May 20, 2024 Presentation Open Access

AI Playground for INFN Scientific Use Cases

Mauro Gattari

Talk given at "Workshop sul Calcolo nell'INFN", Palau (Sassari), 20-24/05/2024

Title: 

AI Playground: a curated collection of technologies offered “as a Service” on top of INFN Cloud for fast prototyping Machine Learning solutions across INFN research areas. 

Abstract: 

The introduction of ChatGPT in November 2022 has gained widespread attention and significantly boosted Generative AI adoption in technological solutions, highlighting the potential of AI to automate tasks, analyze large datasets, and make predictions with high accuracy. 

The fast-paced adoption of AI techniques has also been possible by the development and general availability of AI frameworks, libraries and platforms that provide structured approaches that make it easier to implement AI solutions. 

The integration of AI and Machine Learning (ML) in the Physics domain is also becoming increasingly pervasive, transforming the way scientists approach and solve complex problems. For instance, ML techniques in the High Energy Physics (HEP) domain are ubiquitous, successfully used in many areas and are playing a significant role in LHC Run 3 and in the future High-Luminosity LHC upgrade. 

INFN always stands at the frontier’s edge of the most innovative technological advancements, hence supporting AI as a promising approach across the diverse research areas.  

However, AI adoption requires researchers at INFN to solve not only problems related to the specificity of the application/experiment (e.g., tailored models and specialized domain knowledge), but also requires solving general infrastructure-level and ML-workflow related problems. 

In this regard, we introduce AI Playground, a curated collection of technologies offered “as a Service” on top of INFN Cloud, for fast prototyping Machine Learning solutions across INFN research areas. 

AI Playground leverages INFN Cloud resources and principles by providing an open-source solution to INFN users that can be deployed through the INFN Cloud Dashboard. 

The general idea behind the design of AI Playground is to address common use cases within the institute, collect reliable and consolidated technologies to solve these problems, then offer these technologies within the playground so that scientists can easily prototype their AI solutions for use cases that benefit of the same technologies. 

In this contribution we introduce the principles and high-level architecture of AI Playground and address two use cases in different domains that have been prototyped within the playground. 

The first use case is in the NLP domain: we expose through an INFN Cloud HTTP endpoint a RAG (Retrieval Augmented Generation) pipeline: RAG is a popular technique for injecting knowledge into a Large Language Model (LLM). We describe the RAG pipeline implemented through on-premises model serving of open source LLMs. 

The second use case is in the HEP domain: we expose through an INFN Cloud HTTP endpoint a model for inference related to a signal-vs-noise discrimination problem about data generated by particle collisions. 

The two use cases belong to different research areas but leverage the same AI Playground technologies. 

AI Playground is currently a work in progress, the aim is that its application-agnostic nature will serve as a unified ecosystem where developers, data scientists, and domain experts can leverage a standardized framework for ML model development and deployment. Hopefully, the playground will eliminate the need for extensive domain expertise in every application area, empowering a broader audience to leverage the benefits of machine learning, breaking down barriers and fostering innovation across diverse research domains.