Relační reasoning ve vision-language modelech
Thesis title in Czech: | Relační reasoning ve vision-language modelech |
---|---|
Thesis title in English: | Relational reasoning in vision-language models |
Academic year of topic announcement: | 2023/2024 |
Thesis type: | Bachelor's thesis |
Thesis language: | |
Department: | Department of Theoretical Computer Science and Mathematical Logic (32-KTIML) |
Supervisor: | RNDr. Jakub Bulín, Ph.D. |
Author: | hidden - assigned and confirmed by the Study Dept. |
Date of registration: | 07.03.2024 |
Date of assignment: | 12.03.2024 |
Confirmed by Study dept. on: | 12.03.2024 |
Opponents: | Mgr. Jindřich Libovický, Ph.D. |
Guidelines |
Vision-language models have made significant strides in interpreting and generating descriptions from visual data. However, their ability to perform complex relational reasoning remains a challenge. Relational reasoning involves understanding the relationships between different entities within a context, which is crucial for tasks such as visual question answering and scene understanding. Santoro et al. proposed a simple neural network module designed specifically to enhance relational reasoning in neural networks [1]. This thesis aims to explore the effectiveness of such modules when integrated into vision-language models.
The objective of this bachelor thesis is to implement a vision-language model enhanced with a relational reasoning module as described by Santoro et al. in "A simple neural network module for relational reasoning" (2017) [1]. The student will benchmark this enhanced model against a standard vision-language model, focusing on performance in relational reasoning tasks. This comparative analysis will help in understanding the impact of integrating relational reasoning capabilities into vision-language models. |
References |
[1] Santoro, Adam et al. “A simple neural network module for relational reasoning.” Neural Information Processing Systems (2017).
[2] McCallum, Andrew et al. “Chains of Reasoning over Entities, Relations, and Text using Recurrent Neural Networks.” Conference of the European Chapter of the Association for Computational Linguistics (2016). [3] Hu, Ronghang et al. “Learning to Reason: End-to-End Module Networks for Visual Question Answering.” 2017 IEEE International Conference on Computer Vision (ICCV) (2017): 804-813. |
Preliminary scope of work in English |
The goal of the thesis is to implement and benchmark a vision-langauge model with a relational module as presented in an earlier work by Santoro et al. and compare its performance to a pure vision-language model on a relational reasoning dataset. |