The APi Project

Summary

The APi project (ANR grant ANR-18-CE23-0014) aims to unleash the potential of machine learning methods by addressing the fundamental issue of the pre-image from all angles in kernel methods and deep neural networks. To this end, the main objectives of the project can be summarized around the following research directions (WPs for Working Packages):

  • WP1. Establish a new class of algorithms that overcome the curse of the pre-image, through closed-form solutions obtained using probabilistic models or joint optimization strategies.
  • WP2. Provide metric learning and representation methods for time series, revisiting neural networks or kernel methods in the light of the pre-image problem.
  • WP3. Explore pattern recognition on discrete structured spaces, especially for graph data, with emphasis on the task of synthesizing graphs/molecules from a set of training graphs.
  • WP4. Couple the pre-imaging problem in kernel methods with two classes of neural networks, auto-encoders and generative adversarial networks (GANs), in order to provide more in-depth analyses of their underlying functioning, and to stimulate the development and exchange of ideas for designing new architectures.
  • WP5. To these 4 well-identified scientific and technical objectives is added a transversal research direction that aims at exploring new application domains.

Principal investigators

The project Team…

Research Results

The project team carried out exploratory research as part of the project APi and obtained significant results in all the aforementioned WPs.

The main obtained results are theoretical, methodological and algorithmic results presented here : Research Results

There results have been disseminated in major Machine Learning and Pattern Recognition journals (e.g., Artificial Intelligence, JMLR) and top venue conferences (e.g., several ICML, ICLR, AISTATS, ICASSP ). Complete list at the end of Research Results page.

Outcomes and Impacts

Besides the research results obtained by the team within the project, the project APi led to several outcomes and impacts.

Training new researchers

Several PhD students and Post-doc fellows have been funded by this project. The main non-permanent researchers were Luc Brogat-MOTTE (PhD student for 36 months), Clément GLÉDEL (PhD student for 36 months), Saeed VARASTEH YAZDI (post-doc for 23 months) and
Muhammet BALCILAR (post-doc for more than 12 months thanks to a co-funding from the Programme PAUSE (Collège de France).

Start-ups

As a wider long-term effect on the society, the economy and science, two start-ups have emerged that are directly linked to the APi project:

  • Startup Xpdeep, founded in 2023, aims to make deep learning explainable. Xpdeep’s scientific director is Ahlame Douzal, who was also a scientific leader of the APi project. It is worth noting that, by addressing the resolution of the pre-image, the project APi has paved the way for explainable and interpretable machines, as explained in the work of PhD student Thai Tran Thi Phuong, under the supervision of Ahlame Douzal and Paul Honeine (see also their paper entitled Interpretable time series kernel analytics by pre-image estimation, Artificiel Intelligence, 2020).
  • Startup Tellux, founded in late 2019, aims to develop deep learning methods for hyperspectral imaging for soil pollution analysis. One of the three co-founders of Tellux, Paul Honeine is also the principal coordinator of the APi project. It is worth noting that the APi project has provided some advances in hyperspectral imaging using the pre-image problem, as demonstrated for instance in the work of Zhu et al. entitled Pixel-wise linear/nonlinear nonnegative matrix factorization for unmixing of hyperspectral data (IEEE ICASSP, 2020).

Taming the Beast of the Preimage in Machine Learning for Structured Data: Signal, Image and Graph