Prismer: A Scalable Vision-Language Model Using Pre-Trained Experts
Prismer is an innovative vision-language model that offers multi-modal generation abilities without the need for training large models on massive datasets. By leveraging a diverse ensemble of pre-trained domain experts and freezing their network weights during training, Prismer significantly reduces the training requirements. This article explores the features and advantages of Prismer.
1. Data and Parameter Efficiency:
Prismer inherits network weights from publicly available pre-trained domain experts and only requires training a few components. This approach reduces the need for extensive training data and computational resources, making it a scalable alternative to large pre-trained models.
2. Multi-Modal Expertise:
Prismer excels in visual language tasks by utilizing its projected multi-modal signals. It can perform visual question answering and picture captioning, demonstrating its multi-modal thinking expertise. Prismer splits complex reasoning tasks into smaller, manageable chunks for efficient processing.
3. Visually Conditioned Text Generation:
Researchers have developed a visually conditioned autoregressive text generation model that enhances the utilization of pre-trained domain experts for vision-language reasoning tasks. Despite being trained on only 13M examples, Prismer performs exceptionally well in tasks like image captioning, image classification, and visual question answering.
The Prismer model consists of an encoder-decoder transformer. It employs a visual encoder and an autoregressive language decoder. The vision encoder takes in RGB and multi-modal labels as input and generates RGB and multi-modal features as output. The language decoder generates a sequence of text tokens based on cross-attention training.
Prismer offers several benefits, including efficient use of training data. By leveraging pre-trained vision-only and language-only backbone models, Prismer achieves comparable performance to other state-of-the-art vision-language models with significantly fewer GPU hours. The model also utilizes multi-modal signal input for better semantic understanding of images.
Inclusion of Pre-Trained Specialists:
Prismer incorporates two types of pre-trained specialists: vision-only and language-only models. These experts efficiently translate images and text into meaningful tokens, contributing to the model’s overall performance. As the number of modality specialists increases, Prismer’s performance improves.
Enhancement through Corrupted Experts:
Researchers experiment by introducing corrupted depth experts, where a fraction of predicted depth labels is replaced with random noise. Prismer’s performance remains unaffected by the inclusion of noise-predicting experts.
Prismer is a scalable vision-language model that taps into the expertise of pre-trained domain experts. Its efficient use of data, multi-modal capabilities, and incorporation of specialists make it a powerful tool for various vision-language reasoning tasks. Prismer’s superior performance and learning habits make it a promising solution in the field of AI research.
About the Author:
Dhanshree Shenwai is a Computer Science Engineer with experience in the FinTech industry. She has a deep interest in AI applications and the exploration of new technologies.