Stanford Institute for Human-Centered Artificial Intelligence (HAI)’s Post

A new tool for optimizing compound AI models makes it easy to combine the broad knowledge base of large language models with the specialized capabilities of scientific tools, allowing researchers to leverage the strengths of both. https://lnkd.in/gHudmEmZ

Very useful across many scenarios.

AI agents help explain other AI systems | MIT News | Massachusetts Institute of Technology: Researchers from MIT's CSAIL have developed an innovative approach using AI models to conduct experiments on other systems and explain their behavior. The automated interpretability method, involving an "automated interpretability agent" (AIA) and a new benchmark called "function interpretation and description" (FIND), aims to evaluate and improve interpretability procedures for large language models. The team presented their work at NeurIPS 2023. - Artificial Intelligence topics! #ai #artificialintelligence #intelligenzaartificiale

What if you could understand how complex neural networks behave? MIT's researchers have made it possible! They've created a way for AI to explain complex neural networks easily. Using AI agents, the team built a system where these agents use language models to clarify how trained networks make decisions. Their tool, named FIND, tests these explanations against real-world network functions. This breakthrough makes AI not just powerful but also more understandable and trustworthy. I think we will be seeing a lot more using AI to explain AI systems in this era of intelligence. #AI #NeuralNetworks #MITInnovation #XAI

Can we look inside Neural Network black box? Decoding AI Secrets: MIT Unveils Breakthrough in Neural Network Explanations. MIT's CSAIL introduces an innovative AI-driven approach to demystify complex neural networks. Using automated interpretability agents (AIAs) built from language models, the method explains computations inside trained networks, shedding light on their behavior. Explore the groundbreaking research in AI transparency and interpretability. #AI #AIResearch #NeuralNetworks #MITCSAIL https://lnkd.in/dfDWazuf

Does Ai even know how it works? ⚙️ One big concern many people have over the use of neural networks is we don’t really know why they give the answers they do. To solve this, the CSAIL team (Computer Science and Artificial Intelligence Laboratory) at MIT have developed an Ai tool to understand neural networks such as Large Language Models (LLMs). The ‘automated interpretability agent’ (AIA) acts like a researcher, coming up with a hypothesis and experimentation plan, running the tests, learning and iterating on the process, refining its understanding of the system and providing a detailed analysis of the results. Sounds like a pretty familiar process to me! I like 2024! 😁 https://buff.ly/3NR9JHU #behavioralscience #systemsthinking #Ai #designthinking #sustainability

Interesting. This approach seems similar to Google DeepMind evolving AlphaGo by deleting the 150,000 games used to initially train it to play Go, retaining the rest of its self-learning, and then allowing it to teach itself. We don't completely understand how AI neural networks generate the results they do, but we're starting to learn how the black box functions. Making the models more closely resemble our human memorizing/forgetting/learning processes is more computationally/energy efficient and better at adaptation. --via Quanta Magazine/WIRED #AI #languagelearning #translations #neuralnetworks https://lnkd.in/eXuMjGZ3

It's fascinating to see the progress being made with interpretable AI/ML, as highlighted in this article from MIT Press. Some methods already exist for "unwrapping" Deep Neural Network black boxes into more understandable forms such as collections of interconnected linear models, but this approach essentially introduces an AI agent into the experimentation and explanation loop. I think we are pretty close now to having the ability to train, interpret, and diagnose ML models using semi-autonomous AI agents accessed through chat bots, which has the potential to massively reduce model development time and open many previously locked doors to adoption. https://lnkd.in/egNCymhY

LLM and GNN: How to Improve Reasoning of Both AI Systems on Graph Data: Graph neural networks (GNNs) and large language models (LLMs) have emerged as two major branches of artificial intelligence, achieving… Continue reading on Towards Data Science » #MachineLearning #ArtificialIntelligence #DataScience

At the end of the day it seems that the Feynman Technique is applicable to LLMs after all, and that explaining makes learning better - or at least inference better. Like the Feynman Technique, you understand better what you are capable to explain... The way I see it, the Chain-of-Thought approach shows that the complexity of the world has to be chewed, to make it intelligible. Who would have thought that explainability would come hand in hand with better predicitions? #AI #GenAI #Weekend #Longread #tech #innovation