Welcome to the 20th edition of the WatchTower! In this edition, we explore how the principles of physics can be utilised to advance AI image generation and discuss some impressive capabilities demonstrated by GPT-4 in the field of ophthalmology.

📰 Featured in This Edition:

• How Physics Is Revolutionizing AI Image Generation

• Can GPT-4 Replace Eye Doctors?

How Physics Is Revolutionizing AI Image Generation

In a groundbreaking development, MIT researchers have introduced Poisson Flow Generative Models (PFGM++) that leverage principles of physics to enhance AI image generation. These models integrate concepts from both diffusion models and electrostatics, presenting a novel approach to creating high-quality images with greater efficiency and accuracy.

Traditional diffusion models, while powerful, often require substantial computational resources and are prone to errors. PFGM++ addresses these limitations by using electrically charged particles to streamline the image generation process. This innovative method reduces the number of computational steps needed, making it up to 20 times more efficient than existing diffusion models​​​​.

The PFGM++ model operates by extending the electric field into a higher-dimensional framework, balancing robustness and ease of use. This balance is achieved through a new training method that improves the learning efficiency of the electric field, allowing the model to generate images with superior quality compared to current state-of-the-art techniques​​.

One of the key advantages of PFGM++ is its ability to produce images that are less prone to errors and exhibit higher resistance to perturbations. This robustness is critical for applications requiring high precision and reliability, such as digital content creation and generative drug discovery​​.

The introduction of PFGM++ not only enhances image generation but also uncovers a surprising connection between electrostatics and diffusion models. This breakthrough provides new theoretical insights, paving the way for further advancements in generative AI​​.

Looking ahead, the MIT team aims to refine the PFGM++ model to identify optimal values for various data types and applications. This ongoing research promises to unlock even more potential in the realm of AI-driven image and pattern generation, marking a significant step forward in technology​​.

This innovative approach highlights the transformative power of interdisciplinary research, blending physics and AI to push the boundaries of what's possible in image generation.

_{Published by Ziming, July 08 2024}

Can GPT-4 Replace Eye Doctors?

Introduction

Since the launch of ChatGPT last year, we have seen the power of LLMs from text generating to programming assistance. A recent research paper has shown that GPT4 can be applied to healthcare. In the trials, it is demonstrated that it has the best performance among other existing LLMs and can approach expert-level clinical knowledge and reasoning in ophthalmology.

Details

The goal of the study is to test the performance of GPT4 in medical knowledge and reasoning. In the study, various LLMs, expert ophthalmologists and doctors in training were trialed on a set of questions from the Fellowship of the Royal College of Ophthalmologists (FRCOphth) examinations to gauge the ophthalmological knowledge base and reasoning capability of LLMs compared to experts. The exam questions were not freely available online, which means it is not likely to be trained on the models. Then, the responses were recorded and graded on a Likert scale from 1-5 (representing very bad to very good) by ophthalmologists in terms of the accuracy, relevance and overall preference. Results showed that the performance of GPT4 was superior to other LLMs like GPT3.5, LlaMA and PaLM 2. On the other hand, GPT4’s score compared favorably with junior doctors, trainee doctors and expert ophthalmologists, with significantly superior performance compared to junior and trainee doctors and statistically similar score attained by expert ophthalmologists.

Looking Ahead

This stunning result demonstrated GPT4’s ability in not just remembering facts but also applying and analyzing. Also, LLMs are approaching human expert-level knowledge and reasoning in some domains. Not to mention that more powerful models like GPT-4o have emerged in recent months, not only ophthalmology but also the whole healthcare system is destined to be changed. 

_{Published by David Hung, July 08 2024}

Sponsors

Our ambitious projects would not be possible without the support of our GOLD sponsor, UNOVA.

Closing Notes

We welcome any feedback / suggestions for future editions here or email us at [email protected].

Stay curious,

Sources

• From physics to generative AI: An AI model for advanced pattern generation

• Large language models approach expert-level clinical knowledge and reasoning in ophthalmology: A head-to-head cross-sectional study