Personal web page : http://jstarck.cosmostat.org
Laboratory link : http://www.cosmostat.org
Video digitizes images and audio streams. It is a cost-effective media for remote vision and for
capturing physical events occurring around us. In recent applications, such as video surveillance,
data streams are processed for finding patterns in order to aid human vision with machine
learning. In other domains, such as Astrophysics, observational tools record phenomena appearing
beyond the visual and audio spectrum, producing data that we could compare to video streams. In
both cases, some people need to spend time observing data streams so as to find special events,
which might be physical risks in sensible places or neutrinos before a red giant star collapses in a
supernova somewhere in the Universe . As Big Data and uninterrupted data flows easily escape
from human attention and understanding, computer vision can massively process these data flows
For identified events, such as fire and flames, computer vision is reliable and efficient as it can
distinguish between flames and fire-colored moving objects, thanks to the use of spatial and
temporal wavelet transforms . Computer vision systems also enable the early detection of
identified events and features, for autonomous vehicles and robots . With unidentified events,
the best way to efficiently detect and identify them remains human observation and analysis. As a
result, detecting and recognizing unidentified events in data streams requires man-machine
collaboration, instead of man-machine competition.
D R F / I R F U
Direction de la Recherche Fondamentale– Institut de Recherche sur les lois Fondamentales de l’Univers
Active learning strives towards the collaboration of human intelligence and artificial intelligence.
With supervised models in machine learning and deep learning, items must be previously labelled
by experts and/or people answering to labelling queries – e.g. captcha. Active learning from
relative queries, instead of absolute ones, makes it possible to train machine-learning algorithms
from weak predictors, such as side information or non-experts answers, then to get comparable or
better performance than with the experts’ answers to labelling queries . With such a model,
item recognition relies on human information processing-based inputs, more than on human
The IoT and smart cities are going to produce huge and heterogenous datasets, to be compared
with astronomical observations and collected in data lakes. Therefore, the detection of unexpected
and/or unidentified events in data streams should become a significant task, depending on active
learning for the future of society, science, and economy. Finally, detecting weak signals in videos
streams thanks to active learning or other approaches could represent an important stake, while
there currently appears no or few works published on this topic, to our knowledge. It might allow
us to catch unusual phenomenon in video surveillance before specific events occur, so as to
prevent risks. It might also allow cosmologists to be notified of unobserved phenomena that could
require attention, in order to identify and classify them for an active reinforcement learning
The thesis we propose concerns the research and experimentation of both sparse representations
 and online models of active learning for the detection of weak signals in data streams. It is
focused on application in video surveillance and astronomical data surveillance. It will be
experimented thanks to Big Data systems and Lambda architecture, allowing to run offline and
online analysis tasks in parallel processes. The candidate should be interested in computer vision,
online analysis and research applications. Candidate preferring a permanent position in private
research after the thesis are welcome (Lead Researcher for online analysis @DATA2B).
 Qian B., Wang X., Wang F., Li H., Ye J., and Davidson I. Active learning from relative
queries. In Proceedings of the Twenty-Third International Joint Conference on Artificial
Intelligence, IJCAI ’13, pages 1614–1620. AAAI Press, 2013.
 Donadio F., Frejaville J., Larnier S., and Vetault S. Human-robot collaboration to perform
aircraft inspection in working environment. In Proceedings of 5th International conference on
Machine Control and Guidance,, 2016.
 Antonioli P., Tresch Fienberg R., Fleurot F., Fukuda Y., Fulgione W., Habig A., Heise J.,
McDonald A.B., Mills C., Namba T., Robinson L.J., Scholberg K., Schwendener M., Sinnott
R.W., Stacey B., Suzuki Y., Tafirout B., Vigorito C., Viren B., Virtue C., and Zichichi A. Snews:
the supernova early warning system. New Journal of Physics, 6(1):114, 2004.
 Dedeoglu Y. Güdükbay U. & Cetin A. E Töreyin, B. U. Computer vision based method
for real-time fire and flame detection. 1(27):49–58, 2006.
 Starck J.-L., Murtagh F, and Fadili J, Sparse Image and Signal Processing: Wavelets and
Related Geometric Multiscale Analysis, Cambridge University Press, Cambridge (GB), 2015