Offre en lien avec l’Action/le Réseau : – — –/Doctorants

Laboratoire/Entreprise : Télecom SudParis and Université d’Artois
Durée : 3 ans
Contact : aikaterini.tzompanaki@cyu.fr
Date limite de publication : 2023-12-01

Contexte :
The ANR project EXPIDA (EXplainable and parsimonious Preference models
to get the most out of Inconsistent DAtabases) (Project website: https://www.cril.univ-artois.fr/expida/index.html) aims to develop principled and
rigorous explainable techniques for dealing with imperfect data. More precisely,
EXPIDA aims to design tractable methods for dealing with conflicts in databases
by efficiently exploring novel inconsistency-tolerant semantics and quantifying
contradictions [6] to answer queries and to draw (high level) explanatory information.
While the set of repairs (maximal consistent datasets) is often large for
real databases, we aim to explore preference mechanisms (e.g., [4]) in order to
retrieve meaningful answers and explanations to identify the reasons of query answers,
and to assist end-users to “realize” query outputs. The EXPIDA project
aims in addition to be useful for applications intensively relying on multiple
heterogeneous data sources. Many such applications are nowadays developed
in various domains such as transportation control, health management, social
network analysis, data journalism, etc. This research project will advance the
state-of-the-art in two major ways: innovations in inconsistency management,
preferences and explanation for databases, and developing practical Artificial
Intelligence tools for managing inconsistent databases with validations on real
data. In this context we are proposing two PhD positions.

Sujet :
1st PhD position: This PhD thesis will focus on the following two main aspects. The first one
concerns query answering semantics with inconsistent databases. Despite the
fact that various methods have been studied, for drawing useful information in
presence of conflicts in the propositional and description logic settings [2, 5], we
are not aware of any existing work studying rich conflict-tolerant relations in
the context of databases equipped with flexible preference relations [7], such as
the partially ordered relation (e.g., [1, 3]). Then, investigating database query
answering under such varying conflict-tolerant methods should be accompanied
by studying the computational complexity of the different related problems,
which is among the main objectives of this thesis.
The second objective of the thesis is how to handle preferences in conflicting
databases. In fact, to answer queries over conflicting databases, it is crucial
to express priority among potential repairs in order to select the most optimal
candidates. As the number of potential repairs can be (very) large, one may
choose to rank repairs according to some preference criteria, and select a small
number of the most desirable repairs. Moreover, we note that the preference
among the sources of data could be of different nature. Often, conflict-tolerant
methods aim to find a stratification inducing a total preorder among all pieces
of information. Such a stratification allows to handle more easily inconsistencies
in data sources. Nonetheless, this can lead to a comparison of incomparable and
independent pieces of information. The main objective here is then to develop
a new framework to handle preferences in conflicting databases in order to draw
meaningful answers to user queries.
Applications should be submitted via email to Said Jabbour (jabbour@cril.fr),
Badran Raddaoui (badran.raddaoui@telecom-sudparis.eu) and Yue Ma (ma@lri.fr)
with the subject “Application for EXPIDA PhD 1”.

2nd PhD position: This PhD thesis will focus on the aspect of explainability in two ways, as
presented in what follows.
First, we will consider explanations for query results over inconsistent databases
with different conflict-tolerant semantics (e.g., consistent, brave, intersection repair,
intersection closed repair, non-objection, nonconsensus based semantics,
etc.). To this end, we will adapt the notion of lineage (or provenance) [2] in the
context of uncertain/inconsistent data and devise mathematical formalisations
that will provide the necessary properties for characterising and measuring the
‘quality’ of the explanations. Through the study of causality [7] and argumentation
[4, 9, 8, 3] in our setting, we further aim at improving the acceptability and
usefulness of the provided explanations by the end-user. Second, we will investigate
the complementary problem of explaining missing query results, widely
known as Why-Not explanations, which has not been yet addressed in the context
of inconsistent databases. In the setting of consistent databases, Why-Not
explanations ‘explain’ why certain results are not generated by a query (or a
workflow) by means of instance-based (i.e., source tuples), query-based (i.e.,
query operators) or refinement-based explanations (i.e., corrected query). Close
to our problem, [1] has proposed Why-Not provenance polynomials, which may
account for probabilistic tuples. It would be interesting to check how such formalisations
can be revisited to fit the inconsistent database’s different conflict tolerant semantics. Applications should be submitted via email to Badran RADDAOUI badran.raddaoui@telecomsudparis.
eu and Aikaterini TZOMPANAKI aikaterini.tzompanaki@cyu.fr, with
the subject “Application for EXPIDA PhD 2”.

