Les commentaires sont clos.

VALVACC: An environmental sentinel and AI procedure to monitor marine ecosystems.

23 Novembre 2021

Catégorie : Post-doctorant

As part of the MIAI Grenoble Alpes Institute (Multidisciplinary Institute in Artificial Intelligence), which aims to conduct research at the highest level in artificial intelligence, the post-doctoral student will be asked to answer a fundamental question: "Can we diagnose and predict the effect of a disturbance, climatic or otherwise, on a coastal ecosystem?
We propose to develop a methodology to characterise the signals and biological states of a species (scallop) known to be sensitive to variations in its environment and used as a model for studying the responses of benthic fauna to anthropogenic impacts. These bivalves, monitored in situ by miniature and multimodal sensors (accelerometers and valvometers), are used as indicators of the quality of the environment to ultimately constitute monitoring systems for global changes or warning systems for transitory disturbances.

We have an excellent and large corpus of in situ data from a completely new acquisition system (frequency, autonomy, weight, size, precision, etc.). A set of tools has already been developed (patents), the objective will be to propose a multimodal extension and to predict the stress sequences by CNN1 1D networks. (end of project at T0 =12 months).


Missions: In the framework of the MIAI Grenoble Alpes Institute (Multidisciplinary Institute in Artificial Intelligence) the role of the postdoctoral fellow will be to understand existing algorithms and make an automatic classification of transient signals applied on valvometric or accelerometric signals acquired to classify animals and/or the behaviours of an animal according to its biological state (stressed or not). It will be asked to propose an extension taking into account the multimodal character of the data. Once the temporal sequences linked to stress are identified on the sensors, it will be necessary to predict these stress sequences with 1D CNNs using the concepts developed in the work of Kiranyaz. The objective is to define the best architecture capable of performing the task of predicting the signal or behaviour related to the stress or dynamics under study.

The main activity will be to propose extensions to existing algorithms or to adapt existing algorithms for the detection and classification of different types of behaviour.
Five months will be devoted to task 1 (Classification of accelerometry and valvometry data and use of multimodality).
Five months will be devoted to task 2 on the use of 1D-CNN to predict stress behaviour.

Objective result(s) setting the end of the agent's mission:
It is expected that the results will be disseminated through publication in journals, publications in conferences as well as through publication of the code. It is extremely important that the candidate is able to interact with biologist and ecologist collaborators on both the biology itself and on the signal and AI aspects. (Short) stays in Brest will be necessary.

The expected results will concern the analysis of acquired data and data in the process of being acquired.
The evaluation of the results will be carried out in consultation with the project partners.

Priority expected skills :
Knowledge of machine learning methods, in particular deep learning.
Knowledge of machine learning methods, in particular deep learning.
Possibly basic knowledge of environmental sciences.
Use of current data science tools: Python language and deep learning framework (such as PyTorch, TensorFlow or Keras)