arXiv subjects

Katarzyna Filus

Publications and source records attributed to Katarzyna Filus.

5 recordsLinked to original sources

arXiv2026

Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherence

Within Explainable Artificial Intelligence, mechanistic interpretability uses Sparse Autoencoders (SAEs) to extract more interpretable features from neural representations. However, assessing their monosemanticity, and thus explanation quality, remains challenging. Existing metrics require external concept labels or depend on pretrained embedding models, making them sensitive to encoder's geometry. We introduce the Tversky Monosemanticity Score (TMS), a label-free metric that operationalizes monosemanticity as activation-set coherence of binarized SAE latents, and does not require external embedding encoders. We evaluate TMS on SAEs trained on features from pretrained vision and vision-language models (DINOv3, CLIP, BLIP2), two common SAE regimes (TopK, BatchTopK), multiple sparsity levels, and expansion factors. Our results show that TMS is less affected by encoder anisotropy than its embedding-based alternative, while remaining aligned with established monosemanticity indicators. TMS also reveals distinct SAE training dynamics across base models. Moreover, under encoder anisotropy, TMS provides a stronger indication of probe-based concept deletion effectiveness, while being competitive otherwise.

cs.CV↗

arXiv2026

Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning

Continual learning studies how models can adapt to new tasks while retaining previously acquired knowledge. Although a broad spectrum of methods has been proposed to mitigate catastrophic forgetting, the field remains predominantly performance-driven, with limited insight into what forgetting actually corresponds to within the vision model's representation space. Prior work has primarily analyzed forgetting through task-level performance or coarse measures of representational drift, without disentangling output-level accessibility from changes in finer-grained internal structure. To this end, we propose a diagnostic framework that leverages Sparse Autoencoders (SAEs) to define a task-anchored latent feature space, enabling analysis of how task-specific information evolves at a finer granularity, where individual SAE latents are treated as concept proxies for recurring and relatively disentangled visual patterns in the model's internal computations. Within this framework, we decompose forgetting into apparent concept deletion, recoverability, and decodability. We show that a large portion of seemingly lost concept-level information can often be recovered under linearity assumption, with concept decodability degrading as more tasks are introduced. Overall, our findings suggest that a significant part of concept-level forgetting can be attributed to changes in the representational accessibility rather than complete information erasure.

cs.LG↗

arXiv2025

Semantically Guided Adversarial Testing of Vision Models Using Language Models

In targeted adversarial attacks on vision models, the selection of the target label is a critical yet often overlooked determinant of attack success. This target label corresponds to the class that the attacker aims to force the model to predict. Now, existing strategies typically rely on randomness, model predictions, or static semantic resources, limiting interpretability, reproducibility, or flexibility. This paper then proposes a semantics-guided framework for adversarial target selection using the cross-modal knowledge transfer from pretrained language and vision-language models. We evaluate several state-of-the-art models (BERT, TinyLLAMA, and CLIP) as similarity sources to select the most and least semantically related labels with respect to the ground truth, forming best- and worst-case adversarial scenarios. Our experiments on three vision models and five attack methods reveal that these models consistently render practical adversarial targets and surpass static lexical databases, such as WordNet, particularly for distant class relationships. We also observe that static testing of target labels offers a preliminary assessment of the effectiveness of similarity sources, \textit{a priori} testing. Our results corroborate the suitability of pretrained models for constructing interpretable, standardized, and scalable adversarial benchmarks across architectures and datasets.

cs.CV↗

arXiv2025

Inspecting Training Dynamics of Similarity Development in Supervised Vision Networks

For trustworthy and human-aware artificial intelligence, models should be evaluated beyond accuracy, among others through error predictability and semantic alignment. Similarity is central to these aspects, as it influences which classes a model considers related and confusable. Similarity manifests in multiple forms, including semantic similarity, which can serve as a proxy for human similarity perception. While similarity perception is often imposed in computer vision, little attention has been paid to its natural emergence during supervised training. Existing studies are largely limited to static and qualitative analyzes and lack a systematic, training-time perspective. Therefore, we analyze how similarity perception evolves and aligns with model error patterns and semantics in supervised vision networks. As an enabler, we introduce Deep Similarity Inspector (DSI) - a systematic, training-time framework that unifies complementary views on similarity within a single methodology. Using DSI, we analyzed Convolutional and Transformer-based Networks and showed that both architectures develop rich similarity structures through three phases - initial similarity surge, refinement, stabilization - while exhibiting clear differences. We also identified the mistake refinement phenomenon, in which networks improve mistakes with time.

cs.CV↗

arXiv2025

Semantic Depth Matters: Explaining Errors of Deep Vision Networks through Perceived Class Similarities

Understanding deep neural network (DNN) behavior requires more than evaluating classification accuracy alone; analyzing errors and their predictability is equally crucial. Current evaluation methodologies lack transparency, particularly in explaining the underlying causes of network misclassifications. To address this, we introduce a novel framework that investigates the relationship between the semantic hierarchy depth perceived by a network and its real-data misclassification patterns. Central to our framework is the Similarity Depth (SD) metric, which quantifies the semantic hierarchy depth perceived by a network along with a method of evaluation of how closely the network's errors align with its internally perceived similarity structure. We also propose a graph-based visualization of model semantic relationships and misperceptions. A key advantage of our approach is that leveraging class templates -- representations derived from classifier layer weights -- is applicable to already trained networks without requiring additional data or experiments. Our approach reveals that deep vision networks encode specific semantic hierarchies and that high semantic depth improves the compliance between perceived class similarities and actual errors.

cs.CV↗