Mon Mar 27 2023
Thu Mar 23 2023

Capabilities of GPT-4 on Medical Challenge Problems

Medicine
Natural Language Processing
Medical AI
Medical education
Clinical practice
Medical assessment

GPT-4, without any specialized prompt crafting, exceeds the passing score on USMLE by over 20 points and outperforms domain-expert models like Med-PaLM.

GPT-4 can be used in medical education, assessment, and clinical practice. It has potential to improve accuracy and safety.

The Quantization Model of Neural Scaling

Neural Networks
Artificial Intelligence
AI model scaling
Scaling strategies

The model explains both the observed power law dropoff of loss with model and data size, and also the sudden emergence of new capabilities with scale.

The model can help explain the behavior of large neural networks and may lead to improvements in scaling strategies for AI models.

Ablating Concepts in Text-to-Image Diffusion Models

Deep Learning
Computer Vision
Natural Language Processing
Image generation
Copyrighted images

The paper proposes an efficient method of ablating concepts in a pretrained model to prevent the generation of copyrighted or unwanted images.

The proposed method can help prevent the generation of copyrighted or unwanted images without retraining the model from scratch.

CoBIT: A Contrastive Bi-directional Image-Text Generation Model

Vision and language pre-training models
Computer vision
Natural language processing
Image classification
Text-to-image generation
Captioning

Presents a model that unifies three pre-training objectives (contrastive, image-to-text, and text-to-image) and achieves superior performance in various image and text tasks.

Can improve image understanding, image-text understanding, and text-based content creation in business operations, particularly in zero-shot scenarios.

DreamBooth3D: Subject-Driven Text-to-3D Generation

Personalizing text-to-image models
Text-to-3D generation
Computer vision
3D modeling
Virtual product prototyping
Marketing

Presents an approach to personalize text-to-3D generative models from a few captured images of a subject.

Can help businesses create high-quality, subject-specific 3D assets with text-driven modifications for various applications such as virtual product prototyping and marketing.

Wed Mar 22 2023
Tue Mar 21 2023
Mon Mar 20 2023
Sun Mar 19 2023