual pathologies, using Artificial Intelligence through automated text generation, to carry out large-scale reading tests. The MNREAD acuity chart – a standardized reading test prominently used worldwide in both clinical and research settings – will serve as the foundation for this work.

CONTEXT: In 2015, 405 million people were visually impaired around the globe, against ‘only ‘285 million in 2010 [1]. Earlier interventions could have prevented almost half of it in the form of treatment or rehabilitation. As a consequence, it is crucial to have more efficient solutions to detect visual pathologies at the early stages.

PROBLEM: Since reading speed is a strong predictor of visual ability and vision-related quality of life for patients with vision loss, reading performance has become one of the most used clinical measures for judging the effectiveness of treatments, surgical procedures, or rehabilitation techniques. Accurate measurement of reading performance requires highly standardized reading tests, such as the MNREAD acuity chart [2]. This test, available in 19 languages, allows measuring reading performance in people with normal and low vision. In brief, performance is measured from the time needed to read a series of short sentences that were designed to be equivalent in terms of linguistics, length, and layout. Each sentence must be presented only once to avoid the introduction of a memorization bias and ensure accurate measurement. However, because of their highly constrained nature, MNREAD sentences are hard to produce, leading to a limited number of test versions (only two in French). Given that repeated measures are needed in many applications of MNREAD, there is a high interest from the scientific and medical communities for a much larger pool of sentences.

STATE-OF-THE-ART: Fully automated generation of constrained text is a very complex task. Recently, a method for computer-generated sentences has been explored by the MNREAD creators themselves [3]. However, this semi-automated method presents several major drawbacks: (1) it relies on sentence templates that must be created manually (i.e., sequences of placeholders, each containing a list of possible words that fit into the sentence at that point); (2) it works in two stages (i.e., sentence creation followed by sentence selection) implying additional calculations and longer execution time; (3) it can not be extended to other languages.

DESCRIPTION OF THE WORK: The problem that we must solve here is to generate a vast number of sentences, while taking into account very strict linguistics, length and layout constraints, such as fixed number of characters, restricted vocabulary (3rd grade level), tightly constrained physical layout, etc. To tackle this matter, one approach we will consider is oriented towards the automatic production of constrained text from a corpus. However, to further extend the test capabilities, we must keep the option to modify these constraints along the course of our project. Therefore, it is crucial to consider methods based on constraints satisfaction, such as those developed by J.C. Régin (co-supervisor), in collaboration with the Sony Computer Science Lab [4]. These methods are primarily based on multivalued decision diagrams and the operations that allow them to be manipulated [5,6,7] and have already proven to be efficient [8]. More general methods will also be considered, such as those based on word embedding paired with neural networks or decomposition into singular values. Tools such as OpenAI’s GPT2 or SimpleNLG can also be used. Once created, sentences will be validated experimentally, allowing for fine-tuning of the generator.

SUPERVISION: This Ph.D. will be co-supervised by P. Kornprobst (Inria, Biovision Lab), J.C. Régin (I3S, Constraints and Application Lab), and A. Calabrese (Inria, Biovision Lab).

POTENTIAL IMPACT ON THE FELLOW CAREER: This Ph.D. has the potential to open promising career perspectives for the candidate who will (1) acquire substantial expertise in natural language processing methods, (2) work on a multidisciplinary project with contributions in computer science, cognitive science, and low-vision research, (3) spend one to three months in the world-renowned lab of G.E. Legge (Univ. of Minnesota, USA) working on low-vision research. For all the reasons mentioned above, this Ph.D. will be of high value in academia. Besides, all the acquired skills paired with a high potential for transfer of the work done will be of great interest outside academia, as evidenced by MantuLab’s interest in our project.

BIBLIOGRAPHY:
[1] Bourne, R., et al. (2017) Magnitude, temporal trends, and projections of the global prevalence of blindness and distance and near vision impairment: a systematic review and meta-analysis, The Lancet Global Health, Volume 5, Issue 9, e888 – e897
[2] Mansfield J.S., et al. (1993) A new reading-acuity chart for normal and low vision. Ophthalmic and Visual Optics/Noninvasive Assessment of the Visual System Technical Digest, (Optical Society of America, Washington, DC., 1993.) 3: 232–235.
[3] Mansfield, et al. (2019) Extending the MNREAD sentence corpus: Computer-generated sentences for measuring visual performance in reading. Vision Research, 158, 11–18.
[4] Papadopoulos, A. et al. (2015) Generating all Possible Palindromes from Ngram Corpora. IJCAI 2015: 2489-2495
[5] Perez, G., Régin, J.-C. (2015) Efficient Operations On MDDs for Building Constraint Programming Models. IJCAI 2015: 374-380
[6] Perez, G., Régin, J.-C. (2017) Soft and Cost MDD Propagators. AAAI 2017: 3922-3928
[7] Perez, G., Régin, J.-C. (2018) Parallel Algorithms for Operations on Multi-Valued Decision Diagrams. AAAI 2018: 6625-6632
[8] Perez, G., Régin, J.-C. (2017) MDDs: Sampling and Probability Constraints. CP 2017: 226-24

HOW TO APPLY? 

  • Through BoostUrCAreer@UCA: This Ph.D. offer is part of the BoostUrCAreer project which aims at implementing at Université Côte d’Azur and with the support from the European Commission and the Conseil Region Sud-Provence-Alpes-Côte d’Azur a multidisciplinary doctoral programme in e-health. It will offer excellent conditions for the candidates. Check out this website for more information and to apply. Deadline: March 23, 2020
  • Through Computer Science Doctoral School, via this link. Deadline: May 11, 2020
  • Through Inria Ph.D. Campain: Information coming soon…
  • You may also send us a message with your CV and transcript of record. We will get back to you shortly.

Comments are closed.