PREreview of "The Scattering Compositional Learner: Discovering Objects, Attributes, Relationships in Analogical Reasoning"

This Zenodo record is a permanently preserved version of a PREreview. You can view the complete PREreview at https://prereview.org/reviews/18371910. Summary The research investigates how people use analogical reasoning when they solve Raven's Progressive Matrices (RPM) which contains complex structural elements. The authors introduce the Scattering Compositional Learner (SCL) which uses sequential neural network module composition to represent visual reasoning compositional structures. The proposed method reaches the highest performance level on two benchmark datasets Balanced-RAVEN and PGM through its relative improvements which exceed all previous methods. The research shows that SCL produces performance improvements while simultaneously learning meaningful object attribute patterns which enhance its ability to handle domain changes and achieve excellent performance on new analogical patterns it has not seen before. The research proves that architectural compositionality leads to enhanced visual reasoning abilities through its creation of enhanced generalization and robustness techniques. Major issues The paper needs to demonstrate why its recommended compositional sequence produces superior results than other modular and relational architectures for achieving better generalization performance. The research requires direct model assessment and removal tests with models that have equivalent compositional reasoning abilities to determine cause-and-effect relationships. The research presents promising results about how learned representations of attributes and relations become interpretable but the assessment of compositionality needs more organized methods. The research needs additional quantitative assessment methods and experimental tests to confirm the discovered structural model. The paper presents solid results for robustness and zero-shot generalization but it needs to establish the exact boundaries of this robustness system by showing which domain variations and untested relationships continue to be difficult to handle. Minor issues The architectural description contains complex sections which need a general diagram to show how the sequential composition process works. The research needs to present experimental evidence about data segmentation methods and evaluation procedures which researchers applied to assess domain shift and zero-shot performance. The sections about compositional representations and analysis results need minor language and presentation adjustments to achieve better clarity. Competing interests The author declares that they have no competing interests. Use of Artificial Intelligence (AI) The author declares that they did not use generative AI to come up with new ideas for their review.