Good evening, everyone. My name is Cao Yun.

Introduction

So today I'm going to talk about how to use a large language model to analyze thousands of scientific publications from AI conferences to make the scientific publications easily digestible to the readers with all different kind of background and truly unlock the value of AI conference for everyone.

So before I delve into the specific topic, I want to introduce myself. I have been working in the industry delivering large scale AI solutions including search and conversational AI for more than 10 years. I work as a manager and individual contributor in AI science organization. I had a PhD degree in computer science and I'm very passionate about knowledge sharing and innovations in technology.

Understanding the Value of AI Conferences

1So yes, we want to talk about how to unlock the value of AI conferences and first of all I want to give you an overview on the AI conferences. So every year there are more than 60 top tier AI conferences worldwide.

For every conference, there can be more than 2,500 publications. So by the end of the conference, all the publications will be archived into the conference proceeding, normally in PDF format.

And for every conference there can be more than 9,000 participants from academia, from industry, top lab worldwide. So some of the key players in the conference include leading industry research labs such as Google and Meta Lab, and also include top universities labs worldwide.

Key Highlights of AI Conferences

And I want to show you some of the examples of the publications from the AI conferences. So you can see a lot of hot topics today. They are originally published from the AI conferences.

So you probably heard about this paper, Attention is All You Need. So whenever people talk about large language model, people believe that this is a paper originally presented the large language model architecture. So this is published on Neural Nibs in 2017 by Google.

And then the image generation. So this is the paper published by OpenAI in 2021, ICML 2021. So this is the paper for the product Dolly 2. So today if you get on ChatGPT and prompt ChatGPT to generate an image for you, the ChatGPT actually behind the scene connect to a DALL-E model, but it's the latest DALL-E model and this is a publication in 2021 on DALL-E 2 from OpenAI.

So this is a paper more recently published in 2024 ICML. So this is about video game generation. So this is like less mature technology, so it's probably less popular than the large language model and image generation. But it's a very interesting paper.

Significant Publications in AI

So I'm going to get online. All right.

So this one. So the idea is that, so if you gave the AI system a picture and the picture has a background as an environment and has an object on the picture, then the AI system will automatically generate a video game and in that video game you can use the keyboard to interact with the object and to control the object to interact with the background environment.

Harnessing AI for Scientific Analysis

1So yeah, so we see that the AI conferences are very valuable in many perspectives, but there are thousands of publications from every conference, right? And actually we can leverage large language model to help us easily understand the publications.

So I list three functions here. The first one is comprehensive conference coverage. So you can use large language model to generate a concise summary for thousands of papers.

from the conference and you can classify the papers into different categories such as reinforcement learning, large language models, diffusion models. From a high level, you can understand what's being discussed in the AI conference.

And for individual paper, you can also generate a easily digestible summary to actually help you understand the individual paper.

And you can also use the RAC system, right, so the Retrieval Augmented Generative AI. to build a search engine or chatbot to give you instant answer on any research questions, right? So it's going to search, first to search the relevant papers to your question and then answer the question based on the relevant papers.

Applications and Use Cases of AI

So an example from the large language model analysis on the ICML 2024 paper proceeding.

So in terms of use cases, right, so there's games and robotics being discussed in the conference, right, so we can use AI to create realistic 3D models, videos, and images. And we can also use AI to predict the shape and the properties of tiny materials to assist the process of design new materials.

So AI can foresee the future to predict sales, stock price, or weather. It can help discovering content through smart search for image, videos, and articles.

And AI can also help drug discovery and improving health. by analyzing medical data in text image to detect illness earlier for the patients.

Key Topics Discussed in ICML 2024

In terms of methodology, so the ICML 2024 discussed reinforcement learning, large language models, so those are the hot topics. It also discussed like differential privacy, diffusion models and graph neural networks. So the topics are pretty diverse in the conference.

Introducing Arcnode.com

So I implement all the functions I mentioned earlier on a website called arcnode.com and you can access from internet. So arcnode.com can help you easily understand AI concepts from the conference.

It provides comprehensive overviews of top papers. So currently it has more than 15,000 papers indexed from three different conferences in 2024 and I'm onboarding more conferences onto the website.

