Martin Raison

👉🏼 Chapitre 7. Applications et éclairage éthique des Intelligences Artificielles Génératives linguistiques en Médecine 🤓 Cécile Monteil 👇🏻 https://lnkd.in/e4Fnfwqw 🔍 Focus on data insights: - The integration of Generative Artificial Intelligence systems like Large Language Models (LLM) is reshaping medical practices. - Physicians are required to adapt and acquire new competencies in technology for effective utilization of Gen AI. 💡 Main outcomes and implications: - The evolving ethical framework surrounding Gen AI in medicine necessitates increased ethical vigilance. - Healthcare professionals need to balance technological advancements with preserving their medical expertise and critical thinking skills. 📚 Field significance: - The digitalization of medicine emphasizes the importance of training physicians in new technologies to maintain quality healthcare services. - Gen AI presents both risks and opportunities that require careful consideration within the healthcare sector. 🗄️: [#GenerativeAI #Healthcare #EthicalFramework #DataInsights]

3

1 commentaire

Kolmogorov–Arnold Networks seem like a paradigm shift in ML architectures. Curious to see how they evolve. "While MLPs have fixed activation functions on nodes ("neurons"), KANs have learnable activation functions on edges ("weights"). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. " https://lnkd.in/gdPk_qCQ

29

I couldn't be more proud to write my first official 🤗 blogpost 🎉! I explore the often overlooked yet crucial topic of profiling LLM deployments, specifically with TGI's Benchmarking Tool. It’s a vast area but so vital for understanding the performance nuances of our models, especially for different use-cases. Have you ever encountered surprises or challenges while profiling LLMs? I’d love to hear your experiences and insights! Check out the blog and let’s discuss! 📈 https://lnkd.in/gwK56EXa #LLM #GenAI #TGI #performance

205

7 commentaires

"A team of computer scientists and AI researchers from FAIR at Meta, INRIA, Université Paris Saclay and Google, has developed a possible means for automating data curation for self-supervised pre-training of AI datasets. The group has written a paper describing their development process, the technique they developed and how well it has worked thus far during testing. It is posted on the arXiv preprint server. As developers and users alike have been learning over the past year, the quality of the data that is used to train AI systems is tied very closely to the accuracy of results. Currently, the best results are obtained with systems that use manually curated data and the worst are obtained from systems that are uncurated." #datacuration #aidatasets

2

Embedding models are key for many applications, notably retrieval within RAG systems which are all the hype nowadays ! 🚀 Great initiative by Imene and the MMTEB team to launch embedding leaderboards in languages other than English ! It's pretty fun to see that even by putting minimal efforts in training retrieving models that focus on french (I spent a few hours at most on the sentence-croissants), it's possible to edge out the majority of state-of-the-art proprietary and open-weights multilingual embedding models on these benchmarks. This mainly means that there is still a lot of performance on the table, and it would be great to see both better models, and better evaluation datasets on this leaderboard! The leaderboard: https://lnkd.in/etfHhWBU The current version of the sentence-croissant model, optimized for French and English retrieval : https://lnkd.in/extU7QjD

31

How fast is fast when it comes to commercial open source software traction? Here are some stats to serve as benchmarks that Matthieu Lavergne put together among many other great data and insights in Serena's first report on commercial open source software. It's available for free here: https://lnkd.in/d5uBvjAU

18

3 commentaires

I am pleased to announce the release of 𝐌𝐢𝐬𝐭𝐫𝐚𝐥-𝐑𝐀𝐆 . This model is a refined fine-tuning of the 𝐌𝐢𝐬𝐭𝐫𝐚𝐥-𝐈𝐭𝐚-7𝐛 , specifically engineered to enhance 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐚𝐧𝐬𝐰𝐞𝐫 𝐭𝐚𝐬𝐤𝐬. What sets the Mistral-RAG apart is its innovative 𝐝𝐮𝐚𝐥-𝐫𝐞𝐬𝐩𝐨𝐧𝐬𝐞 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐲, offering both 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 and 𝐞𝐱𝐭𝐫𝐚𝐜𝐭𝐢𝐯𝐞 𝐦𝐨𝐝𝐞𝐬. - 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐦𝐨𝐝𝐞: is particularly valuable when complex, synthesized responses are needed, which involve integrating information from multiple sources or providing expanded explanations 📖. This makes it ideal for educational, advisory, and creative scenarios where depth and detailed understanding are crucial 🔍. - 𝐄𝐱𝐭𝐫𝐚𝐜𝐭𝐢𝐯𝐞 𝐦𝐨𝐝𝐞: excels in situations where speed and precision are paramount, delivering direct and concise answers by pulling specific data from texts 🏁. This mode is indispensable for factual queries in research, legal, and professional contexts where accuracy and direct evidence are necessary ✅.

