Section: New Results

Association to other scientific domains

Concerning Machine Learning and our work on reservoir computing, X. Hinaut is collaborating with Michael Spranger (Sony Lab, Tokyo, Japan) on grounding of language, adapting Hinaut's previous Reservoir parser (ResPars) with the representational system of Spranger: IRL (Incremental Recruitment Language) [17]. He is also collaborating with Hamburg on the use of reservoir models for robotic tasks (cf. § 9.3). In this work, we have shown that the RLM can successfully learn to parse sentences related to home scenarios in fifteen languages [5]. This demonstrates that (1) the learning principle of our model is not limited to a particular language (or particular sentence structures), and (2) it can deal with various kinds of representations (not only predicates), which enable users to adapt it to their own needs. Some people can mix two languages within the same sentence: this is known as intra-sentential code-switching. With M1 intern Pauline Detraz, we collected data from human subjects that were required to mix pairs of given sentences in French and English. The corpus obtained have some very complex mixed sentences: there can be until eleven language switches within the same sentence. Then, we trained our Reservoir-based sentence Parsing model, with the collected corpus. Surprisingly the model is able to learn and generalize on the mixed corpus with performances nearly as good as the unmixed French-English corpus [16]. A post-doc joined the team in Nov 2019 to work on the project HuRRiCane ("Hierarchical Reservoir Computing for Language Comprehension") project founded by Inria. This project aims at extending the ResPars model to work from speech inputs to sentence comprehension including coherancy checking. In other words, the objective is to experiment how a sentence comprehension model, based on reservoir computing, can learn to understand sentences by exploring which meanings can have the sentences, implying several steps from stream of phonemes to words and from stream of words to sentence comprehension. The model will be implemented on a virtual agent first and then on the Nao humanoid robot.

This project is linked to other projects in the team on the hierarchical organization of the prefrontal cortex (including Broca's area, involved in language). This hierarchy corresponds to an increasingly higher abstraction, which is made by different sub-areas. We will therefore be able to link this post-doc project to existing projects of the team, where different levels of abstractions are necessary for sentence comprehension.

As explained in  § 7.6, song segmentation and classification has been done by L3 intern (ENS) Pierre Marcus.

The on-going work on an original prototype based approach of deep-learning considering not so big data sets, and targeting also interpretability of the result, has been finalized [4], including a fine study on metaparameter adjustment in this context, while both standard learning and meta-learning paradigms have been considered. The capability to easily the "how it works" mechanism to no specialist of the field is an important outcome of the paper.

Co-leaded by Margarida Romero scientific director of the LINE laboratory of the UCA and researchers of our team, a preliminary work regarding artificial intelligence devoted to education (AIDE) was developed to study applications to educational science. This first year has been devoted to study to which extents the existing collaboration between Inria science outreach regarding computational thinking initiation and educational science research in order to be understand the underlying cognitive processes of the former actions and evaluate them, could be enlarged to multi-disciplinary research in both fields. The first outcome of this collaboration has been an analysis of a computational thinking unplugged activity under the perspective of embodied cognition [9] and deep and large review in the field, analyzing how computational thinking in K-12 education could be developed [22], within the scope of studies regarding co-creativity, robotics and maker education. [20], while a qualitative analysis one very large audience (more then 18000 inscriptions) on-line course outcomes has been published [18], with some operational outcomes regarding enlarging comptational thinking training from teachers to all citizens [12].

Software is a fundamental pillar of modern scientific research, across all fields and disciplines. However, there is a lack of adequate means to cite and reference software due to the complexity of the problem in terms of authorship, roles and credits. This complexity is further increased when it is considered over the lifetime of a software that can span up to several decades. Building upon the internal experience of Inria, the French research institute for digital sciences, we provide in this paper a contribution to the ongoing efforts in order to develop proper guidelines and recommendations for software citation and reference. Namely, we recommend: (1) a richer taxonomy for software contributions with a qualitative scale; (2) to put humans at the heart of the evaluation; and (3) to distinguish citation from reference.