Filters








4 Hits in 1.7 sec

AdaVQA: Overcoming Language Priors with Adapted Margin Cosine Loss [article]

Yangyang Guo and Liqiang Nie and Zhiyong Cheng and Feng Ji and Ji Zhang and Alberto Del Bimbo
2021 arXiv   pre-print
Experimental results demonstrate that our adapted margin cosine loss can greatly enhance the baseline models with an absolute performance gain of 15\% on average, strongly verifying the potential of tackling  ...  To this end, an adapted margin cosine loss is designed to discriminate the frequent and the sparse answer feature space under each question type properly.  ...  AdaVQA: Overcoming Language Priors with Adapted Margin Cosine Loss -Supplementary Material-1 Proposed AdaVQA Formulation of Partial Derivatives Let p i = s(cos θ i − m i ) = s( W T i ||Wi||2 · x ||x|  ... 
arXiv:2105.01993v1 fatcat:dnuoxucxt5ctbc6ztdp5ik6upa

AdaVQA: Overcoming Language Priors with Adapted Margin Cosine Loss

Yangyang Guo, Liqiang Nie, Zhiyong Cheng, Feng Ji, Ji Zhang, Alberto Del Bimbo
2021 Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence   unpublished
Experimental results demonstrate that our proposed adapted margin cosine loss can enhance the baseline models with an absolute performance gain of 15\% on average, strongly verifying the potential of tackling  ...  An adapted margin cosine loss is designed to discriminate the frequent and the sparse answer feature space under each question type properly.  ...  Based upon this observation, we argue that an adapted cosine margin is more favorable for overcoming the language priors in VQA.  ... 
doi:10.24963/ijcai.2021/98 fatcat:ng2365jpx5be7eh6dypushkp5q

Visual Perturbation-aware Collaborative Learning for Overcoming the Language Prior Problem [article]

Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, Yan Yan
2022 arXiv   pre-print
Several studies have recently pointed that existing Visual Question Answering (VQA) models heavily suffer from the language prior problem, which refers to capturing superficial statistical correlations  ...  prior problem by learning the instance-level characteristics.  ...  [39] addressed the problem by introducing an adapted margin cosine loss for different answers in the angular space under the corresponding question type, which effectively separates the answer embeddings  ... 
arXiv:2207.11850v1 fatcat:by74nzm64zaljbztkjh4xgg5g4

On Modality Bias Recognition and Reduction [article]

Yangyang Guo, Liqiang Nie, Harry Cheng, Zhiyong Cheng, Mohan Kankanhalli, Alberto Del Bimbo
2022 arXiv   pre-print
In addition, to overcome this problem, we propose a plug-and-play loss function method, whereby the feature space for each label is adaptively learned according to the training set statistics.  ...  After stepping into several empirical analysis, we recognize that one modality affects the model prediction more just because this modality has a spurious correlation with instance labels.  ...  In the next subsection, we will introduce a more sophisticated adapted margin cosine loss to overcome this issue.  ... 
arXiv:2202.12690v2 fatcat:bw2wbi4ombfwrbdudxbawfv3lq