Sebastian Raschka, PhD

Yet another indicator about why LLMs are not databases. They are non-deterministic (in other words, they are stochastic), meaning that even with the same prompts and contexts, different host GPUs can change the output of the LLM in different ways. Even with RAG, replicability is not guaranteed. The underlying message here? LLMs value as generators is unquestionable, and it is a good tool for natural language prompts. However, LLMs likely will never be reliable enough to serve as databases, precisely because they are chaos machines: they are extraordinarily dependent upon starting conditions. https://lnkd.in/gBDei_ye

276

35 Comments

There is often an unspoken assumption by ML researchers that all we need to make progress and model every aspect of our world is data. We do not ascribe to this in my lab. In fact, we believe that data will never be enough. Our experiences with the nuances and complexities of scientific problems have informed us to the insufficiency of data to capture continuous physical processes, which afterall govern our biological and physical world. An example of this is this series of two papers led by my wonderful PhD student, Anowarul Kabir and advanced by a precious multi-year collaboration of my lab with Los Alamos National Lab: Anowarul Kabir, Manish Bhattarai, Kim Rasmussen, Amarda Shehu, Anny Usheva, Alan R Bishop, and Boian S Alexandrov. Examining DNA Breathing with pyDNA-EPBD. Bioinformatics 39(11):btad699, 2023. https://lnkd.in/gZaHE4hS Anowarul Kabir, Manish Bhattarai, Selma Peterson, Yonatan Najman-Licht, Kim Ø Rasmussen, Amarda Shehu, Alan R Bishop, Boian Alexandrov, Anny Usheva. DNA Breathing Integration with Deep learning Foundational Model Advances Genome-wide Binding Prediction of Human Transcription Factors. Nucleic Acids Research: gkae783, 2024. https://lnkd.in/gfTTqcPy Our goal: advance an exceptionally challenging problem in molecular biology, prediction of transcription factor binding sites. Our first step: capture the underlying physics that is missing in the data. Our second step: integrate that now with the data we have in a foundation model for predicting transcription factor binding sites. Performance improves. Most importantly, when we look at the sequence motifs that constitute "signatures" of what makes a transcription factor binding site, we obtain answers. All in the open, nothing opaque.

33

1 Comment

Shashank Sonkar and colleagues are back with another great paper and matching GitHub repo, "Pedagogical Alignment of Large Language Models". You can read the paper on arXiv here - https://lnkd.in/enZfCJgi, and browse the code here - https://lnkd.in/gyW5hSzp. They demonstrate an interesting approach to generating synthetic data to train open weights models to behave more pedagogically, as well as developing some new evaluation metrics. I think there's lots here for those of us interested in this kind of work to learn from and build on. (If synthetic data generation is new do you, check out this post from answer.ai earlier this week - https://lnkd.in/ecsG7R4b.)

38

4 Comments

LLM Self-Refinement in Machine Translation with TEaR Framework! Large Language Models (LLMs) have been making waves in Machine Translation (MT), but human evaluation reveals room for improvement. This research paper introduces TEaR (Translate, Estimate, Refine), a systematic LLM-based self-refinement framework that addresses this challenge! Here's what TEaR brings to the table: * Improved Translation Quality: TEaR helps LLMs refine their translations across various languages, from high-resource to low-resource settings, and across different language pairs (not just English-centric). * Superior Systematicity and Interpretability: TEaR offers a clear and understandable approach to self-refinement, making it easier to analyze and improve. * Impact of Estimation Strategies: they explore how different estimation methods influence the effectiveness of the final corrections. * General-Purpose LLMs for Translation & Evaluation: Unlike traditional models, TEaR leverages the power of general-purpose LLMs to handle both translation and evaluation tasks simultaneously. here's the link to explore #machinelearning #nlp #llms #machinetranslation #research Feel free to share your thoughts and comments below!

33

Looking how to select better representatives from your clusters? Take a look at Lexin Chen's PRIME repo: https://lnkd.in/enM_Btbx Now with updated functionality and tutorials! Great collaboration with Arup Mondal and Alberto Perez! #moleculardynamics #clustering #machinelearning

