November 25, 2026
Europe/Bratislava timezone

Concept Exploration and Modeling in Deep Neural Networks

Not scheduled
20m
Študenti Informatika Poster session + coffee: works of students of the faculty Informatics

Speaker

Mhd Walid Al Jallad

Description

Deep neural networks (DNNs) learn rich visual representations, but the concepts encoded in those representations and the relationships among them can be difficult to inspect and manipulate. We explores a topology-preserving, prototype-based framework for modeling visual concepts in DNN feature spaces. It places a self-organizing map (SOM) over features extracted from convolutional or foundation-model encoders, organizing prototypes so that nearby locations on the map correspond to similar representations. The resulting structure is intended to provide a navigable concept space: a way to examine how visual patterns cluster, relate, and vary, while connecting individual prototypes to the broader geometry of the representation.

We investigate how this structure could support concept identification and relations between concepts, and provide a foundation for compositional reasoning. In particular, graded prototype activations can be treated as fuzzy memberships, allowing logical operations to combine concepts while taking their map-based relationships into account. The framework also motivates future extensions to top-down concept learning, few-shot and semi-supervised learning, and continual or open-world recognition, where new concepts must be incorporated alongside existing ones. We will evaluate the proposed methods on benchmark vision datasets and compare their concept organization and interpretability with relevant prototype-based approaches. By bringing topographic organization and prototype-based modeling to DNN feature spaces, this work aims to provide a shared representational basis for exploring, relating, and composing visual concepts.

Tlač postru/ Print poster Budem požadovať tlač /I hereby required to print the poster in faculty

Authors

Dr Kristina Malinovska (Department of Applied Informatics, FMPI CU) Mhd Walid Al Jallad

Presentation materials

There are no materials yet.