Michael Sergi-Curfman

Frontiers | Global explanation supervision for Graph Neural Networks The paper discusses various methods for explaining and understanding neural networks. Several authors have proposed techniques to generate explanations for graph neural networks, including GNNExplainer, XGNN, and RelEx. Other researchers have focused on developing methods for attributing the output of a model to specific input features or relationships within the data, such as few-shot learning with per-sample rich supervision. Graph Neural Networks: Foundations, Frontiers, and Applications provides an overview of the current state of graph neural networks. The book covers the basics of GNNs, as well as their applications in various fields. The automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain is discussed in relation to neuroimage analysis. A probabilistic graphical model explanation for graph neural networks, PGM-explainer, is also introduced. Learning deep attribution priors based on prior knowledge is another technique that has been explored by researchers. The Wu-Minn Human Connectome Project provides an overview of the project's objectives and methods. The project aims to investigate human brain development and function across various life stages. Various authors have contributed to this research, including N. Tzourio-Mazoyer, D. Papathanassiou, and B. Landeau. The work presented in these papers showcases the diversity of techniques being developed to explain and understand neural networks, from graph neural networks to probabilistic graphical models. https://ift.tt/iMDS8aU #news #cto #tech

1

The idea behind model2vec (distill-pca) encoder library is incredible. 15x less storage, 500x faster inference. We can quickly ingest tons of documents and get query results in seconds. One of the standout benefits of our lean 𝗿𝗮𝗴𝗽𝗶𝗽𝗲 framework is its ability to simplify the model evaluation process, making it possible to rapidly test and compare different configurations. Adding a new model is quick and DIY! No need to wait for the RAG framework maintainers to add the new model. 1. Create a ragpipe plugin yourself by directly copying encoding pipeline from the new model's README code.  2. Modify few lines of the older project config yaml to point to the new encoder plugin 3. The new pipeline is ready to run/evaluate! So here's some quick results with the new, blazing fast model2vec encoder on a benchmark with 40k documents. Compared with the popular baai/bge-small embeddings, see how the time to ingest documents diverges between baai/bge-small and model2vec, as the number of documents increases. model2vec encodes even 40k documents in a few seconds. — Join us as an early contributor and help shape this lean and flexible RAG framework! If you’re facing challenges with LLM-driven RAG, don’t hesitate to reach out in the ragpipe Discord channel for a conversation. Discord: 👋https://lnkd.in/g4J_eqZW Github: github.com/ekshaks/ragpipe

18

3 Comments

For those engineers that don't know, AI is revolutionizing control systems with high-performance models like neural networks, but there's a catch—these "black-box" models make decisions we can't easily explain. In critical fields like autonomous vehicles or healthcare, knowing why an AI acted a certain way is just as important as the decision itself. While #Black-Box AI, delivers high accuracy by learning complex patterns but lacks transparency, it’s great when performance is the top priority, but it’s hard to trust in safety-critical situations since we can’t fully explain its actions. On the other hand, #Explainable AI (XAI) sacrifices some performance but provides clear, understandable reasons behind decisions. In industries where accountability and safety matter, like industrial automation or healthcare, XAI builds trust by making AI actions interpretable. In control applications, unexplained failures can lead to serious consequences. XAI ensures we can track, diagnose, and correct issues, making AI more reliable and trustworthy. So, what's your take—performance or transparency or do you think there should be something new that has a tradeoff? Let’s discuss! #ExplainableAI #ControlSystems #AI #Transparency #XAI #NeuralNetworks

1

We are often asked the (good :)) question of how the latest Gen AI advances will impact autonomous driving. Here's a quick sneak peak. From the start of Nuro we have invested in leveraging and building off of the latest and greatest AI advancements. And today our on-road behavior is almost exclusively AI driven -- with "almost" being very important, based on a robust parallel real time validation system we will geek out over in subsequent posts. We're now expanding beyond integrating the fundamental AI breakthroughs behind language models (which we have done for years) towards integrating language models themselves into the stack. It is so cool to see where we have collectively gotten as an industry and what we can already do with this technology. And just wait for what's next...

58

1 Comment

In this episode, we discuss Cycle3D: High-quality and Consistent Image-to-3D Generation via Generation-Reconstruction Cycle by Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, Li Yuan. Recent 3D large reconstruction models often generate low-quality and inconsistent multi-view images, which harm the final 3D output. To resolve this, the proposed Cycle3D framework integrates a 2D diffusion-based generation module and a 3D reconstruction module to iteratively enhance texture quality and multi-view consistency. Experiments show that Cycle3D outperforms state-of-the-art methods in creating high-quality and consistent 3D content.