Profil du candidat :
The PhD candidate should have a background on computer science and on at least on of the following domains: databases, logics, artificial intelligence, algorithms and complexity. They should be excellent students, self-motivated and eager learners.

Formation et compétences requises :
Engineering degree or Bachelor and Master’s degree in Computer science or (applied) mathematics.

Adresse d’emploi :
1st PhD: Université d’Artois, Lens-France
2nd PhD: Télécom SudParis, Paleseau-France

Document attaché : 202310032054_EXPIDA_PhD_subjects.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LIRIS
Durée : 3 ans
Contact : stefan.duffner@insa-lyon.fr
Date limite de publication : 2023-11-30

Contexte :
Research on machine learning and Deep Neural Networks (DNN) has made considerable progress in the past decades. State-of-the-art DNN models usually require large amounts of data to be trained and contain a tremendous number of parameters leading to overall high resource requirements, in terms of computation and memory and thus energy. In the past years, this gave rise to approaches to reduce these requirements, where, for example, during or after training, parts of the model are removed (pruning) or stored with lower precision (quantisation) or surrogate models are trained (knowledge distillation) or where the best configuration is searched by testing different parameters (Neural Architecture Search, NAS). Also, concerning the hardware, many optimisations have been proposed to accelerate the inference of DNNs on different architectures.
But these accelerators are usually specific to a given hardware and are optimised to satisfy certain static performance criteria. However, for many applications, the performance requirements of a DNN model deployed on a given hardware platform are not static but evolving dynamically as its operating conditions and environment change. In the context of this ANR-funded project we propose an original interdisciplinary approach that allows DNN models to be dynamically configurable at run-time on a given reconfigurable hardware accelerator architecture, depending on the external environment, following an approach based on feedback loops and control theory.

Sujet :
At software (SW) level, for a given DNN model, different variants with incremental precision levels can be obtained by setting parameters along different dimensions: (i) data precision or quantization (increasing/decreasing bit-width of activations and/or weights), (ii) degree of sparsification (e.g., pruning, tensor decomposition), (iii) depth of the NN (number and type of network layers to execute). Depending on the chosen SW precision level, the mean output accuracy will change, as well as energy consumption and timing. The key observation is that, for some particularly “easy” inputs, using high-precision energy-hungry computations is an “overkill”. Conversely, for “hard” inputs, low-precision energy-efficient computations are not enough. Therefore, being able to dynamically change the SW precision is key to enable energy-efficient and accurate NN computations.
At the same time, at the hardware (HW) level, the DNN accelerator needs to be configurable at runtime to satisfy SW processing requirements. Different degrees of existing HW strategies have an impact on energy consumption and runtime, which need to be taken into account when designing NN architectures and SW compression schemes.
This PhD thesis will concentrate on the SW (i.e. machine learning) side. There are two major challenges that need to be addressed. First, an effective method is required to train different DNN model variants for the same task but responding to different performance criteria, i.e., providing different levels of classification accuracy, throughput, energy consumption etc. Also, to reduce memory requirements, the more complex variants should be able to mutualize as much parameters as possible with the simpler models. Based on existing approaches and results on structured pruning and multi-exit models, one objective is to develop new algorithms and DNN architectures that can be parameterized at inference time to execute only the parts of the model that are really needed. For some layers, different parameter or activation precisions (8, 16, 32-bit) may be necessary and should be able to be switched at run time depending on the sensitivity of the different parts of the model, that should be quantified. The second challenge concerns the influence of these architectural changes on the performance. The impact of different possible SW configurations must be assessed in terms of final accuracy of the obtained results and in terms of the corresponding effort (e.g., in terms of number of operations, their datatype, potential computational over-head etc.).

Profil du candidat :
– A strong background in machine learning and in particular neural networks
– Capacity to work autonomously and within a research team
– Scientific curiosity and creativity
– Very good English proficiency

Formation et compétences requises :
A Master degree in computer science or applied mathematics or similar.

Adresse d’emploi :
INSA Lyon, laboratoire LIRIS, Lyon

Document attaché : 202310032028_thesis_proposal_insa_lyon.pdf