Yann Dubois

Reflections on Sparse Representations in Language Models Reading about the use of sparse autoencoders in language models revealed an intriguing balance between interpretability and efficiency. Sparse representations enhance feature disentanglement, aiding in understanding model behaviors, especially in AI safety and bias detection. However, for practical applications where interpretability is secondary, sparse representations might introduce computational redundancy and inefficiency. Thus, while sparse autoencoders offer valuable insights for research and safety, more compact representations could be preferable for deployment. Balancing these aspects is crucial, potentially through adaptive approaches that optimize for both interpretability during research and efficiency in real-world applications. https://lnkd.in/egPP7dEg

1

These findings motivate our design of a novel framework to investigate hidden representations in LLMs with LLMs, which we call Patchscopes. The key idea behind this framework is to use LLMs to provide natural language explanations of their own internal hidden representations. Patchscopes unifies and extends a broad range of existing interpretability techniques, and it enables answering questions that were difficult or impossible before. For example, it offers insights into how an LLM's hidden representations capture nuances of meaning in the model's input, making it easier to fix certain types of reasoning errors. While we initially focus the application of Patchscopes to the natural language domain and the autoregressive Transformer model family, its potential applications are broader. For example, we are excited about its applications to detection and correction of model hallucinations, the exploration of multimodal (image and text) representations, and the investigation of how models build their predictions in more complex scenarios. Google DeepMind

9

Sparse Autoencoders (SAEs) are one of the most promising and popular methods for LLM interpretability and explainability. I wrote a simple, intuitive introduction to SAEs, complete with diagrams and reference PyTorch code. Among other things, I cover why we use SAEs, how they are used for interventions, and challenges with using SAEs. https://lnkd.in/gzrvi6Kg

15

Unlocking Big Potential: How Repeated Sampling Empowers Small Models Reading about repeated sampling in LLMs was eye-opening. The concept that multiple attempts can elevate a smaller model's performance, rivaling larger, more expensive models, is fascinating. This method doesn't just improve accuracy—it democratizes access to powerful AI by making high-level performance achievable with less computational cost. The analogy to multi-agent systems further solidifies the idea: multiple perspectives or attempts converge to enhance overall results. For anyone interested in efficient AI, this approach is a game-changer, offering a practical path to optimizing smaller models and getting more out of limited resources. https://lnkd.in/eZ7Wmd5r

2

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

5,305

163 Comments

🌍 Thrilled to have virtually presented our paper, "Information Parity: Measuring and Predicting the Multilingual Capabilities of Language Models," at EMNLP 2024 last week! This work builds upon and extends our earlier research on Compression Parity (CP), which was presented at the ICML 2024 Theoretical Foundations of Foundation Models workshop. While the metric itself remains fundamentally the same—now renamed as Information Parity (IP)—this new paper introduces significant advancements in analysis and evaluation. What’s new in this iteration? We conduct extensive statistical analysis to validate IP and demonstrate its superiority over alternative metrics, such as tokenization-based measures (e.g., Tokenization Parity and Tokenizer Fertility) and the proportion of languages in training data. Notably, we show that: • IP consistently outperforms these metrics in predicting multilingual model benchmarks performance, such as multilingual ARC and MMLU. • IP overcomes predictive biases inherent to tokenization, such as those associated with different script families, providing reliable multilingual performance predictions. • Tokenization Parity and Tokenizer Fertility are captured within the explained variance of IP, further highlighting its ability to holistically represent multilingual performance. In essence, IP offers a robust, scalable, and task/prompt-agnostic way to evaluate LLMs, even for low-resource languages where existing methods often fall short. A huge thank you to my co-author, Alon Kipnis, for his invaluable contributions and mentorship throughout this work! Read the full paper here: https://lnkd.in/dzreCQ3N

23

1 Comment

A very interesting and useful paper: IndicGenBench is a benchmark for evaluating LLMs on different generation tasks across 29 Indic languages. The funniest and most surprising stuff that I found in the paper was (and I quote from the paper), "We find that Hindi in-context exemplars are much more useful for all models as compared to their English counterparts". In other words, LLMs understand Hindi in context examples better than English ones (considering the tasks of cross-lingual summarisation, machine translation, multilingual QnA, and cross-lingual QnA)!