And it also has a chatbot to provide instant answer for any research questions.

Features and Capabilities of Arcnode

So yes, I'll show you the arcnode.com and also I want to show you some learning I got from the conference content and I use that learning to build a animal image recognition, the animal image recognition system. So the problem statement is actually inspired by the first Lego competition.

So it is a Lego robotics competition for students. And this year the topic is submerged. So it encouraged the student to use technology to improve the ocean's ecosystem.

So that's why I picked that problem setting.

All right. All right.

Demonstration of Arcnode Interface

So this is the website, right?

And when you get on arcnode.com, you will land on ICML 2024. And there are three conferences, NeurONIPS and CIGIR.

So you can see that I categorized the papers into like the different categories like reinforcement learning, large language model. So next to the reinforcement learning, there is an explain button.

And if you click on that explain button, It's going to give you an intuitive explanation on what is reinforcement learning. So it talks about reinforcement learning through a use case. and also talks about the key architecture components and process for reinforcement learning, so a little bit deeper dive into the concept.

Yeah, and you can check out the explain for other concepts, large language model, graph neural networks, etc. So, and yeah, and there is a overview for all the papers in the conference.

You can see the use cases, right, including designing new materials, etc. And it has the problem statements, the methodology being discussed in the conference, and the future directions.

And if we, under the reinforcement learning, if we expand on the methods tab, so you can see, right, so there are 587 papers talking about reinforcement learning. And among those papers, 276 mentioned reinforcement learning family, of course, and 90 mentioned policy optimization methods. So it's going to give you a deep dive into the AI methods and the reinforcement learning category.

And so there is a copy button. So if we copy policy optimization method and input into the search bar. So then So it's gonna give you a explanation of, uh, policy optimization method, right, in detail.

And, um, on the bottom, So in the references section, you will see the list of papers relevant to the policy optimization method. And then from here, you can deep dive into individual papers.

And on the top, I also provide a Hugging Face tutorial for the method. The reason I provide the Hugging Face tutorial because I think Hugging Face tutorial provide you a very sort of good hands-on lab, right? So if you learn about this method and you learn about the papers and you really want to implement the method, right, so you can go to the tutorial and it's going to provide you like some sample code and for you to actually take actions.

And now let's say we want to deep dive into individual paper, right? So we can copy the paper name for the video generation paper and input in the search bar. So then you will have a paper summary for the paper.

So it's concise and very structured. So all the papers have the very similar structure of the summary. So you don't have to like switch between different writing styles.

You can directly understand the paper here. And if you have further question after reading the summary, you can also ask a question in the below search bar. So for example, we can ask what is, okay, what is mask GIT, right?

So, yeah, so if we ask a specific question, uh, it's going to give you an answer only based on this paper as the context. And then, so say, uh, we have a general question about all the papers in the, um, in the conference, we can ask, um, what are the easy methods for image classification? So yeah, so it's gonna give you like the different methods such as CMF classifier, zero-shot classification using vision language models such as CLIP,

um, and deep image clustering method. Um, so you are- you're gonna get the different, um, methods and also in the references you get the, uh, list of papers relevant to the question. And actually, if you go to multi-model machine learning, so you can also see the CLIP model being mentioned there.

So the CLIP model actually being mentioned in many places in the overview and in the AI methods. So it's actually like a very popular method in computer vision, but because my background is also not computer vision, so this is something new that I learned actually from this conference.

So if you go to the Hugging Face tutorial, right, so you can see... All right. So you can see some sample code there. And the sample code is, so this is very simple, right?

So you just give an image and you give the labels that you want to assign the image to, and it's going to achieve the classification task, right? So this is a pre-trained model. And then, so I just want to simply show you, right?

So I copy, paste, this code, right, so into Google Colab. So all I did was replacing the image with the specific sea animal image and I found the sea animal dataset from Kaggle dataset and yeah, and then you can achieve the AI system for sea animal image classification.

Conclusion

So yeah, that's all for my talk and you can visit arcnode.com to check it out. You can also connect with me through email or connect me on LinkedIn.

Yeah. Thank you.