Agustin Pecorari

The paper "Polynomial Time Cryptanalytic Extraction of Neural Network Models" by Isaac Andrés Canales Martínez and colleagues introduces a new method for efficiently extracting all parameters of ReLU-based deep neural networks (DNNs). Unlike previous methods that required exponential time, this approach uses techniques that only need polynomial time and queries, making it much more practical. This advancement helps us better understand and audit DNNs, while also revealing the vulnerabilities of current black-box models. At Datalytica, we turn cutting-edge research like this into practical tools. Our solutions provide unmatched insights and actionable intelligence, transforming AI transparency and trust. Empower your organization with these advancements to stay ahead in the fast-changing world of AI.

13

1 comentario

Our third inference acceleration technique recommendation is pretty straightforward: ⚙️ 𝐂𝐨𝐧𝐬𝐢𝐝𝐞𝐫 𝐚𝐝𝐨𝐩𝐭𝐢𝐧𝐠 𝐚 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞 𝐚𝐧𝐝 𝐩𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧-𝐚𝐰𝐚𝐫𝐞 𝐚𝐩𝐩𝐫𝐨𝐚𝐜𝐡 𝐭𝐨 𝐦𝐨𝐝𝐞𝐥 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭. Production-aware model development takes into account the inference hardware very early in the development stage, particularly, during model selection. Also known as hardware in a loop development approach, it considers the hardware the deep learning models are going to be deployed on, the most suitable software stack for the target hardware, and the limitations. – What’s the #11DaysofInferenceAccelerationTechniques? The Deci team is posting, for 11 days, a series of inference acceleration techniques for deep learning applications. If you’re looking for practical tips and best practices for improving inference, follow Deci AI so you won’t miss an update. #deeplearning #machinelearning #neuralnetworks #computervision

17

From our AI CTO, Miguel Amigot II: 🚀 By leveraging Prompt Adaptation, LLM Approximation, and LLM Cascade, ibl.ai optimizes AI performance while cutting costs. 🎯 Precise prompts minimize computational load, fine-tuned smaller models handle specialized tasks, and our tiered querying system ensures efficient resource use. 🌿 Transcript: At ibl.ai, we harness the power of AI efficiently and responsibly. Let’s explore three advanced techniques we use to optimize costs while enhancing the capabilities of large language models. Prompt Adaptation is our first technique. By crafting concise, optimized prompts, we minimize computational load. For example, instead of using lengthy few-shot prompts with multiple examples, we select the most effective ones. This not only reduces processing costs but also enhances the model's focus and effectiveness. Let’s consider email classification. We transform a bulky prompt into a lean one by selecting the top five most relevant past examples using semantic similarity tests. This approach reduces token count significantly, slashing costs by up to 70%. Next is LLM Approximation. Rather than defaulting to high-cost models for all tasks, we employ a strategic use of cached responses and fine-tune smaller models for specific functions. This method maintains high accuracy but with lower operational costs. For instance, when recurrent queries or similar tasks arise, we retrieve answers from a cache, which saves up to 95% in costs. Additionally, by fine-tuning smaller models on specialized tasks, we achieve or surpass the performance of larger models without the associated expense. The third technique is LLM Cascade. We start with the simplest model that might provide a satisfactory answer and escalate to more complex models only if necessary. This approach ensures we use the minimal computational resources required for each query. Using an email sorting task, we first query a basic model. If the confidence score is sufficient, we stop there. If not, the query moves up to more sophisticated models. This tiered querying significantly cuts costs while ensuring quality responses. At ibl.ai, these techniques—Prompt Adaptation, LLM Approximation, and LLM Cascade—are more than just cost-saving measures. They are part of our commitment to sustainable AI use, ensuring that our technologies not only lead in innovation but also in efficiency and responsibility.