10

A recent breakthrough titled "Matrix Multiplication-Free LLMs" demonstrates a huge advancement in the area of Large Language Models (LLMs) by reducing computational costs. The authors have eliminated MatMul operations from LLMs, claiming to 10 times reduction in memory usage and a 25.6% increase in training speed, all while maintaining strong performance at billion-parameter scales. Paper link: https://lnkd.in/ggph8qXc #AI #machinelearning #deeplearning #LLMs

16

Revolutionizing Spreadsheet Processing: A New Approach with LLMs After reading the SPREADSHEETLLM paper, I was inspired to develop standalone tools for structural-anchor extraction, inverted-index translation, and data-format-aware aggregation. These tools can be independently utilized and dynamically invoked by an LLM agent based on specific requirements. By integrating these methods with retrieval-augmented generation (RAG), we can significantly enhance token efficiency and maintain clear data representation. This innovative approach promises to revolutionize spreadsheet processing, making it more efficient and intelligent. Combining traditional algorithms with cutting-edge LLM capabilities offers a powerful solution for complex data management and analysis. https://lnkd.in/eFT6SUAq

4

I just looked at this paper that claims to have removed the multiplication of matrices from deep learning. The claim is that one can reduce the electricity demands by a huge margin. I needed to know how they got rid of the multiplication of matrices. The trick is simple to understand. Normally. all the multiplications are between inputs and weights. If they give weights only three possible values: 0, 1, or -1, they can "multiply" any number with those values without actually multiplying. To multiply with zero you return zero. To multiply with 1, you just return the input value. And to multiply with -1, you just change the sign on the input (which is a one-bit operation) and return the changed value. This saves a lot of electricity because the normal floating-point multiplications are expensive -- a lot of work for the processor. This confirms that there is still a lot to optimize in the current DNN implementations. We probably do not need a 16-bit addition either. This can be reduced too, probably again without much loss in model performance. I may have missed something as I ran through the paper quickly.

16

I tried an experiment! Ever wondered how to systematically evaluate LLMs beyond the 'vibe test'? In my latest blog post, I attempt to write LLM evaluations in pytest, a popular testing framework. I share insights on setting up evals that mimic human checks and discuss the challenges and solutions in automating these tests. Curious to learn more? Read the full story here and let's discuss! How do you handle evaluations in your LLM work? #ai #machinelearning #pytest #llm #softwaretesting https://lnkd.in/e6a3Gg_s Feel free to share your thoughts in the comments or share this post with others who might find it interesting!

35

6 Comments

🔍 Ever wondered if there's a unified framework for Retrieval-Augmented Generation (RAG)? We've formalized the retrieval enhancement paradigm with consistent notations, synthesizing research across all machine learning domains. Check out our new preprint: https://lnkd.in/e5iJv99z 📍Retrieval augmentation isn't just for language models! It can be generalized across machine learning, extending its impact far beyond language generation to areas like speech processing, computer vision, time series prediction, and computational biology. 🌟 We introduce Retrieval-Enhanced Machine Learning (REML), a formal framework that integrates retrieval mechanisms into machine learning models to enhance their performance. Our key contributions: 1️⃣ Synthesis and Unification We've synthesized existing research and unified it into a coherent framework, providing consistent notations that bridge various ML domains. 2️⃣ Modularization of REML Retrieval-Enhanced Machine Learning (REML) is modularized into distinct components: Querying, Searching, Presentation, Consumption, Storing, Optimization, and Evaluation. This modular approach makes it easier to understand and implement. 3️⃣ Enriching with IR Research Information Retrieval (IR) research is often overlooked by NLP researchers. Each modularized part of REML is enriched by the rich history and foundational principles of IR research. 4️⃣ Proposing Future Research By synthesizing and unifying existing work, we identified numerous gaps and areas for improvement. We propose numerous future research opportunities in this emerging field of REML. Incredible work with Alireza Salemi, Andrew Drozdov, Fernando Diaz, and Hamed Zamani

69

🚀 Excited to share my latest in-depth blog post exploring a cutting-edge AI framework that combines Hyperdimensional Memory Matching (HDMM), Neuro-Symbolic Context Mapping (NSCM), Evolutionary Reinforcement Scoring (ERS) powered by Genetic Algorithms (GAs) and Linear Regression for advanced feature prioritization and decision-making. This framework isn’t just another iteration of AI — it’s a paradigm shift. By leveraging hyperdimensional computing for high-capacity memory representation, symbolic reasoning for human-like interpretation and evolutionary algorithms for dynamic optimization it creates a system that’s scalable, interpretable, and robust. Key highlights: 📈 HDMM enables unmatched scalability for high-dimensional data encoding and matching. 🧠 NSCM bridges symbolic reasoning with data-driven models for enhanced decision-making. ⚙️ ERS with GAs optimizes features in a way that’s both adaptive and efficient. 🧮 Linear Regression integrates seamlessly for actionable insights and predictions. This isn’t just about better AI — it’s about creating systems that mimic cognitive processes, adapt dynamically, and deliver transparent and interpretable results. Let’s push the boundaries of what’s possible in AI! 💡 Read the full article here #ArtificialIntelligence #MachineLearning #HyperdimensionalComputing #NeuroSymbolicAI #GeneticAlgorithms #Innovation #FutureOfAI

