Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-Language Models

By Jize Cao, Zhe Gan, Yu Cheng, Licheng Yu, Yen-Chun Chen, Jingjing Liu. 2020

143 citations recorded by the source corpus.

Recent Transformer-based large-scale pre-trained models have revolutionized vision-and-language (V+L) research. Models such as ViLBERT, LXMERT and UNITER have significantly lifted state of the art across a wide range of V+L benchmarks with joint image-text pre-training. However, little is known about the inner mechanisms that destine their impressive success. To reveal the secrets behind the scene of these powerful models, we present VALUE (Vision-And-Language Understanding Evaluation), a set of meticulously designed probing tasks (e.g., Visual Coreference Resolution, Visual Relation Detection, Linguistic Probing Tasks) generalizable to standard pre-trained V+L models, aiming to decipher the inner workings of multimodal pre-training (e.g., the implicit knowledge garnered in individual attention heads, the inherent cross-modal alignment learned through contextualized multimodal embeddings). Through extensive analysis of each archetypal model architecture via these probing tasks, our key observations are: (i) Pre-trained models exhibit a propensity for attending over text rather than images during inference. (ii) There exists a subset of attention heads that are tailored for capturing cross-modal interactions. (iii) Learned attention matrix in pre-trained models demonstrates patterns coherent with the latent alignment between image regions and textual words. (iv) Plotted attention patterns reveal visually-interpretable relations among image regions. (v) Pure linguistic knowledge is also effectively encoded in the attention heads. These are valuable insights serving to guide future work towards designing better model architecture and objectives for multimodal pre-training.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Multimodal Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Multimodal Learning

Citations in the corpus

TAP: Text-Aware Pre-training for Text-VQA and Text-Caption (2020, methodology): “For the former, we reﬁne the pre-training tasks in VLP [37, 34] to support the extra scene text input.”
Large-Scale Adversarial Training for Vision-and-Language Representation Learning (2020, methodology): “To provide a more direct measurement on how well our adversarial pre-trained model captures such multimodal signals, we conduct a probing analysis following [7].”
Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts (2021, background): “Supporting evidence for this phenomenon is found in the recent work of [9], which observe that image+text pre-trained models exhibit a preference for attending text rather than images during inference (in image and text understanding task).”
Compressing Visual-linguistic Model via Knowledge Distillation (2021, background): “[7] show that certain attention matrices of the pre-trained VL models contain extensive intra-modal and cross-modal co-reference relations.”
ROSITA: Enhancing Vision-and-Language Semantic Alignments via Cross- and Intra-modal Knowledge Integration (2021, unknown): “The effect of fine-grained semantic alignments across modalities can be inferred from the attention maps of the learned Transformer model [5].”
History Aware Multimodal Transformer for Vision-and-Language Navigation (2021, background): “The reason of using such dual-stream architecture rather than one-stream is that the length of different modalities can be highly imbalanced, and the dual-stream architecture can balance the importance of intra-and inter-modal relationships by model design [47].”
Multimodal Learning With Transformers: A Survey (2022, methodology): “Why and how Transformers perform so well in multimodal learning has been investigated [106], [299], [300], [301], [302], [303], [304], [305], [306].”
Playing Lottery Tickets with Vision and Language (2021, background): “As shown in [3], the pre-trained weights indeed have captured rich visual coreference and visual relation knowledge.”
Effect of Vision-and-Language Extensions on Natural Language Understanding in Vision-and-Language Models (2021, result): “Cao et al. (2020) analysed linguistic probing results as a subsection of a comprehensive investigation of V&L models and showed that pretrained V&L model generally performed worse than the original LM, but some questions still remain following their study.”
Probing Contextual Language Models for Common Ground with Visual Representations (2020, background): “Closer to our use of probing for better understanding grounding is the work of Cao et al. (2020), that design probes for examining multi-modal models.”
Unsupervised Vision-and-Language Pre-training Without Parallel Images and Captions (2021, background): “Probing tasks (Cao et al., 2020) confirm that they capture”
UNITER: UNiversal Image-TExt Representation Learning (2019, background): “VALUE [6] developed a set of probing tasks to understand pre-trained models.”
A Closer Look at the Robustness of Vision-and-Language Pre-trained Models (2020, methodology): “Pre-trained models [39, 75, 59, 35, 37, 24] such as ViLBERT [43], LXMERT [64] and UNITER [13] have demonstrated great generalizability over diverse V+L tasks [80, 23, 47, 70, 4, 22, 46, 9], such as Visual Question Answering (VQA) [6], Visual Commonsense Reasoning (VCR) [78], and Referring Expression”
Seeing past words: Testing the cross-modal capabilities of pretrained V&L models on counting tasks (2020, background): “In the context of the VALUE benchmark, Cao et al. (2020) report results on multiple V&L tasks, some of which we also corroborate, notably, the dominance of textual features compared to visual features in the model’s predictions.”