3

⚡️ PyTorch Lightning AI is the #DeepLearning framework for professional #AI researchers and #MachineLearning engineers who need maximal flexibility without sacrificing performance at scale. 🤝 And it's seamlessly integrated with Comet! 👇 Check out the full code tutorial to get started:

1

An Overview of Advancements in Deep Reinforcement Learning (Deep RL) https://buff.ly/3Jy0c5W #datascientist #engineering #automation #software #programmer

1

New Post: Deepdive: Custom Collate in DataLoader, Gemini cookbook, Build LLM from scratch Optimize your large language model (LLM) training with PyTorch by using a custom collate function in DataLoader. This technique dynamically pads text https://buff.ly/3RSNvXS

1

📢 #HighlyCited 1. Lemurs Optimizer: A New Metaheuristic Algorithm for Global Optimization https://lnkd.in/gZU-UvM8 2. Investigating Explainability Methods in Recurrent Neural Network Architectures for Financial Time Series Data https://lnkd.in/gq5ssMWb 3. Prediction of Machine Failure in Industry 4.0: A Hybrid CNN-LSTM Framework https://lnkd.in/gYNeDGx8 4. OTSU Multi-Threshold Image Segmentation Based on Improved Particle Swarm Algorithm https://lnkd.in/gK9dRA4k MDPI Encyclopedia MDPI

1

We discuss the limitations of current deep learning models in achieving true Artificial General Intelligence (AGI) and highlight more viable approaches such as hybrid neural-symbolic systems. Learn more: https://lnkd.in/eFhxvKF4 #AI #AGI #DecentralizedAI #NeuralSymbolicAI #DeepLearning #Technology

18

1 comentario

Get ready to optimize your NLP models with vLLM! In his latest blog post, Sergio Morales Esquivel explores the practicalities of streamlining NLP models for real-world efficiency. In it, you'll learn: ✅​ About the vLLM library for efficient model serving ✅​ Practical strategies for conserving computing resources during inference ✅​ The art of batched inference for seamless AI deployments Read on and discover how to make your models perform smarter, not harder: https://lnkd.in/egnUbWsk #Data #DataEngineering #DataScience #ArtificialIntelligence #NLP

11

New Post: Technow: Lightning LLM booster, Anthropic Prompt Library, AI Agent Thunder by Lightning offers a 40% faster training speed for Large Language Models, using operation fusion techniques in PyTorch. Meanwhile, Anthropic' https://buff.ly/3yA5Cep

1

Our DRLearner (https://drlearner.org/) 0.2 release is now available. Importantly, this version delivers on the promise of independence from a single cloud infrastructure (GCP). Whereby, we are now suggesting running an implementation in Docker or AppTainer. This release also includes a bit of code refactoring, and minor bug fixing. We hope that this version is more user-friendly to our fellow researchers. Of note, this version converges on superior performance in Montezuma's Revenge in half the compute time. We attribute this to improvements in the dependent libraries (e.g. JAX). Finally, co-author Phil Tabor intends to put out a video on this topic in the near future. Project: DRLearner is an open-source project to broaden AI access and innovation by distributing AI/Machine Learning (reinforcement learning) code that rivals or exceeds human intelligence across a diverse set of widely acknowledged narrow AI benchmarks. Finally, the code is largely based on “Agent 57”, a curiosity-enhanced Deep Reinforcement Learner by Badia et al (DeepMind). And we are continually grateful for the initial guidance from Adria Badia, of DeepMind. Source Code: https://lnkd.in/eTNiuYwj Project Site: https://drlearner.org/ Machine Learning with Phil: https://lnkd.in/eU8pahCH Docker: https://www.docker.com/ AppTainer: https://apptainer.org/ DeepMind: https://lnkd.in/e--dFWt6

2

On June 20, OpenAlex will host a webinar dedicated to data: ‘Finding, tracking, and analyzing research datasets in OpenAlex’. At the webinar, OpenAlex will cover the work they are doing to index datasets, and Iratxe Puebla from Make Data Count will present on the importance of data citations to understand connections between data and other outputs. Join the webinar on June 20 to hear more about the importance of open infrastructure to evaluate the impact and reach of open data: https://lnkd.in/egTcgu3r #opendata #datametrics

13

🚀 Want to employ the power of #PyTorch for your scientific models without a steep learning curve? 👀 Check out our latest blog, where Travis Oliphant discusses how PyTorch's integration with the Array API makes it easy to use with familiar libraries like NumPy and scikit-learn.🔥 https://lnkd.in/dDRkEY3v #Innovation

10

Let's rewatch LlamaIndex's webinar with Tomaz Bratanic about Advanced RAG with Knowledge Graphs - learn how to use #LlamaIndex property graph! #Neo4j #RAG #llms #GenAI #knowledgegraphs

41

1 comentario