Michael Benayoun

MisguidedAttention : a collection of prompts to challenge the reasoning abilities of large language models. In this github repo, authors introduced  prompts with slight variations of commonly known thought experiments or paradoxes ("trick questions"), which can challenge the reasoning abilities of LLMs. 𝗘𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹 𝗢𝗯𝘀𝗲𝗿𝘃𝗮𝘁𝗶𝗼𝗻𝘀 𝘂𝘀𝗶𝗻𝗴 𝘁𝗵𝗲𝘀𝗲 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 𝗳𝗼𝗿 𝗟𝗟𝗠 𝘁𝗲𝘀𝘁𝗶𝗻𝗴 - Expected behavior would be that the LLMs solving the problems with logical deduction, but many LLMs will mistakenly recognize the unmodified problem due to frequent occurrence in their training data. - When LLMs not working, they responded with a solution to the unmodified problem instead of going through the details step-by-step to find a solution for the modified problem. -  it's also possible to observe intertwined strings of reasoning where conflicting thoughts are alternating in the same text. -  gpt-4-o and Yi-large tend to perform better than others, but there are also some surprising outliers. 𝗣𝗿𝗼𝗯𝗹𝗲𝗺𝘀 𝘁𝗵𝗮𝘁 𝗳𝗼𝗿𝗺𝗲𝗱 𝗯𝗮𝘀𝗶𝘀 𝗼𝗳 𝗺𝗼𝗱𝗶𝗳𝗶𝗲𝗱 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 - Trolley problem: https://lnkd.in/euQHgvVg - Monty Hall problem: https://lnkd.in/e9cC9wdb - Barber paradox: https://lnkd.in/edHhQxPw - Schrödingers cat: https://lnkd.in/e_vYW9CY - Unexpected hanging paradox: https://lnkd.in/eZ642wmZ - River crossing puzzle: https://lnkd.in/eZMXJryn - Two doors problem, apparently a variant of Knights and Knaves https://lnkd.in/enQQKpcr from this movie 𝗣𝗿𝗼𝗺𝗽𝘁𝘀 𝗿𝗲𝗽𝗼: https://lnkd.in/eGP9Shes #generativeai

23

🚀 Exciting News! Our team’s report, “A Partial Replication of MaskFormer in TensorFlow on TPUs for the TensorFlow Model Garden,” is now available on arXiv! We’ve tackled the challenging task of replicating the MaskFormer model, originally developed in PyTorch, within the TensorFlow ecosystem, specifically optimized for TPUs. Check out our findings and the full paper here:

18

Hey, Data Scientists and Test Engineers, we would like to introduce you to our article: Why It’s Time to Analyze Your Acoustic Test Data with Machine Learning — Now! 🚀 You can find it at medium: https://lnkd.in/eChVvUV9 The reasons for this are: - Maturity of ML Algorithms: Advanced algorithms are now capable of processing audio data more efficiently than ever before. - Accessible Expertise: The technology has become more accessible. This allows a wider range of test engineers to use ML tools. For instance, suitable pre-trained models and test data can be easily found on Huggingface. - Models with Low Data Requirements: Foundation models require only a few data points to provide valuable insights 💡. This includes, for example, the Audio Spectrogram Transformer. This can be used without training to quickly and easily generate audio embeddings, which can be used to gain initial insights into the data distribution using 2D UMAP. The animation below shows this using the emdbd dataset as an example. The article describes common challenges and how to overcome them, and also shows which system architecture is suitable to start with your own system. References: Yuan Gong, Yu-An Chung, James Glass: AST: Audio Spectrogram Transformer (2021), arXiv Felix Burkhardt, Astrid Paeschke, Miriam Rolfes, Walter Sendlmeier und Benjamin Weiss: A Database of German Emotional Speech (2005), Proceedings Interspeech 2005, Lissabon, Portugal --- Discover how AI is transforming test data analysis in automotive, aerospace, and mechanical engineering at Renumics Test Data Analysis: https://lnkd.in/ePfTyrRh

