Feature Learning by Least Generalization
Hien D. Nguyen and Chiaki Sakama
30th International Conference on Inductive Logic Programming as an event of the
1st International Joint Conference on Learning & Reasoning (IJCLR),
Lecture Notes in Artificial Intelligence, vol. 13191, 2021.
This paper provides an empirical study for feature learning based on induction.
We encode image data into first-order expressions and compute their least generalization.
An interesting question is whether the least generalization can extract a common pattern of input data.
We also introduce two different methods for feature extraction based on symbolic manipulation.
We perform experiments using the MNIST datasets and show that the proposed methods successfully capture features
from training data and classify test data in around 90% accuracies.
The results of this paper show potentials of induction and symbolic reasoning to feature learning or pattern recognition from raw data.
Full Paper (PDF 1646K)