Symbolic machine learning techniques for explainable AI

Machine learning (ML) approaches such as encoder-decoder networks and LSTM have been successfully used for numerous tasks involving translation or prediction of information (Otter et al, 2020). However, the knowledge obtained by these techniques is learned in an implicit form, which makes it difficult to review or verify. In addition, large correlated datasets are needed for training using these techniques, which may not be available for some applications (Transcoder, 2021).

Current research at King’s is applying symbolic function discovery techniques to the learning of data transformation and code generation functionality from input/output examples (Lano et al., 2020). The project will investigate extensions of this approach to define symbolic machine learning approaches for general sequence-to-sequence and tree-to-tree mapping tasks, and evaluate these in comparison to implicit knowledge learning techniques such as LSTM.

Applications include learning of machine translation rules from examples, learning of program translation and code generation rules, and programming-by-example.

(Transcoder, 2021) Facebook Research, github.com/ facebookresearch/TransCoder, 2021.

(Lano et al, 2020) K. Lano, et al., “Enhancing model transformation synthesis using natural language processing”, MDE Intelligence workshop, MODELS 2020.

(Otter et al, 2020) D. Otter et al., “A survey of the usages of deep learning in natural language processing”,
IEEE Trans. Neural network learning systems, April 2020, pp. 1–21.

Z. Balogh, D. Varro, Model transformation by example using inductive logic programming, SoSyM, 2009.

R. Bodik, B. Jobstmann, Algorithmic program synthesis: introduction. International Journal on Software Tools for Technology Transfer volume 15, pages 397–411 (2013).

K. Lano, et al., “Enhancing model transformation synthesis using natural language processing”, MDE Intelligence workshop, MODELS 2020.

Project ID

STAI-CDT-2021-KCL-15

Supervisor

Kevin Lanohttps://nms.kcl.ac.uk/kevin.lano

Category

AI Planning, Verification