92

3 commentaires

Wouldn’t it be great if we could align LLMs at inference time to a custom set of guidelines while maintaining the helpfulness of the undefended model? 🤔 Meet 𝗣𝗿𝗶𝗺𝗲𝗚𝘂𝗮𝗿𝗱 🤺: the new, most performant method to do just that at a marginal additional cost 💲 Together with my awesome co-authors, Eliott Zemour and Eric Lin we had the pleasure of presenting our work at the 𝘐𝘊𝘔𝘓 2024 𝘕𝘦𝘹𝘵𝘎𝘦𝘯𝘈𝘐𝘚𝘢𝘧𝘦𝘵𝘺 𝘸𝘰𝘳𝘬𝘴𝘩𝘰𝘱 workshop, organized by OpenAI and Google DeepMind in Vienna 🇦🇹, proudly representing Dynamo AI 🫶🏼 Here’s some background and a method overview (links below)👇: 𝗠𝗼𝘁𝗶𝘃𝗮𝘁𝗶𝗼𝗻: We observed that models equipped with current inference-time safety interventions suffer from a 𝘀𝗶𝗴𝗻𝗶𝗳𝗶𝗰𝗮𝗻𝘁 𝗱𝗿𝗼𝗽 𝗶𝗻 𝘂𝘀𝗲𝗳𝘂𝗹𝗻𝗲𝘀𝘀 on safe queries, a phenomenon we refer to as a "Guardrailing tax." PrimeGuard provides you with a tax exemption! 💰 𝗦𝗲𝘁𝘂𝗽: Apply PrimeGuard on top of models of different sizes to ensure its responses abide by a significantly more stringent safety policy than the built-in safety mechanisms.  𝗠𝗲𝘁𝗵𝗼𝗱 • 🔍 𝗥𝗼𝘂𝘁𝗲𝗿 𝗠𝗼𝗱𝗲𝗹: Analyzes the user query and policy to identify conflicts, instructing the Main Model to refuse if there's a direct conflict, provide an optimal answer if there's no conflict, and generate additional guidance if there's a potential conflict.  • 🛠️ The router can be trained, but as we show in the paper, performant 𝗿𝗼𝘂𝘁𝗶𝗻𝗴 𝗰𝗮𝗻 𝗯𝗲 𝗶𝗻𝗱𝘂𝗰𝗲𝗱 𝘄𝗶𝘁𝗵 𝗮𝘀 𝗳𝗲𝘄 𝗮𝘀 𝟲 𝗳𝗲𝘄-𝘀𝗵𝗼𝘁 𝗲𝘅𝗮𝗺𝗽𝗹𝗲𝘀. 🚀 𝗖𝗼𝗼𝗹 𝗥𝗲𝘀𝘂𝗹𝘁𝘀 𝗳𝗼𝗿 𝗠𝗶𝘅𝘁𝗿𝗮𝗹 𝟴𝘅𝟮𝟮𝗕, a frequently self-hosted model (see more in the paper): • 📉 Reduces the unsafe response rate by a factor of 13x. - 📈 Increases the helpfulness score over the undefended model.  • 🛡️ First method with a noticeable decrease in attack success rate against a SOTA interactive attack TAP, reducing the success rate from 100% to 8%.  • 🟢 Does not increase the refusal rate to innocuous samples.  🔍 𝗙𝘂𝗻 𝗳𝗮𝗰𝘁: Mistral 7B with the Mistral 7B PrimeGuard Router Model achieves similar safety performance to the extensively aligned Llama 3 8B. Thanks for making it this far! Please check out our paper: https://lnkd.in/eBhkxwCR Code: https://lnkd.in/eNcY8YhG For a limited time, we are also hosting a demo of PrimeGuard where you can experience the magic with your own queries 🪄: https://lnkd.in/extyAxbr If this research sounds interesting to you, my team is hiring: https://lnkd.in/epJzUX-c

132

5 commentaires

It's crazy that we dwarf models to achieve good metrics on benchmarks. 🤦‍♂️ While developing Docmatix, we got a great model from fine-tuning Florence-2 on it. But it didn't do well with the benchmarks. For that, we had to further fine-tune on DocVQA to force the model to output very short, machine-like answers. The resulting model sucks from a human perspective but does great in the benchmark 😩 The amazing Dana Aubakirova saw this and decided that we needed to find a better solution! In comes LAVE, a novel metric that uses an LLM to judge whether the generated answer matches the ground truth. We tested it using Llama-2-Chat-7b as a judge and went from 0% ANSL to 58% Lave score on the model trained only on Docmatix. Showing that the problem was the metric! Check out the full analysis in the blog: https://lnkd.in/d_8uvkw7