65

Text embeddings have become crucial for transforming textual data into dense vector representations (with context), enabling efficient processing across various modalities like text, images, and audio. This article provides a good summary on some of these (not all) embeddings available for text https://lnkd.in/gKp4ta9a 1. Gensim Word Embeddings (Free) Why Use It: Specializes in NLP with open-source models like Word2Vec and FastText, promoting transparency and ease of custom training. 2. Facebook Embeddings (Free) Why Use It: Trained on extensive corpora and supports over 100 languages, making it versatile for global applications. 3. AllenNLP Embeddings (Free) Why Use It: Provides NLP-focused embeddings like BERT and ELMo, with tools for fine-tuning and visualization. 4. OpenAI Embeddings (Paid) Why Use It: Comprehensive training on massive datasets allows for effective semantic understanding, making it suitable for advanced NLP tasks. Ideal for tasks requiring zero-shot learning. 5. MultiLingual BERT (Free) Why Use It: Handles multilingual datasets effectively, offering robust support for multiple languages. Perfect for applications requiring consistent performance across different languages, such as global customer support systems. 6. DistilBERT Base Uncased (Free) Why Use It: Smaller and faster than BERT, maintaining near-comparable performance with reduced resource usage. 7. MPNet V2 (Free) Why Use It: Excels in text similarity tasks with a Siamese architecture, providing accurate similarity measurements. 8. HuggingFace Embeddings (Free/Paid) Why Use It: Offers versatility with a wide range of embeddings (text, image, audio) and the ability to fine-tune on custom data. HF also provides wrappers to almost all open sourced models. 9. Longformer Base 4096 (Free) Why Use It: Supports long texts up to 4096 tokens, making it unique in handling lengthy documents without truncation. 10. RoBERTa (2022) (Free) Why Use It: Offers updated text representations with comprehensive training on recent data. 11. GTE-Base (Free) Why Use It: Offers a good balance of performance and resource requirements with a 768-dimensional embedding. 12. GTE-Large (Free) Why Use It: Provides high-quality embeddings with more detailed representations, useful for nuanced text analysis. Best for applications requiring in-depth text understanding, such as detailed content recommendations. 13. GTE-Small (Free) Why Use It: Optimized for speed with a smaller model size, making it efficient for applications needing quick processing. Useful for real-time systems where fast text similarity searches are critical. 14. E5-Small (Free) Why Use It: Combines compact size with good performance, making it suitable for fast, general-purpose text embedding tasks.

11

Can't believe I'd not put this paper on Research Gate or Google Scholar until today! Data Compression and Feature Selection: "Feature Selection for Supervised Learning and Compression" https://lnkd.in/e_USPuea My first paper from this work is still not able to be published! I finished it in 2019.

9

2 Comments

🚀 **Just Published: Unlocking the Power of Large Language Models with Chunking Techniques!** 🚀 🔍 Ever wondered how large language models (LLMs) handle long documents while keeping context intact? The answer lies in **chunking**! In my latest blog, I explore the key chunking methods that boost LLM performance, covering everything from **simple sentence splits** to **adaptive AI-based chunking**. With each technique, I’ve included **code samples** to help you put these methods into action! 💡 Whether you're working in **Generative AI**, **NLP**, or **text summarization**, this guide will help you unlock new potential with LLMs. 📈 **Curious to know how it works? Check out the full blog!** https://lnkd.in/gniDQYwW --- #GenerativeAI #AI #MachineLearning #LLM #NLP #AIResearch #DataScience #Chunking #TextProcessing #AIEngineering --- For maximum engagement, tag a few companies or groups if possible, like: - **@OpenAI** - **@Google AI** - **@Hugging Face** - **@DeepMind**    Excited to share these insights and can’t wait to hear your thoughts! #llm #openai #genai #chunking #langchain

12

2 Comments

Self Alignment using Instruction Backtranslation Recently, it's become common practice to fine-tune LLMs on proprietary datasets or domain-specific data. However, this process requires labeled data – often referred to as ground truth – and getting humans to label this data is both time-consuming and expensive. A recent paper, 'Self Alignment by Instruction Backtranslation',  where researchers have come up with an innovative approach to generate high-quality training data. Check it out here at: https://lnkd.in/dfx5yyha In simple terms, the process starts with a model fine-tuned on a small amount of seed data. The seed model constructs training examples by generating instruction prompts for web corpus. High-quality examples from among these candidates are selected and used to fine-tune a more aligned model. The cool part? When they tried this method with LLaMA, the resulting model outperformed all other LLaMA-based models on the Alpaca leaderboard not relying on data distillation. I've put together a code implementation in my GitHub repo. Feel free to check it out and let me know what you think! #GenAI #LLMs #Llama #Fine-tuning #MachineLearning #NLP #DataScience

47

Another interesting and again very practical article from Mozilla AI About model selections for summarization but really about evaluation A lot of factors in the model evaluation are taken into account I do not a sub eval for hallucination (autoregressive models do it on summarization tasks) Good tips if you do model selection for such tasks as summarization https://lnkd.in/d_yUzmab

