Machine Learning Street Talk (MLST)’s Post

The future is NeuroSymbolic. The inefficiency of linear approximation by brute force is just too overwhelming. Trying to multiply that inefficiency by using a system that only uses the most inefficient information encoding, namely "Human Language", only amplifies that problem. Its like inefficiency on top of inefficiency. This is why intelligence or autonomous systems will get more and more out of reach, the more people try to solve it end-to-end gradient-based. The future is intelligently designed cognitive architectures.

In our research, we believe that Reinforcement Learning can have a huge impact on discovering new algorithms: in our specific case, distributed algorithms. And it seems that Google DeepMind shares our vision! Last month, at EPFL, Dr. Alhussein Fawzi, a Google DeepMind researcher, presented his latest results on using Reinforcement Learning to discover new matrix multiplication algorithms. A very interesting and insightful talk that I share here for the ones interested:

🔍 Today, in The Mindmap 🧠, I'm exploring Neuralink—a fascinating challenge that's changed my understanding and sparked my imagination. Exploring this innovative technology has opened my eyes to a future that feels like it's been plucked from the pages of science fiction. Join me to uncover its possibilities, and dive into my mind map 🦾 See link to the full mindmap in link below 👇 https://bit.ly/3usA4pb

One of the major potential issues working with top models like gpt2-chatbot can be the level of confidence these modes give to you because of their extreme capacity and knowledge. These models give you a sense of new achievement even if it is not real. Here, we invented a new physics explaining quantum paradoxes in a novel way. The key thing about it was not not plausible text, but a verifiable experiment setups with predictions against regular model. Very impressive!

Last weeks recap of our Arxiv Dive into Mechanistic Interpretability of Transformers is up. Next we are looking at diving into some computer vision papers or multi-modal language models, let us know any papers you would like to see! https://lnkd.in/gPxtWbiq

Hello ML enthusiasts, In this video, we explore a fascinating use case of novel vision transformers for indoor navigation. The content emphasizes the significance of feature engineering and demonstrates how data can be tailored using cutting-edge deep learning methods, specifically the vision transformer. While vision transformers offer advantages over traditional CNNs, it's worth noting their bulkiness and higher computing power requirements. The VITAL framework introduces a pruned vision transformer tailored for edge devices, striking a balance between accuracy and resource efficiency. https://lnkd.in/gt4AWF3q Access the full paper here :

The dualization problem of positive (or elsewhere called monotone) boolean functions is a prominent problem with applications on machine learning, AI, game theory and databases. Determining the complexity status of the dualization problem was an open problem for more than 30 years. We shed a new light on this problem by showing classical and quantum computing algorithms that have polynomial time complexity on the asymptotic dimensions of the problem. Here is the full paper published on Quantum Machine Intelligence https://rdcu.be/dYrkq https://lnkd.in/dcr_CVYu

If you're working with LLMs this is essential reading. LLMs are a black box. Up until we haven't really been able to figure out how they function internally. Think about that for a moment... The only real comparison we have is how we function. Data in, *some brain magic* Information out. We're only just starting to understand how our brains do this magic, and the same can be said for LLMs.

Excited to share my technical blog post with you all! 🚀 Check out the insightful content I've crafted on the topic "Neural networks have been used in Scientific Machine Learning (SciML) to forecast black hole dynamics with the outcomes in a Generative adversarial network (GAN)." Looking forward to your feedback and thoughts on these articles! Please click here for the blog: https://lnkd.in/gZyYwnrZ

🚀 Excited to share a groundbreaking new blog post on the intricacies of understanding neural network features! Check out "Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning" where we delve into the core challenge of mechanistic interpretability. Discover how E2E sparse dictionary learning offers a significant leap, elucidating network behavior concisely and accurately. Read the full post and access our library for training E2E SAEs and reproducing our analysis at https://bit.ly/4bOvMZe. Let's unlock the secrets of network interpretability together!