150

7 commentaires

!! 𝐖𝐞 𝐚𝐫𝐞 𝐇𝐢𝐫𝐢𝐧𝐠 𝐟𝐨𝐫 𝐀𝐈/𝐌𝐋 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐒𝐜𝐢𝐞𝐧𝐭𝐢𝐬𝐭𝐬 !! Dear AI/ML Researchers ! At Eluvio we are on a mission to build world's first decentralized platform serving a global share of internet videos. In this journey we have been lucky to tackle many complex research and engineering spanning multi-field multimodal learning to Generative AI with Large / (Compressed) Multimodal Models. Our ML research is a part of our Casablanca Release that won the **𝐍𝐀𝐁 𝟐𝟎𝟐𝟒 𝐏𝐫𝐨𝐝𝐮𝐜𝐭 𝐨𝐟 𝐭𝐡𝐞 𝐘𝐞𝐚𝐫 𝐀𝐰𝐚𝐫𝐝 **  If you love working on advanced ML research with unique multimodal datasets; join our ML Research team ! Your research will have direct impact on existing products and published to top ML/AI conferences! Please apply to the link below:- https://lnkd.in/gyanngZv #GENAI  #hiring  #machine_learning  #Eluvio  #multimodal

135

8 commentaires

I’m excited to learn LLMops and MLops from Nicolas Thiebaut in the upcoming module!

1

"Refining Large Language Models: Analyzing Content Alignment with Cosine Similarity and ROUGE Scores" When it comes to Large Language Models, it is crucial that the generated content closely resembles the expected output to ensure the reliability and utility of these models. In my current project centered around LLMs, I employ two key metrics, cosine similarity and ROUGE scores, to precisely measure and validate the similarity between generated responses and the expected answers. Cosine similarity is a mathematical measure that evaluates the similarity between two vectors by computing the cosine of the angle between them. In our context, these vectors represent the textual representations of generated responses and the corresponding expected answers. A cosine similarity score close to 1 indicates a high degree of alignment between the two vectors, suggesting that the answers share similar directional trends and content. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) scores offer a comprehensive evaluation of summary quality by measuring the overlap in n-grams between generated responses and reference answers. These scores encompass a range of metrics, including ROUGE-N (measuring overlap of n-grams), ROUGE-L (measuring the longest common subsequence), and ROUGE-W (incorporating weighted n-gram overlap). By analyzing different aspects of content overlap, ROUGE scores provide nuanced insights into the precision and recall of the generated content, highlighting areas of content relevance, coherence, and structural similarity. The significance of these metrics cannot be overstated. They serve as critical benchmarks for quality assurance, ensuring that the generated responses faithfully reflect the expected content. By quantitatively assessing the efficacy of language models in producing coherent and contextually relevant responses, cosine similarity and ROUGE scores facilitate informed decision-making and iterative model refinement efforts. In summary, the utilization of cosine similarity and ROUGE scores is indispensable in the development and evaluation of Large Language Models. These analytical tools not only validate the similarity between generated and expected content but also drive continual improvement in model performance, ultimately enhancing the utility and effectiveness of language models in various applications. #LLM #AI #GenerativeAI

4

[NEW PAPER!!!] I am thrilled to announce that our latest research paper, "Evaluating the Adversarial Robustness of Semantic Segmentation: Trying Harder Pays Off," has been accepted for presentation at ECCV 2024! 🎉 In this study, we investigate the vulnerability of semantic segmentation models to adversarial attacks, highlighting the crucial importance of robust segmentation in ensuring model reliability. We developed an evaluation tool that generates adversarial examples with unprecedented efficiency. Our results reveal that many previous solutions are not as robust as previously thought. Even for more effective defenses, we managed to lower the known lower bounds of robust accuracy. We created new metrics and a comprehensive methodology for evaluation. Excitingly, we will also be releasing the evaluation tool we developed during this research very soon, allowing others to access and utilize the new methods and metrics we’ve introduced. You can read more about our findings here: https://lnkd.in/gmqcvr8M This acceptance for ECCV marks a significant milestone in my career, and I am immensely grateful to everyone who supported and collaborated on this work, especially Bálint Mohos and Márk Jelasity. I look forward to presenting our findings and discussing them with the esteemed community at ECCV! #research #science #publication #ECCV2024 #arXiv #career

22

🚀 Some great bits of advice from the leaders at the VivaTech conference, Paris Meta AI chief Yann LeCun, advised students looking to work in the AI ecosystem, to not work on LLMs- “If you are a student interested in building the next generation of AI systems, don’t work on LLMs. This is in the hands of large companies, there’s nothing you can bring to the table” Some more key points to take away from the conference:- 1️⃣ LeCun asked people to focus on developing next-generation AI systems that overcome the limitations of large language models. 2️⃣ Muffed Vh the creator of Devika told people should move away from Transformer models and start researching on new architectures. He believes that with this approach, it is even possible to build something nearly as impressive as GPT-4 3️⃣ They also spoke about the growth of AI assistants and how all our interactions with the digital world will be mediated by AI assistants in the coming future. A great article on what we should focus on building next, to create an impact. Link - https://lnkd.in/dfb7g4PF #generativeai #ai #llm #openai #machinelearning #meta #analyticsindia