47

2 commentaires

⭐️ New video release 📺: Performant, scientific computation in 🐍Python and 🦀 Rust Watch Stefan Ulbrich as he guides you through building performant scientific computation packages in Python and Rust, offering insights into efficient computation and optimization for machine learning and data science applications. 📺 Watch the video on YouTube: https://lnkd.in/eq-BhHdH Stefan Ulbrich, Senior Data Scientist, offers a tutorial on enhancing scientific computation in Python and Rust. The session explores packaging numerical calculus and algorithms using modern tools. Highlighting the efficiency of vectorization in Python for computational tasks within Machine Learning, Ulbrich presents the integration of Rust to optimize package performance. Rust, recognized for its performance and memory safety, complements Python, particularly for instances prioritizing speed and modern language features. The talk demonstrates creating a scientific Python package, gradually transitioning components to Rust for enhanced performance. Through an example of implementing the "Expectation maximization for Gaussian Mixture Models" algorithm, attendees learn how to leverage modern tools to build, optimize, and extend packages efficiently. Ulbrich emphasizes that while Rust knowledge is not mandatory, the session provides a high-level understanding of its benefits for Python extension and performance improvement. For a hands-on experience, participants are encouraged to clone Ulbrich's repository for swift installation and to optimize session time effectively.

9

1 commentaire

📣 New paper announcement with some new and old collaborators : "Magnituder Layers for Implicit Neural Representations in 3D" https://lnkd.in/eHkNgaG4 In this work, we design new exotic neural network layers building on my earlier work from this year "Scalable Neural Network Kernels". As the name suggests, these layers only depend on the magnitude of the inputs. 🤔 You may ask that this layers may be quite useless as it throws away so much information? - We swap usual MLP layers with these layers in various Implicit Neural Representations (INR) (NeRF and iSDF) and show we can achieve similar quality as the baseline with a reduction of training parameters and an improvement of inference speed. - We show that this layer approximates certain MLPs quite well and we can write down a closed form formula for the optimal layer weights. So we can simply distill the baselines with our efficient variants without any backpropagation. This is my first foray in INR and I am excited what we can achieve with these kernel methods where MLPs dominate most of the computation.

18

2 commentaires

LONGEMBED : Benchmarking Extension of Embedding Models for Long-Context Retrieval In this paper, authors introduced a novel LONGEMBED benchmark comprising of two synthetic tasks and four carefully chosen real-world tasks, featuring documents of varying length and dispersed target information. 𝐍𝐨𝐭𝐚𝐛𝐥𝐞 𝐨𝐮𝐭𝐜𝐨𝐦𝐞𝐬 𝐨𝐟 𝐭𝐡𝐢𝐬 𝐩𝐚𝐩𝐞𝐫 𝐚𝐫𝐞: - observing effective extension of the context window of existing embedding models by several folds, regardless of their original context being 512 or beyond 4k. - For models employing absolute position encoding (APE), showing the possibility of harvesting further improvements via fine-tuning while strictly preserving original behavior within the short context. - For models utilizing Rotary Position embeddings (RoPE) substantial enhancements on LONGEMBED are observed when employing methods that fully leverage RoPE’s advantages, such as NTK and SelfExtend. 𝐋𝐢𝐦𝐢𝐭𝐚𝐭𝐢𝐨𝐧𝐬 𝐨𝐟 𝐄𝐱𝐢𝐬𝐭𝐢𝐧𝐠 𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐬 - candidate documents not long enough -  target information to answer user query not uniformly distributed across the document as possible. 𝐓𝐚𝐬𝐤𝐬 𝐢𝐧 𝐋𝐨𝐧𝐠𝐄𝐦𝐛𝐞𝐝 - Synthetic tasks like Personalized Passkey Retrieval and Needle-in-a-haystack Retrieval - Real tasks for evaluations on long-form QA and summarization tasks across datasets like NarrativeQA , 2WikiMultihopQA , QMSum and SummScreenFD 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐖𝐢𝐧𝐝𝐨𝐰 𝐄𝐱𝐭𝐞𝐧𝐬𝐢𝐨𝐧 𝐟𝐨𝐫 𝐀𝐏𝐄-𝐛𝐚𝐬𝐞𝐝 𝐌𝐨𝐝𝐞𝐥𝐬 Strategies used: - 𝐏𝐚𝐫𝐚𝐥𝐥𝐞𝐥 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐖𝐢𝐧𝐝𝐨𝐰𝐬 (𝐏𝐂𝐖) :  PCW divides the long document into multiple short chunks, processes each chunk in parallel, and aggregates their results - 𝐆𝐫𝐨𝐮𝐩𝐞𝐝 𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝐬 (𝐆𝐏) & 𝐑𝐞𝐜𝐮𝐫𝐫𝐞𝐧𝐭 𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝐬 (𝐑𝐏) : position reorganization accommodates longer context by reusing the original position ids. - 𝐋𝐢𝐧𝐞𝐚𝐫 𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧 𝐈𝐧𝐭𝐞𝐫𝐩𝐨𝐥𝐚𝐭𝐢𝐨𝐧 (𝐏𝐈) : introduces position embeddings via linear interpolation of existing ones. 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐖𝐢𝐧𝐝𝐨𝐰 𝐄𝐱𝐭𝐞𝐧𝐬𝐢𝐨𝐧 𝐟𝐨𝐫 𝐑𝐨𝐏𝐄-𝐛𝐚𝐬𝐞𝐝 𝐌𝐨𝐝𝐞𝐥𝐬 Strategies used: - 𝐒𝐞𝐥𝐟 𝐄𝐱𝐭𝐞𝐧𝐝 (𝐒𝐄): re-introduces normal relative positions within the nearest neighbor window. - 𝐍𝐓𝐊-𝐀𝐰𝐚𝐫𝐞 𝐈𝐧𝐭𝐞𝐫𝐩𝐨𝐥𝐚𝐭𝐢𝐨𝐧 (𝐍𝐓𝐊): scales high frequencies less and low frequencies more to spread out the interpolation pressure across multiple dimensions 𝐑𝐞𝐬𝐮𝐥𝐭𝐬 - APE based models: Plug and play methods like GP, RP, LPI and PCW strategies yield comparable results with no significant disparities. - RoPE-Based Models: RoPE-specific methods including NTK and SE yield significant improvements for both models across all datasets, surpassing PCW, PI and GP by a large margin. 🏆 For my detailed analysis of this paper, please refer to my blogpost on it at: https://lnkd.in/dRMbdH39 #generativeai #largelanguagemodels