1

New algorithm from NVIDIA speeds up LLMs by up to 11x. Transformer-based Large Language Models (LLMs) have revolutionized natural language processing, but their quadratic complexity in self-attention poses significant challenges for inference on long sequences. Star Attention addresses this issue by introducing a two-phase block-sparse approximation that shards attention across multiple hosts while minimizing communication overhead. Here's how it works: 1. The context is split into blocks, each prefixed with an anchor block. 2. Context tokens are processed using blockwise-local attention across hosts, in parallel. 3. Query and response tokens attend to all prior cached tokens through sequence-global attention. The technique seamlessly integrates with most transformer-based LLMs trained with global attention, significantly reducing memory requirements and inference time (by up to 11x). As Star Attention is scaled to even longer sequences (up to 1M) and larger models, the speedups become even more impressive. However, there are still open questions around the optimal size of anchor blocks and performance on more complex long-context tasks. But the trend is clearly ongoing: faster models that require less memory. ↓ Liked this post? Join my newsletter with 50k+ readers that breaks down all you need to know about the latest LLM research: llmwatch.com 💡

486

13 Comments

Robotics and AI will touch all aspects of life and society. This one “could” be a path in reducing or eliminating the herbicides used on crops. On the technical side, from the article, this is astounding. Who would have thought even 10 years ago GPUs would be utilized in this manner. “Powered by 24 NVIDIA GPUs, Carbon Robotics’ vehicles use deep-learning-based computer vision models to autonomously identify and eliminate weeds via CO2 lasers. LaserWeeder, as the product is called, processes 4.7 million high-resolution images per hour, zapping 5,000 weeds per minute, according to the company. To date, its dataset includes 25 million labeled plants and more than 30,000 crop and weed models” #AI #agritech #robotics #nvidia

3

1 Comment

This week at [ICML] Int'l Conference on Machine Learning catch up with Anders Christensen and Nooshin Mojab at the GRaM Workshop https://lnkd.in/d9b76E-h Anders will be presenting results of our work on “Geometry Fidelity for Spherical Images” https://lnkd.in/dPScMk6y Part of the problem for immersive generative content is that, because of its rare format needs, often we rely on small datasets that hinder training, adding that traditional optimization functions don’t transfer so well to other formats. To achieve better results we must improve the way we evaluate our models. Here we introduce OmniFID to improve evaluation for spherical images. If you can’t make it to ICML, don’t worry we will have a more detailed paper with that and other improvements for spherical images at the main track of the European Conference on Computer Vision. Anders Christensen, Nooshin Mojab, Khushman Patel, Karan Ahuja, Zeynep Akata, Ole Winther, Mar Gonzalez-Franco, Andrea Colaco

22

2 Comments

In-Context Learning at Extreme Scales: My Thoughts As large language models (LLMs) continue to expand their context capacities, the domain of in-context learning (ICL) at extreme scales has piqued my interest. A recent paper by researchers at Carnegie Mellon University and Tel Aviv University provides a detailed exploration into this area, unveiling both impressive capabilities and notable challenges. 🔍 Key Insights: Performance Improvement: The study demonstrates that ICL can improve significantly by integrating hundreds to thousands of examples into the context window—far beyond traditional few-shot settings. This is particularly impactful in complex tasks like 77-way banking intent classification, where long-context ICL matches or surpasses traditional finetuning. Input Sensitivity: Long-context ICL shows reduced sensitivity to input shuffling compared to shorter contexts. However, organizing examples by label dramatically reduces performance, emphasizing the need for diverse context configurations. Source of Gains: Performance gains primarily arise from accessing more relevant examples, rather than aggregating knowledge across many examples during encoding. 📉 Challenges & Drawbacks: Sensitivity to Example Order: Despite reduced sensitivity, the performance still varies with input order, indicating potential instability in practical scenarios. Performance Saturation: The models often do not fully leverage their extended context capacities, showing performance saturation before reaching the maximum context length. Computational Cost: The increased computational cost associated with using large context windows could limit practical deployment, particularly in resource-constrained environments. Efficiency Issues: There is a noted inefficiency in how models utilize the encoded information within large datasets, suggesting that current architectures may need refinement to better use available data. 🚀 Future Directions: The paper suggests several exciting avenues for future research, including optimizing the trade-offs between computational costs and performance, exploring model adaptation strategies, and refining ICL approaches for specific applications or datasets. As we push the boundaries of what LLMs can achieve, effectively leveraging their full context could unlock new capabilities. This area is ripe for further exploration and could lead to major advancements in how we deploy AI. For those at the cutting edge of language model development, I highly recommend this paper for its thorough analysis and forward-looking perspective. Check out the full work at https://lnkd.in/d7AmGh4y. What are your thoughts on the future of in-context learning at such scales and the challenges discussed? #AI #MachineLearning #DeepLearning #LanguageModels #InContextLearning

