Doktorandské kolokvium KAI - Tamara Bíla (28.4.2025)
v pondelok 28.4.2025 o 13:10 hod. v miestnosti I/9
Prednášajúci: Tamara Bíla
Názov: In Search of Meaning: Unveiling Interpretable Concepts in Neural Representations
Termín: 28.4.2025, 13:10 hod., I/9
Abstrakt:
In the atmosphere of the ever-growing presence of artificial neural networks in the public as well as private domains, the demand and need for explainability are self-evident. From the perspective of decision transparency, the aim is to introspect and report on the reasoning and key factors behind the network's output. To this end, we investigate several neural network architectures trained for classification on image datasets. Our approach builds on the previously developed Concept Activation Regions (CAR) method combined with hierarchical agglomerative clustering. By employing these two techniques in tandem, we aim to gain a multi-faceted understanding of the structure within the networks' internal representations. In particular, by comparing supervised against unsupervised learning algorithms, we intend to expose the potential human bias when assessing the extracted concepts from the latent representation space.