15

🚀 Breaking New Ground in Neural Networks – Part 2 is Here! 🚀 Following up on my last article, I'm thrilled to present the next phase in the journey of Dynamic Neural Networks! In this new post, we dive deep into how Spectral Clustering with Nonlinear Kernel Embedding and Functional Data Analysis (FDA) can supercharge neural network architecture, allowing networks to evolve and adapt their structure dynamically during training. 🔥 Why does this matter? - Dynamically adjust connections to handle complex, nonlinear data effortlessly. - Enhance model efficiency by pruning unnecessary connections, reducing overfitting. - Unlock new potential in industries like bioinformatics, natural language processing, and healthcare diagnostics. This is a continuation of my exploration into how neural networks can adapt, evolve, and redefine problem-solving. If you’re passionate about the future of AI and deep learning, this is a must-read! Check out Part 2 for all the exciting details and practical code applications🔗 The future of adaptable, intelligent networks is here! 💡 #NeuralNetworks #MachineLearning #DataScience #Innovation #DeepLearning #Bioinformatics #NLP #HealthcareTech

3

3 commentaires

Thinking Tokens for Language Modeling : Enhancing LLMs generalization capabilities to perform calculations for complex problems Language models often make mistakes in doing calculation for simplest of problems like "How much is 56 times 37?". Since language models rely on large training sets and great memorization capability, naturally they are not equipped to run complex calculations. Authors of this paper propose to use special ’thinking tokens’ which allow the model to perform much more calculations whenever a complex problem is encountered. 𝗞𝗲𝘆 𝗜𝗱𝗲𝗮 - reasoning abilities could be improved by employing a large amount of supervision by providing labeled examples to the model , but authors aimed to find much faster unsupervised approach. - Also uses the premise that humans also cannot perform complex calculations right away and require a considerable amount of time to provide a solution. - To enhance the generalization capability of language models, authors proposed to use special ’thinking tokens’ which allow the model to perform much more calculations whenever a complex problem is encountered. 𝗧𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝘁𝗼𝗸𝗲𝗻𝘀 𝗳𝗼𝗿 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 - approach is to introduce special ’thinking tokens’ (< T >) after each word in a sentence whenever a complex problem is encountered. - core idea is that each ’thinking token’ would buy more time for the model before an answer is expected, which would be used to run additional computations to better answer a complex problem that was presented. 𝗘𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹 𝗥𝗲𝘀𝘂𝗹𝘁𝘀 - Experiments performed on sampled sentences from Maths dataset - Preliminary results show that sentences that require non-trivial reasoning, have the biggest improvement in perplexity when ’thinking tokens’ are used compared to the standard model - authors also observed that introduction of ’thinking tokens’ is also successful for sentences that include specific numbers or representative symbols of numerical values. 𝗣𝗮𝗽𝗲𝗿 𝗹𝗶𝗻𝗸: https://lnkd.in/ecU_YDDX #generativeai