17

Our work on time series anomaly detection - Explainable time series anomaly detection using masked latent generative modeling - is now available in Pattern Recognition (https://lnkd.in/dMTcvqjM). We also release the source code for our model and experiments: https://lnkd.in/dx6Q9_qd The method represents a major improvement on the previous SOTA in anomaly detection and introduces mechanisms to interpret and understand the results. Improved interpretability and understanding make it easier for practitioners to use the methodology in real-world settings. Daesoo Lee, Sara M. #machinelearning #ai #timeseries #anomalydetection

71

4 Comments

The next big thing that may replace diffusion in the field of image generation may be coming -- next-scale prediction. This paper, "Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction", is probably the most exciting paper I've read so far this year. Like always, I've got you covered with a 🔥1-minute read summary. ● What is VAR about? It proposes a new autoregressive image generation paradigm. Instead of predicting the next image tokens, it predicts the next image scale. ● What is the evaluation task and metric? Task: ImageNet Image Generation Benchmark. Metric: FID and IS. ● Why is it a milestone work? This is the first time that GPT-style image generation has beaten diffusion models on both image quality and inference speed. ● Why it has huge potential? Just like LLM, this model is proven to follow the scaling law. ● What's next? I can already foresee text-control VAR, video-based VAR, VAR for world models, super super-large VAR coming... For more insights, check out the paper here: https://lnkd.in/gGqdiS9H Modeling and the code source: https://lnkd.in/gTiaMjKG

310

In this episode, we discuss CYCLE-CONSISTENT LEARNING FOR JOINT LAYOUT-TO-IMAGE GENERATION AND OBJECT DETECTION by The paper's authors are listed as "Anonymous authors" since it is under double-blind review.. The paper introduces a new generation-detection cycle consistent (GDCC) learning framework that simultaneously optimizes layout-to-image generation and object detection, highlighting the inherent duality of these tasks. GDCC employs cycle losses to guide both tasks, enhancing data efficiency without requiring paired datasets, and achieves computational efficiency through novel sampling strategies while keeping inference cost unchanged. Experimental results demonstrate that GDCC improves diffusion model controllability and object detector accuracy, with plans for code release.

2

RAG of 2024 (ColbertV2 Reranking) !!! I've just published an in-depth guide on how to build an efficient local semantic search system using the innovative technologies of ColBERTv2, DSPy, and Qdrant . If you're interested in advancing your search capabilities with the latest AI technology, this article is for you. 🔍 Discover how the deep contextual understanding of ColBERTv2 and the dynamic optimization capabilities of DSPy can transform your data retrieval processes. Plus, learn about integrating Qdrant for scalable and precise search results. 📖 Check out the full article here: https://lnkd.in/gi_mQNH8 #SemanticSearch #AI #MachineLearning #DataScience #ColBERTv2 #DSPy #Qdrant #TechInnovation

20

2 Comments

So, is it as BIG as they say? 🤭 🚀 Excited to share a new LLM benchmark! RULER 😏 🔍 What happens to model performance as context length increases? RULER by NVIDIA dives deep, testing long-context capabilities in language models. 📉 Many models boast large context sizes, but how many hold up under pressure? How does it work? 🤔 RULER tests the models' performance degradation as context length increases and provides a more nuanced understanding of each model's capabilities. It reveals that while many models claim large context sizes, only a few can maintain satisfactory performance when tested with RULER's comprehensive and challenging tasks. RULER reveals only a few can truly perform well with extended texts. 🔗 Dive into the full paper: https://lnkd.in/dihNR7ZR 💡 Consider this benchmark when choosing a model for a task that requires a long context length. I've seen Gemini's 1.5 Pro with 1 mil token context length responses degrade after a while. Sometimes using a vector DB with similarity search on indexed content and a better model could lead to better results. Developed by @NVIDIA 👏 Big thanks to the NVIDIA team for pushing the boundaries of AI research! #AI #MachineLearning #NLP #LanguageModels #Innovation

