Section: New Results

Toward Good AI

Causal Modeling

Participants: Philippe Caillou, Isabelle Guyon, Michèle Sebag;

Post-docs and PhDs: Olivier Goudet, Diviyan Kalainathan

Collaboration: David Lopez-Paz (Facebook).

The search for causal models relies on quite a few hardly testable assumptions, e.g. causal sufficiency [152]; it is a data hungry task as it has the identification of independent and conditionally independent pairs of variables at its core. A new approach investigated through the Cause-Effects Pairs (CEP) Challenge [107] formulates causality search as a supervised learning problem, considering the joint distributions of pairs of variables (e.g. (Age, Salary)) labelled with the proper causation relationship between both variables (e.g. Age "causes" Salary) and learning algorithms apt to learn from distributions have been proposed [109]. An edited book is in preparation  [64].

In D. Kalainathan's PhD and O. Goudet's postdoc, the search for causal models has been tackled in the framework of generative networks [44], trained to minimize the Maximum Mean Discrepancy loss; the resulting Causal Generative Neural Network improves on the state of the art on the CEP Challenge. However, due to the shortage of real-world variable pairs for which the causation type is known, the CEP challenge has been enriched using artificial pairs (e.g. considering variations on pairs of entities involved in biological regulatory networks), biasing the causation training process. On-going studies investigate how the use of such artificial pairs (the so-called Mother Distribution) to train a causation model aimed at real pairs can be cast as a domain adaptation problem [97], [78].

An attempt to circumvent the need for a large dataset of variable pairs, sampled for the Mother Distribution, we proposed the Structural Agnostic Model approach [57]. Working directly on the observational data, this global approach implements a variant of the popular adversarial game [97] between a discriminator, attempting to distinguish actual samples from fake ones, obtained by generating each variable, given real values from all others. A sparsity $L_1$ penalty forces all generators to consider only a small subset of their inpu variables, yielding a sparse causal graph. SAM obtains state-of-the-art performances on synthetic data.

An innovative usage of causal models is for educational training in sensitive domains, such as medicine, along the following line. Given a causal generative model, artificial data can be generated using a marginal distribution of causes; such data will enable students to test their diagnosis inference (with no misleading spurious correlations in principle), while forbidding to reverse-engineer the artificial data and guess the original data. Some motivating applications for causal modeling are described in section 4.1.

Explainability

Participants: Isabelle Guyon, François Landes, Marc Schoenauer, Michèle Sebag.

Causal modeling is one particular method to tackle explainability, and TAU has been involved in other initiatives toward explainable AI systems. Following the LAP (Looking At People) challenges, Isabelle Guyon and co-organizers have edited a book [29] that presents a snapshot of explainable and interpretable models in the context of computer vision and machine learning. Along the same line, they propose an introduction and a complete survey of the state-of-the-art of the explainability and interpretability mechanisms in the context first impressions analysis [56].

The team is also involved in the proposal for the IPL HyAIAI (Hybrid Approaches for Interpretable AI), coordinated by the LACODAM team (Rennes) dedicated to the design of hybrid approaches that combine state of the art numeric models (e.g., deep neural networks) with explainable symbolic models, in order to be able to integrate high level (domain) constraints in ML models, to give model designers information on ill-performing parts of the model, and to provide understandable explanations on its results.

Finally, a completely original approach to DNN explainability might arise from the study of structural glasses (7.2.3), with a parallel to CNNs with rotational invariances, that could become an excellent non-trivial example for developing explainability protocols.

Experimental Validation of the Autonomous Vehicle

Participants: Guillaume Charpiat, Marc Schoenauer; PhD and Engineers: Marc Nabhan, Nizham Makhoud, Raphaël Jaiswal

Collaboration: Hiba Hage, Philippe Reynaud, and Yves Tourbier (Renault)

As said (Section 3.1.2, Tau is considering two directions of research related to the certification of MLs. The first direction, toward experimental validation, focuses on the coverage of the datasets (more particularly here, used to train an autonomous vehicle controller), and is the subject of this section, while the second one, related to formal approaches, has just started with the beginning of Julien Girard's PhD and has not yet lead to results.

Statistical guarantees (e.g., less than ${10}^{-8}$ failure per hour of operation) are obtained by empirical tests, involving millions of kilometers of driving in all possible road, weather and traffic conditions as well as intensive simulations, the only way to full control of the driving conditions. The validation process thus involves 3 steps: i) making sure that all parts of the space of possible scenarios are covered by experiments/tests with sufficiently fine grain; ii) identify failures zones in the space of scenarios; iii) fix the controller flaws that resulted in these failures.

TAU is collaborating with Renault on steps i) (topic of a one-year POC) and ii) (Marc Nabhan's CIFRE PhD). In both cases, the current target scenario is the insertion of a car on a motorway, the "drosophila" of autonomous car scenarios.

Note that another approach toward experimental robustness is investigated in Nizam Makdoud's PhD (CIFRE Thalès), started in March 2018, where Reinforcement Learning is used to find ways to fool some security system.

Clustering of scenarios A first one-year Proof of Concept (ending Oct. 2018) has demonstrated the feasibility and the usefulness of scenario clustering, assuming the availability of data describing the scenarios, i.e., the trajectories of all vehicles involved. Publicly available datasets (e.g., NGSIM were used in a first step. The difficulties met are the following. Firstly, trajectories are varying-length time series, requiring the use of recurrent NNs or LSTMs. Secondly, a scenario is invariant under permutations of the different vehicles involved; neural architectures are taking inspiration from social LSTMs [67]. Lastly, most recorded real-world scenarios are uninteresting (all vehicles drive on in their lanes).

The results of this POC have been duly delivered to Renault, but will remain internal at this point. The follow-up collaboration will explore metrics (in the latent space, or learned via Siamese networks), to complete the clustering in a semi-supervised setting (exploiting human feedback to select "typical" scenarios).

Detection of controller flaws Marc Nabhan's PhD (CIFRE Renault) is concerned with the identification of the conditions of failures of the autonomous car controller. Only simulations are considered here, with one scenario being defined as a parameter setting of the in-house simulator SCANeR. The goal is the detection of as many failures as possible, running as few simulations as possible.

A key difficulty, beside that of getting actual data, is the very low probability of failure.On-going work builds upon TAU expertise in active learning using Monte-Carlo Tree Search [140] and evolutionary optimization, in particular taking inspiration from Novelty Search [121] to focus the exploration on unexplored regions of the scenario space, as well as portfolio optimization and instance-based algorithm selection (see Section 3.3.1).