87

9 commentaires

𝗪𝗵𝗮𝘁 𝗣𝗿𝗼𝗺𝗽𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗧𝗲𝗮𝗰𝗵𝗲𝘀 𝗨𝘀 𝗔𝗯𝗼𝘂𝘁 𝗛𝗼𝘄 𝗪𝗲 𝗧𝗵𝗶𝗻𝗸 In the rapidly evolving world of artificial intelligence, prompt engineering has become a crucial practice, shaping our interactions with language models. Coined in 2021, prompt engineering is more nuanced than many realize, as revealed in the recent paper 𝘛𝘩𝘦 𝘗𝘳𝘰𝘮𝘱𝘵 𝘙𝘦𝘱𝘰𝘳𝘵: 𝘈 𝘚𝘺𝘴𝘵𝘦𝘮𝘢𝘵𝘪𝘤 𝘚𝘶𝘳𝘷𝘦𝘺 𝘰𝘧 𝘗𝘳𝘰𝘮𝘱𝘵𝘪𝘯𝘨 𝘛𝘦𝘤𝘩𝘯𝘪𝘲𝘶𝘦𝘴, published in July 2024 (https://lnkd.in/edNTP46X) This paper meticulously surveys and categorizes 58 text-only prompting techniques and 40 techniques for other modalities, offering a comprehensive taxonomy that highlights the complexity of crafting effective prompts. The survey uncovers some surprising findings. For instance, few-shot learning with exemplars can significantly improve outcomes, but results vary widely based on factors like the number, order, labeling, format, and even the similarity of the exemplars to the task. Another intriguing discovery is that simply asking a model to “read the question again” can enhance performance on complex tasks. The survey emphasizes the sensitivity of model completions to seemingly minor details. Even an extra space can degrade results, underscoring the importance of experimentation in prompt formulation. Innovative techniques like testing for consistency—by having the model reconstruct a prompt from its own completion and comparing the two—reveal inconsistencies that can help refine the original prompt. The survey’s methodology is as rigorous as its findings are insightful. The authors developed a data-scraping pipeline to identify 1,565 relevant papers from arXiv, ACL, and Semantic Scholar, which were then classified and mined by the authors. Beyond technical insights, the paper raises an intriguing question: Are some of these prompting techniques valuable in human-to-human interactions as well? If so, what does this reveal about our cognitive processes and communication methods?

46

6 commentaires

I wrote two blog posts about SN, Leon Bottou and Yann LeCun's 1988 𝘚𝘪𝘮𝘶𝘭𝘢𝘵𝘦𝘶𝘳 𝘥𝘦 𝘕𝘦𝘶𝘳𝘰𝘯𝘦𝘴. One is an English translation of the original paper, for which I've reproduced the figures. The other is a tutorial on how to run their code on Apple silicon. https://lnkd.in/dMFxMZeT

77

2 commentaires

OpenAI has just released the Multilingual Massive Multitask Language Understanding (MMMLU) dataset on Hugging Face, providing a valuable resource for evaluating multilingual large language models (LLMs). The MMLU test set is available in 14 languages, including Arabic, German, Spanish, and French. It spans 57 categories, ranging from elementary to advanced professional subjects, and has been translated by professional human translators. This dataset allows for the evaluation of AI models' general knowledge across diverse cultures and is used in the openai/simple-evals project. However, the licensing details remain unclear at this time. Explore the dataset: https://lnkd.in/er9iPrNA Kudos to OpenAI for this significant contribution! #OpenAI #MultilingualAI #LLM #NaturalLanguageProcessing #MMLU #AIResearch #HuggingFace #DatasetRelease #LanguageModels #AITranslation #DeepLearning #AIInnovation #GlobalAI #MachineLearning #AIEvaluation Umar Iftikhar

3

📣 Happy to announce a new work with some amazing collaborators "Fast Tree-Field Integrators (FTFI) : From Low Displacement Rank to Topological Transformers". In this work, we introduce new fast (polylog) algorithms to compute the action of the Shortest Path Kernel on a Tensor Field on a graph. Other than new theoretical results on graph theory (our method is numerically exact on trees), this method can be applied to (Linear) ViTs to improve their classification accuracy. This work is an extension of our ICML 2023 paper where the graphs in question were meshes. Paper : https://lnkd.in/gpSHWYGj

29

4 commentaires

"LLM Compression" - Analogous to "Image Compression". AKA Quantization. Lots of resources in the comments on this post! #llm #genai #machinelearning #startup

5

One particularly powerful thing about open-source LLMs with Ollama and Timescale's pgai for #Postgres is that you can now host your own models without needing to pay or share private data with closed-sourced providers like OpenAI.

24