5

**Day 9 of 30: Parameter-Efficient Fine-Tuning (PEFT) for LLMs** Training LLMs is computationally intensive. Full fine-tuning requires memory not just for the model but also for optimizer states, gradients, and activations, making it impractical on consumer hardware. 👉 PEFT to the Rescue: Parameter-Efficient Fine-Tuning (PEFT) addresses this by updating only a small subset of model parameters. Here's how: 🏆 Benefits: ♦ Memory Efficiency: By updating fewer parameters, PEFT reduces the memory needed for training. ♦ Faster Training: Quicker training times enable rapid prototyping. ♦ Avoids Overfitting: Maintains the generalization capabilities of the pre-trained model. ♦ Reduced Storage: Each task requires only a small set of new parameters, not a full model copy. 🎊 Popular PEFT Methods: ♦ Selective Fine-Tuning: Updates only specific layers or parameters, balancing parameter efficiency and compute efficiency. ♦ Reparameterization: Techniques like LoRA create low-rank transformations of the model's weights, reducing the number of trainable parameters. ♦ Additive Methods Introduce new trainable components while keeping the original weights frozen. This includes: ♦ Adapter Layers: Add new layers within the model's architecture. ♦ Soft Prompts: Manipulate the input or add trainable parameters to the prompt embeddings. PEFT methods make it possible to train LLMs on a single GPU and avoid the catastrophic forgetting seen in full fine-tuning. 👉 Coming Up: Tomorrow, we'll dive into specific PEFT techniques like LoRA and Soft Prompts to see how they reduce memory requirements and improve training efficiency. How have you used PEFT in your projects? Share your experiences in the comments! #AI #MachineLearning #DataScience #LLMs #PEFT #FineTuning #NLP #TechInnovation #AIInsights

33

1 Comment

I think AI agentic machine translation has huge potential for improving over traditional neural machine translation, and am releasing as open-source a demonstration I'd been playing with as a fun weekend project. Using an agentic workflow, this demonstration (i) Prompts an LLM to translate from one language to another, (ii) Reflects on the translation to come up with constructive suggestions, (iii) Uses the suggestions to refine the translation. In our limited testing, this is sometimes competitive with, and sometimes worse than, leading commercial providers. But it gives a highly steerable translation system where by simply changing the prompt, you can specify the tone (formal/informal), regional variation (do you want Spanish as spoken in Spain or as spoken in Latin America?), and ensure consistent translation of terms (by providing a glossary). This is not mature software. But I hope the open-source community can make agentic translation work much better. Given how a simple reflection workflow already gives decent results, I think there's significant headroom to make agentic translation much better. Releasing an early software prototype like this is something new I decided to try to see if it is helpful to the developer community. I'd love any feedback on this. Thanks to Joaquin Dominguez, Nedelina Teneva, PhD and John Santerre PhD for help with this. https://lnkd.in/gjGANH6H

5,304

163 Comments

Artifacts Claude! Big News: I love the new Artifacts feature in Claude 3.5 that allows you to run code and build things while you figure it out back and forth with the LLM. This has been a big feature of the ChatGPT4 model series, allowing in-line code to run for inference. This takes it to the next level. I made a fully working Tetris game in under 30 seconds. Anthropic - Well done! #ai #claude #newmodel #innovation #llm

24

Very insightful work by Anthropic’s interpretability team. And an amazing paper, with outstanding writing and figures. The idea is very simple: interpret LLMs by leveraging sparse autoencoders as surrogate models of the MLP of transformer blocks, which allow one to disambiguate the superposition of features captured by a single neuron. A simple idea, but a very careful and complex execution, as it is often the case in our line of work. The paper goes into many details and provide a large array of insights, although the gist of the implementation remains obfuscated due to the closed source nature of Claude. Too bad, because this is the kind of work that we need to better understand and eventually trust LLMs. This is demonstrated by the authors in the section ‘Influence on Behavior’, where they show that clamping some features to either high or low value during inference is “remarkably effective at modifying model outputs in specific, interpretable ways”. Hopefully this kind of work is going to be replicated and generalized to open-weights models, such that we have new ways to steer their behavior. https://lnkd.in/eVym7f_f #interpretability #xai #explainableai #steerableai #anthropic #claude #anthropic

3

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

DocETL: Agentic Query Rewriting and Evaluation for Complex Document Processing Shankar et al.: https://lnkd.in/eHsJCY33 #ArtificialIntelligence #DeepLearning #MachineLearning

16