Dynamiczna ocena zbieżności podobieństwa między kategoriami postrzeganego przez głębokie sieci neuronowe w wizji z podobieństwem semantycznym

The goal of the project is to develop methods for evaluating the quality of deep visual networks in terms of the convergence of perceived similarity between classes with semantic relationships during the training process. This method will enable the testing of models beyond simple accuracy, resulting in more predictable decisions even in the event of errors, and will potentially yield further benefits (e.g., the ability to assess resource utilization efficiency by mapping the structure of the world as accurately as possible, including through similarity [2]). The developed methods will have a positive impact on the interpretability of testing and will increase public trust in artificial intelligence. The project will include:

• an overview of available methods for measuring the similarity between (1) the categories perceived by networks and (2) reference categories (e.g., semantic relations in WordNet),
• development of methods for (1) extracting class patterns from within video networks, (2) determining similarity matrices for the networks and reference data, (3) comparing similarity matrices,
• impl. mechanizmów inspekcji zbieżności percepcji podobieństwa podczas trenowania,
• Implementation of mechanisms for inspecting the convergence of similarity perceptions during training.

Unlike other studies examining how networks perceive similarity (e.g., [1]), this project plans to use network parameters to describe known categories (such as the adversary attack evaluation metric in [3]), rather than error matrices generated on large-scale datasets. This approach allows for the evaluation of models without the use of test sets, making it possible to apply the method already during the training process. The project will result in a contribution to a proposal aimed at obtaining resources available within the PLGrid infrastructure, as well as to a publication.

Related publications:

[1] Bilal, Alsallakh, et al. " Do convolutional neural networks learn class hierarchy?" IEEE trensactions on visualization and computer graphics 24.1 (2017): 152-162

[2] Rosch, Eleanor, and Barbara B. Lloyd "Cognition and categorization" (1978)

[3] Filus Katarzyna, and Joanna Domańska "NetSat: Network Saturation Adversarial Attack" 2023 IEEE International Conference on Big Data (BigData). IEEE, 2023. 

Machine-translated
Przetłumaczone maszynowo