1

I wanted to share some insights from a recent paper from Amazon Science in which authors tackled an important challenge in language modeling: minimizing sequence truncation. In language models, truncating input sequences might lead to a loss of crucial information, especially when the models are dealing with long-context tasks such as document summarization or story generation. The authors introduce a clever solution by leveraging a heuristic from the Best Fit Bin Packing problem, an NP-hard problem, to reorder input documents and reduce truncation. The proposed approach helps retain more information compared to other strategies by optimizing how sequences are packed into the model's context window. Authors show that the bin packing solution had biggest performance gain reading comprehension (+4.7%), natural-language inference (+9.3%), context following (+16.8%), and program synthesis. They also found that best-fit packing helped prevent closed-domain hallucination. Authors claim that while the experiments primarily focused on LLM pretraining, best-fit packing is broadly applicable to fine tuning as well. https://lnkd.in/gWquw3ud

28

3 Comments

Just dove into a super cool study, "ASTRAIOS: Parameter-Efficient Instruction Tuning of Code Large Language Models" and guess what? It's part of the 2024 LLMs paper list ! Here’s the scoop: - ASTRAIOS tested 28 models with different tuning tricks on tasks from code comprehension to generation. - Found out some methods like LoRA really save on costs without skimping on performance. - Big models are great but can be a bit clumsy when it comes to robustness and security. Check it out for a full rundown on making AI smart yet efficient: https://lnkd.in/eRZdktKU #AI #MachineLearning #SoftwareEngineering #Coding #Tech #Innovation #ArtificialIntelligence #DeepLearning #TechCommunity #2024LLMsPapers #CodeGeneration #AITuning #EfficiencyInAI

1

Hi, fellow scientists/researchers! I tried to post my research paper on arXiv.org, and it says that I need to get an endorsement from someone who has posted their paper in the cs.AI category. If you have your paper posted under cs category, please help with this. The endorsement link and code are attached below. Thanks! [My info and research] Asuka Inoshita:  https://lnkd.in/gRvz92Qd Submitting paper: https://lnkd.in/gADwfadp [arXiv endorsement request info and forms] Asuka Inoshita requests your endorsement to submit an article to the cs.HC section of arXiv. To tell us that you would (or would not) like to endorse this person, please visit the following URL: https://lnkd.in/gHQWExwj If that URL does not work for you, please visit https://lnkd.in/g6UJAfVi and enter the following six-digit alphanumeric string: Endorsement Code: IV7IGW Have a great weekend! - Asuka

1

I'm hiring - I need a few researchers who have a passion for building small language models in addition to software engineers who prefer Python. We are constructing an ensemble of tiny domain-specific models that work together synergistically to solve complex problems for the enterprise at the Edge. This approach is akin to swarm intelligence, many task-specific models working together in real-time as it were. We will build AI without political biases (unlike Big Tech). We will be open and transparent - from our datasets to our models' weights and architectures. We will focus on heavily regulated industries that must understand "how the sausage is made" in order to use AI safely and responsibly per our models will become the backbone of major operating systems in the near future. In addition, we will support the DoD and our Warfighters in order to safeguard Democracy and protect Western principles. We will remain unwavering in our support of American values. Moreover, if you want to solve mission-critical and challenging problems, join us. If you'd rather remain a cog in a massive machine at Google, Meta or Open AI while building politically biased and potentially dangerous AI - please stay there. We have venture backing, a large pipeline of committed contracts and line of sight to extreme commercial traction. Shoot me a DM and let's chat.

113

16 Comments

We just released a blog on whether the use of multimodal feature transfer can detect deception (phenomena like sarcasm, irony, and condescension) in conversation. The short answer... Yes. This particular write-up profiles a subset of our work at Symbl.ai on multimodal large language models, and is research led by one of our summer interns, Benjamin R. (from Prof. Larry Heck's AVA Lab at Georgia Institute of Technology). We have a paper currently under submission on this work, and we can't wait to share the full details with all of you soon. Blog Link: https://lnkd.in/g6m_wFDi #multimodal #featuretransfer #LLMs #deception #conversation #sarcasm #irony #language #AI #ML #NLP #LargeLanguageModels #SummerInternship #ResearchPapers #ACL #Llama2 #SymblDotAI

16