Modular visual question answering via code generation – Google Research Blog

Visual question answering (VQA) is a machine studying process that requires a mannequin to reply a question about a picture or a set of pictures. Conventional VQA approaches want a considerable amount of labeled coaching knowledge consisting of hundreds of human-annotated question-answer pairs related to pictures. In latest years, advances in large-scale pre-training have led to the event of VQA strategies that carry out nicely with fewer than fifty coaching examples (few-shot) and with none human-annotated VQA coaching knowledge (zero-shot). However, there may be nonetheless a major efficiency hole between these strategies and state-of-the-art absolutely supervised VQA strategies, reminiscent of MaMMUT and VinVL. In specific, few-shot strategies wrestle with spatial reasoning, counting, and multi-hop reasoning. Furthermore, few-shot strategies have typically been restricted to answering questions on single pictures.

To enhance accuracy on VQA examples that contain advanced reasoning, in “Modular Visual Question Answering via Code Generation,” to look at ACL 2023, we introduce CodeVQA, a framework that solutions visual questions utilizing program synthesis. Specifically, when given a question about a picture or set of pictures, CodeVQA generates a Python program (code) with easy visual capabilities that enable it to course of pictures, and executes this program to find out the reply. We display that within the few-shot setting, CodeVQA outperforms prior work by roughly 3% on the COVR dataset and a couple of% on the GQA dataset.

CodeVQA

The CodeVQA method makes use of a code-writing massive language mannequin (LLM), reminiscent of PALM, to generate Python packages (code). We information the LLM to accurately use visual capabilities by crafting a immediate consisting of an outline of those capabilities and fewer than fifteen “in-context” examples of visual questions paired with the related Python code for them. To choose these examples, we compute embeddings for the enter question and of all the questions for which now we have annotated packages (a randomly chosen set of fifty). Then, we choose questions which have the best similarity to the enter and use them as in-context examples. Given the immediate and question that we need to reply, the LLM generates a Python program representing that question.

We instantiate the CodeVQA framework utilizing three visual capabilities: (1) question, (2) get_pos, and (3) find_matching_image.

• Query, which solutions a question a few single picture, is carried out utilizing the few-shot Plug-and-Play VQA (PnP-VQA) methodology. PnP-VQA generates captions utilizing BLIP — an image-captioning transformer pre-trained on hundreds of thousands of image-caption pairs — and feeds these right into a LLM that outputs the solutions to the question.
• Get_pos, which is an object localizer that takes an outline of an object as enter and returns its place within the picture, is carried out utilizing GradCAM. Specifically, the outline and the picture are handed by means of the BLIP joint text-image encoder, which predicts an image-text matching rating. GradCAM takes the gradient of this rating with respect to the picture options to search out the area most related to the textual content.
• Find_matching_image, which is utilized in multi-image questions to search out the picture that finest matches a given enter phrase, is carried out by utilizing BLIP textual content and picture encoders to compute a textual content embedding for the phrase and a picture embedding for every picture. Then the dot merchandise of the textual content embedding with every picture embedding signify the relevance of every picture to the phrase, and we decide the picture that maximizes this relevance.

The three capabilities will be carried out utilizing fashions that require little or no annotation (e.g., textual content and image-text pairs collected from the net and a small variety of VQA examples). Furthermore, the CodeVQA framework will be simply generalized past these capabilities to others {that a} consumer may implement (e.g., object detection, picture segmentation, or information base retrieval).

Illustration of the CodeVQA methodology. First, a big language mannequin generates a Python program (code), which invokes visual capabilities that signify the question. In this instance, a easy VQA methodology (question) is used to reply one a part of the question, and an object localizer (get_pos) is used to search out the positions of the objects talked about. Then this system produces a solution to the unique question by combining the outputs of those capabilities.

Results

The CodeVQA framework accurately generates and executes Python packages not just for single-image questions, but additionally for multi-image questions. For instance, if given two pictures, every exhibiting two pandas, a question one may ask is, “Is it true that there are four pandas?” In this case, the LLM converts the counting question concerning the pair of pictures right into a program during which an object rely is obtained for every picture (utilizing the question operate). Then the counts for each pictures are added to compute a complete rely, which is then in comparison with the quantity within the authentic question to yield a sure or no reply.

We consider CodeVQA on three visual reasoning datasets: GQA (single-image), COVR (multi-image), and NLVR2 (multi-image). For GQA, we offer 12 in-context examples to every methodology, and for COVR and NLVR2, we offer six in-context examples to every methodology. The desk beneath reveals that CodeVQA improves constantly over the baseline few-shot VQA methodology on all three datasets.

We discover that in GQA, CodeVQA’s accuracy is roughly 30% increased than the baseline on spatial reasoning questions, 4% increased on “and” questions, and three% increased on “or” questions. The third class contains multi-hop questions reminiscent of “Are there salt shakers or skateboards in the picture?”, for which the generated program is proven beneath.

img = open_image("Image13.jpg")
salt_shakers_exist = question(img, "Are there any salt shakers?")
skateboards_exist = question(img, "Are there any skateboards?")
if salt_shakers_exist == "sure" or skateboards_exist == "sure":
    reply = "sure"
else:
    reply = "no"

In COVR, we discover that CodeVQA’s acquire over the baseline is increased when the variety of enter pictures is bigger, as proven within the desk beneath. This pattern signifies that breaking the issue down into single-image questions is useful.

Conclusion

We current CodeVQA, a framework for few-shot visual question answering that depends on code generation to carry out multi-step visual reasoning. Exciting instructions for future work embody increasing the set of modules used and creating the same framework for visual duties past VQA. We word that care must be taken when contemplating whether or not to deploy a system reminiscent of CodeVQA, since vision-language fashions like those utilized in our visual capabilities have been proven to exhibit social biases. At the identical time, in comparison with monolithic fashions, CodeVQA gives extra interpretability (by means of the Python program) and controllability (by modifying the prompts or visual capabilities), that are helpful in manufacturing programs.

Acknowledgements

This analysis was a collaboration between UC Berkeley’s Artificial Intelligence Research lab (BAIR) and Google Research, and was carried out by Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein.