We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models.

We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models. Our main focus is int8 as GPU support for float8 is rare, though we also analyze float8 training through simulation. While SwitchBack proves effective for float8, we show that standard techniques are also successful if the network is trained and initialized so that large feature magnitudes are discouraged, which we accomplish via layer-scale initialized with zeros. As a result, we recommend an AdamW-Adafactor hybrid which avoids loss spikes when training a CLIP ViT-Huge model and outperforms gradient clipping at the scales we test.

The rapid advancement of large-scale vision-language models has showcased remarkable capabilities across various tasks. However, the lack of extensive and high-quality image-text data in medicine has greatly hindered the development of large-scale medical vision-language models. In this work, we present a diagnosis-guided bootstrapping strategy that exploits both image and label information to construct vision-language datasets. Based on the constructed dataset, we developed MedDr, a generalist foundation model for healthcare capable of handling diverse medical data modalities, including radiology, pathology, dermatology, retinography, and endoscopy. Moreover, during inference, we propose a simple but effective retrieval-augmented medical diagnosis strategy, which enhances the model's generalization ability. Extensive experiments on visual question answering, medical report generation, and medical image diagnosis demonstrate the superiority of our method.

In this study, we develop a simple daily assistive robot that controls its own vision according to linguistic instructions. The robot performs several daily tasks such as recording a user's face, hands, or screen, and remotely capturing images of desired locations. To construct such a robot, we combine a pre-trained large-scale vision-language model with a low-cost low-rigidity robot arm. The correlation between the robot's physical and visual information is learned probabilistically using a neural network, and changes in the probability distribution based on changes in time and environment are considered by parametric bias, which is a learnable network input variable. We demonstrate the effectiveness of this learning method by open-vocabulary view control experiments with an actual robot arm, MyCobot.