[PhD subject] Reasoning over Data: Analogy-based and Transfer Learning to improve Machine Learning

Context and motivation: Analogical reasonning and transfer learning.
Analogy establishes parallels between two situations, which are similar in many respects and dissimilar in others. Analogical inference relies on the idea that if four items a, b, c, d are in analogical proportion for each of the n attributes describing them, it may still be the case for another attribute. For instance, if class labels are known for a, b, c and unknown for d, then one may infer the label for d
as a solution of an analogical proportion equation. This idea has been successfully applied in different inference tasks including classification problems. The effectiveness of analogical classifiers [2, 8] looks quite challenging and mysterious. In the case of Boolean attributes, a first step for explaining this state of facts was to characterize the set of functions for which the analogical inference is sound, i.e., no error occurs, no matter which triples of examples are used. In [3], it was proved that these so-called “analogy-preserving” (AP) functions coincide exactly with the set of affine Boolean functions. More recently, it has been shown that AP functions are “quasi-linear” in case of nominal attributes [5]. Analogical inference support classification tasks in either predicting or recommending classes for new items, or enlarging training sets [2] for learning classifiers [3], especially in environments with few labeled examples.
Transfer learning can be viewed as a kind of analogical reasoning performed at the meta level. The idea is to take advantage of what has been learnt on a source domain in order to improve the learning process in a target domain related to the source domain. When studying a new problem or a new domain, it is natural to try to identify a related, better mastered, problem or domain from which, hopefully, some useful information can be called upon for help. The emerging area of transfer learning is concerned with finding methods to transfer useful knowledge from a known source domain to a less known target domain.
This thesis proposal is motivated by recent advances in deep learning where neural networks achieve very good results in image classification, speech recognition, and natural language processing. Many methods are developed to analyze and explore this data in a supervised setting. However, most of the produced data does not come with metadata and/or labeling. This raises two possible directions of investigation: (i) learning to classify using a few labeled examples, and (ii) transferring what was learned in a given domain into another domain. Here we will consider transfer learning as the process of training a model on a large-scale dataset and then using that pre-trained model to perform learning in another domain and for another task. Transfer learning was popularized in the field of pattern recognition and language understanding [9]. It has also been used for transferring galaxy morphology information from one large survey to another [10].

Thesis work: reasoning and transferring.
This research work is original and lies at the intersection of Knowledge Representation and Reasoning (KRR) and of Knowledge Discovery (KD) [1]. It is aimed at reasoning over data and transferring and reusing reasoning process as they are applied to knowledge bases over data, for facilitating and improving machine learning procedures. Reasoning takes advantages of domain knowledge and applies inference rules for producing new knowledge units, or implicit units are made explicit. When reasoning by analogy [3, 4, 11], the goal is to adapt the solution of a known or source problem, which is sufficiently similar to the actual or target problem, involving a transfer between the context of a source problem (the problem and its solution)and the context of the target problem. Following such a line, we want to develop a similar approach in knowledge discovery, following the work done in transfer learning for deep learning, i.e. transferring solutions to new target problems, by reusing the adaptation functions that have been designed for reasoning (by analogy or case-based reasoning).
Two applications are envisioned for such a system, namely, in the context of the ANR AstroDeep project for recognizing celestial objects in large astronomy surveys and in the context of the IPL Project HyAIAI where we are investigating different means for integrating knowledge and reasoning in learning processes.

Organization of the thesis work:
Task 1. A study of analogy and of case-based reasoning, and their adaptation for reasoning over data. We want to understand the behaviour of analogical inference in wider domains and determine which
classifiers are compatible with analogical inference (i.e., analogy preserving).
Task 2. A framework combining deep learning with analogy-based and transfer learning for analyzing complex data. Implementation of analogy relations and transfer operations in the system.
Task 3. Consolidation of the KRR-KD based learning system. Application to image and text data and evaluation of the capabilities of the system.
Task 4. Writing of the thesis and related research papers, and seminar and conference presentations.

Thesis Environment:
Team: Orpailleur (Inria Nancy Grand Est/LORIA)
Supervision and contacts:
Miguel Couceiro (Miguel.Couceiro@loria.fr)
Alain Gély (Alain.Gely@loria.fr)
Amedeo Napoli (Amedeo.Napoli@loria.fr)

Keywords: data mining, knowledge discovery, deep learning, active learning, transfer learning, analogy-based reasoning.

Skills and profile:
A master’s degree in Computer Science. Elements of knowledge discovery and data mining algorithms (numerical and/or symbolic approaches), and statistical and deep learning methodologies are appreciated. Experience with programing in Python and deep learning environments is expected.

Application guidelines: Curriculum Vitae, transcript of records and letter of motivation. To be send to one of the contact persons by 15 of May 2020. Start: 01.10.2020 (3 years)

References:
[1] Z. Bouraoui, et al. From Shallow to Deep Interactions Between Knowledge Representation, Reasoning and Machine Learning (Kay R. Amel group), 2019. Preprint http://arxiv.org/abs/1912.06612
[2] M. Bounhas, H. Prade, G. Richard. Analogy-based classifiers for nominal or numerical data. Int. J. Approx. Reasoning, 91, 36-55, 2017.
[3] M. Couceiro, N. Hug, H. Prade, G. Richard. 2017. Analogy-preserving functions: A way to extend Boolean samples. Proc. 26th Int. Joint Conf. on Artificial Intelligence, IJCAI’17, Melbourne,Aug. 19-25, 1575- 1581.
[4] M. Couceiro, N. Hug, H. Prade, G. Richard. 2018. Behavior of analogical inference w.r.t. Boolean functions. Proc. 27th Int. Joint Conf. on Artificial Intelligence, IJCAI’18, July 13-19, Stockholm, 2057-2063.
[5] M. Couceiro, E. Lehtonen, L. Miclet, L.; H. Prade, G. Richard. When nominal analogical proportions do not fail. Submitted.
[6] J. Lieber, E. Nauer, H. Prade. Improving analogical extrapolation using case pair competence. Proc. 27th Int. Conf. on Case-Based Reasoning (ICCBR’19), (K. Bach, C. Marling, eds.), Otzenhausen, Germany, Sept. 8-12, Springer, LNCS 11680, 251-265, 2019.
[7] S. Lim, H. Prade, G. Richard. Solving word analogies: A machine learning perspective. Proc. 15th Europ. Conf. on Symbolic and Quantitative Approaches to Reasoning with Uncertainty (ECSQARU’19), (G. Kern-Isberner, Z. Ognjanovic, eds.), Belgrade, Sept. 18-20, 2019, Springer, LNCS 11726, 238-250, 2019.
[8] L. Miclet, S. Bayoudh, and A. Delhay. Analogical dissimilarity: definition, algorithms and two experiments in machine learning. JAIR, 32, 793-824, 2008.
[9] Zhilin Yang, et al.. GLoMo: Unsupervised Learning of Transferable Relational Graphs, Proc. Neural Information Processing Systems (NIPS’18), 8964–8975, 2018.
[10] Domìnguez-Sánchez, et al.. Transfer learning for galaxy morphology from one survey to another, Monthly Notices of the Royal Astronomical Society, 484(1), 93–100, 2018.
[11] B. Fuchs, J. Lieber, A. Mille, A. Napoli. Differential adaptation: An operational approach to adaptation for solving numerical problems with CBR. Knowledge Based Systems, 68, 103–114, 2014.

 

Logo d'Inria