10

AdvPrompter : Fast Adaptive Adversarial Prompting Approach for LLMs Red-teaming test In a recent paper, authors introduced a novel automated red-teaming method which uses another LLM called AdvPrompter, that rapidly generates diverse human readable adversarial prompts in seconds, ∼800× faster than existing optimization-based approaches. 𝗦𝘂𝗺𝗺𝗮𝗿𝘆 𝗼𝗳 𝗔𝗱𝘃𝗣𝗿𝗼𝗺𝗽𝘁𝗲𝗿 i) Core idea - train another LLM, called AdvPrompter, that aims to generate adversarial suffixes against a TargetLLM, given user instructions. ii) Training method overview Training method called as “AdvPrompterTrain” alternates between - AdvPrompterOpt, a novel efficient optimization algorithm that iteratively generates adversarial suffixes that jailbreak the TargetLLM and maintain human readability, measured by the perplexity of a BaseLLM, and  - supervised fine-tuning of the AdvPrompter with the generated adversarial suffixes as targets. iii) Key advantages - Human-readability : generates coherent human-readable adversarial prompts that mimic human-written adversarial prompts - High attack success rate: generates attacks with higher ASR(attack success rate) and lower perplexity than competing methods. - Adaptivity to input: suffixes generated by AdvPrompter are conditioned on the instruction, even when generalizing to previously unseen test instructions as compared to single universal nonadaptive prefix in other methods. - Fast generation: generates adversarial suffixes simply through next-token prediction, whereas previous methods such as GCG and AutoDAN require solving an entirely new optimization problem for every generated suffix. - No gradient from TargetLLM :  does not use any back-propagated gradient information from the TargetLLM, but only its log probability output 𝗠𝗲𝘁𝗵𝗼𝗱𝗼𝗹𝗼𝗴𝘆 i) AdvPrompter: Predicting Adversarial Prompts -  given a set of harmful instruction-response pairs, training the AdvPrompter with idea of  finding a universal adversarial suffix to a conditional setup. ii) AdvPrompterTrain: Training AdvPrompter via Alternating Optimization - computing “target” adversarial suffixes that can be used as regression targets for the AdvPrompter. iii) AdvPrompterOpt: Generating Adversarial Targets - does not require backpropagated gradient through TargetLLM. - selects best token in greedy manner followed by by stochastic beam-search 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻𝘀 i) Attacking Whitebox TargetLLM - achieves higher overall ASR(Attack success rate) with low perplexity score as compared to AutoDAN and GCG baselines. ii) Transfer-attacking Blackbox TargetLLM -  find a universal adversarial suffix on a whitebox TargetLLM,with ASR evaluation of the resulting adversarial prompts on blackbox TargetLLM. - higher ASR achieved as compared to baseline 🏆 For detailed analysis of this paper, please refer to my blog on it at: https://lnkd.in/dYjy5b9i #largelanguagemodels

23

📢 Our latest blog post in partnership Argilla is out. We cover SimPO a simplified and faster 🏎️ variant of DPO for aligning LLMs to your use cases.

2

The much anticipated open-source partner package huggingface-langchain🚀 is here. It provides a way to seamlessly integrate the latest open Models from Hugging Face into LangChain, supporting both local models and hosted models! 🤗🦜 🛠️ Easy Installation: Install with a simple `pip install langchain-huggingface`. 🚪 Open Models: Easily access open LLMs and Embedding models. 🌐 Flexibility: Utilize Hugging Face models via API, Inference Endpoints, or self-hosted Text Generation Inference. 💬 Chat Models: Support conversational models using ChatHuggingFace. 🧠 Embeddings: Leverage sentence-transformers embeddings locally or via Hugging Face endpoints. 🏎️ Fast Integration: Models can be loaded directly using the from_model_id method or by defining the pipeline manually. Get started here: https://lnkd.in/dmj6xmpU

9

I was delighted to read BABILong which had a much better alternative to the all too common "Needle in a Haystack" approach of evaluating context length for LLMs. 🧠 BABILong includes a diverse set of 20 reasoning tasks, that were designed as prerequisites for any system that aims to be capable of conversing with a human 🔗 The method can be extended to evaluate any length 🛡️The facts can be generated to make it leak proof. My only criticism is that I do think that some of these facts might be out of domain and not quite how we pre-train/instruction tune making it harder than necessary for LLMs. But this is a HUGE improvement over its common predecessor. https://lnkd.in/ddVE5te7

13

1 commentaire

📢 Hugging face is directly integrated with GenAI developers favorite tool "langchain"🦜 Get set go!🚀 💡 pip install langchain-huggingface https://lnkd.in/ddHreVrs #huggingface #GenerativeAI #langchain

22