1

DeepSeek-V2.5 Released by DeepSeek-AI: A Cutting-Edge 238B Parameter Model Featuring Mixture of Experts (MoE) with 160 Experts, Advanced Chat, Coding, and 128k Context Length Capabilities Read our full take on this: https://lnkd.in/g4MdekFs DeepSeek-AI has released DeepSeek-V2.5, a powerful Mixture of Experts (MOE) model with 238 billion parameters, featuring 160 experts and 16 billion active parameters for optimized performance. The model excels in chat and coding tasks, with cutting-edge capabilities such as function calls, JSON output generation, and Fill-in-the-Middle (FIM) completion. With an impressive 128k context length, DeepSeek-V2.5 is designed to easily handle extensive, complex inputs, pushing the boundaries of AI-driven solutions. This upgraded version combines two of its previous models: DeepSeekV2-Chat and DeepSeek-Coder-V2-Instruct. The new release promises an improved user experience, enhanced coding abilities, and better alignment with human preferences. Key Features of DeepSeek-V2.5 🔰 Improved Alignment with Human Preferences: One of DeepSeek-V2.5’s primary focuses is better aligning with human preferences. This means the model has been optimized to follow instructions more accurately and provide more relevant and coherent responses. This improvement is especially crucial for businesses and developers who require reliable AI solutions that can adapt to specific demands with minimal intervention. 🔰 Enhanced Writing and Instruction Following: DeepSeek-V2.5 offers improvements in writing, generating more natural-sounding text and following complex instructions more efficiently than previous versions. Whether used in chat-based interfaces or for generating extensive coding instructions, this model provides users with a robust AI solution that can easily handle various tasks. 🔰 Optimized Inference Requirements: Running DeepSeek-V2.5 locally requires significant computational resources, as the model utilizes 236 billion parameters in BF16 format, demanding 80GB*8 GPUs. However, the model offers high performance with impressive speed and accuracy for those with the necessary hardware. For users who lack access to such advanced setups, DeepSeek-V2.5 can also be run via Hugging Face’s Transformers or vLLM, both of which offer cloud-based inference solutions. Model: https://lnkd.in/g84ADnNV

33

At Voice of Reason AI, where we focus on interfaceless personal aids for individuals with memory and brain disorders, inference speed is not just a luxury—it’s a necessity. Our AI may remain silent most of the time, yet it must continuously listen, understand, and analyze context so it can respond instantly when help is needed. Think of it like the cymbal player in an orchestra, silent for long stretches but ready to deliver a perfectly timed crash. For us, it's essential that our AI “cymbal” sounds precisely when a user needs guidance. This makes Cerebras' advancements incredibly exciting, as they could enable us to create real-time, ambient aids that seamlessly enhance daily life without unnecessary delays. Such speed is crucial for AI that doesn’t merely react but anticipates needs in real-time, helping individuals maintain independence and dignity.

2

Combine contextual retrieval and reranking to maximize retrieval accuracy in RAG systems 👾I converted it to a podcast using NotebookLM and heard it on my commute https://lnkd.in/gXRKBgPs this is my usual way now to read papers/tech blogs 🃏Highlights: 1. Embeddings + BM25 is better than embeddings on their own 2. Passing the top-20 chunks to the model is more effective than just the top-10 or top-5 3. Adding context to chunks improves retrieval accuracy a lot 4. Reranking is better than no reranking Summary: Combine contextual embeddings with contextual BM25, plus a reranking step, and adding the 20 chunks to the prompt 🔗https://lnkd.in/g2PGg_ca

8

"NeuralVDB presents a novel method for high-resolution sparse volume representation using hierarchical neural networks. This approach efficiently encodes the sparse structures inherent in volumetric data, enabling scalable storage and rapid processing. By leveraging hierarchical architectures, NeuralVDB enhances performance in tasks such as rendering, simulation, and 3D data manipulation. The framework offers improved compression and retrieval capabilities, making it ideal for applications in computer graphics, virtual reality, and scientific visualization. Overall, NeuralVDB advances sparse volume representation by integrating neural network hierarchies to manage complexity while maintaining high fidelity."

1