New Flamingo Visual Language Model Masters Few-Shot Learning Tasks
Researchers introduce Flamingo, a visual language model that seamlessly processes interleaved images, videos, and text to achieve state-of-the-art few-shot learning. The model outperforms competitors by rapidly adapting to new tasks using only a handful of examples.
Researchers introduce Flamingo, a family of Visual Language Models (VLM) designed to rapidly adapt to novel tasks using only a handful of annotated examples. The model bridges powerful pretrained vision-only and language-only systems to process sequences of arbitrarily interleaved visual and textual data. This architectural innovation allows Flamingo to seamlessly ingest both images and videos as inputs.
The system trains on large-scale multimodal web corpora that contain mixed text and images, which is key to endowing it with in-context few-shot learning capabilities. By simply prompting the model with task-specific examples, a single Flamingo model adapts to a wide spectrum of open-ended and close-ended tasks. These tasks include visual question-answering, scene captioning, and multiple-choice visual queries.
Evaluations show that Flamingo achieves a new state of the art in few-shot learning across numerous benchmarks. Remarkably, this flexible model outperforms other AI systems that undergo fine-tuning on thousands of times more task-specific data. The research demonstrates a significant leap forward in creating multimodal machine learning systems that learn efficiently from limited information.