Sahil Arora

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

Stuck in A/B testing purgatory? ⏰ There's a better way. Reinforcement learning (RL) personalizes user experiences in real-time, ditching the wait for statistical significance. RL handles complex interactions & learns from smaller data sets, leading to faster results & higher conversions. #CRO #MachineLearning #ExperienceOptimization #UX https://lnkd.in/gtCFGcKh

6

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

5,270

164 Comments

👉🏼 Assessing the Risk of Bias in Randomized Clinical Trials With Large Language Models 🤓 Honghao Lai https://lnkd.in/eBZ2cduS 🔍 Focus on data insights: - Large language models (LLMs) facilitate systematic reviews. - Methods and reliability of LLMs for assessing risk of bias (ROB) in randomized clinical trials (RCTs) are explored. 💡 Main outcomes and implications: - LLMs demonstrated high correct assessment rates for ROB in RCTs. - LLMs showed substantial accuracy and consistency in evaluating RCTs. - Potential of LLMs as supportive tools in systematic review processes. 📚 Field significance: - Enhancing efficiency and accuracy in assessing ROB in RCTs. - Advancing the use of LLMs in systematic review methodologies. 🗄️: #datainsights #riskofbias #RCTs #systematicreviews

1

I am currently working with some of the best Machine Learning Engineers in the US who are looking for new roles. Here are some of the common themes why they are on the market: ➕ Losing confidence in the technical direction of their company ➕ Wanting to move towards a GenAI-focused role ➕ 4+ years, RSUs have vested and want a new challenge ➕ Big tech work is repetitive; they want more high-impact work ➕ Still a lot of restructures and instability ➕ Lack of long-term company vision. Short-term panic affecting work Generally, candidates are very open to startups or large companies. As long as the ML work has a high impact If you are a candidate open to new roles. Get in touch; I have clients actively interviewing this week. If you are a hiring manager, now is a good time to pick up some of the best talent. Get in touch with People In AI 🙂 📧 samjones@peopleinai.com

20

The LLM will see you now: Google and DeepMind released a paper on Monday describing Med-Gemini, a group of advanced AI models targeting healthcare applications. The authors claim that Med-Gemini is outperforming competing models such as OpenAI’s GPT-4. During the pandemic we saw the rise of Telehealth which has largely come back to earth but could doctors be consulting more often with AI? (We know they already Google stuff) #Healthcare #AI

12

9 Comments

If only you could have fixed hallucinations of a pre-trained LLM through contextually relevant prompts. It would have been a solved problem by now! Contextual relevance (RAGs) improve domain specificity, not hallucination! Nice article. #llms #chatgpt4 #aiml #mlops #llmops #ml #rag https://lnkd.in/d44hdifr

11

Is In-Context Learning (ICL) equivalent to Gradient Descent (GD)? There is a common belief that applying ICL in #LLM functions like GD-based fine-tuning. But does this hold in real-world LLMs? I'm super excited that our work got picked for an Oral presentation at #ICML this year! This would be impossible without an amazing collaboration with Aayush M. and Daniel Khashabi at The Johns Hopkins University. Find out in our paper: https://lnkd.in/eXb_sqmM.

81

4 Comments

A recent study (Nori, et. al) shows that a well prompted foundation model, like GPT-4, can achieve results comparable to fine-tuned models such as BioGPT or Med-PaLM. This is significant because effective prompts can be created in a much more affordable manner than fine-tuning an LLM. Check out the study here: https://lnkd.in/eCsitTA8

22

The one USAJobs listing which mentions PyTorch is not on ai.usajobs.gov. I'm not sure if this site is a good idea. Between the federal jobs that aren't on USAJobs at all and the ones that are AI-related but aren't listed at ai.usajobs.gov, my concern is that people find the site, think it's comprehensive, and then don't look elsewhere. The larger ai.gov site describes specific programs/agencies, and that's really helpful, but you can wind up at ai.usajobs.gov without that context. Or you can click through to ai.usajobs.gov from ai.gov and reasonably assume that it has all the listings, since it's described as "across all federal agencies". The data science job that mentions PyTorch (Arlington, Office of the Inspector General, USPS, closes when they receive 50 applications) is here, btw: https://lnkd.in/e8XuNV-C

15

6 Comments

LLMs being autoregressive from left to right is one cause for problems regarding arithmetic, in addition to tokenization. Right-to-left languages aren't actually a problem because they are serialized in the writing order, right to left, and flipped only at the presentation layer. Although it can be trippy trying to read them with a hex editor. Anyway, people read numbers, that is, *Arabic numerals* from right to left. If you have a number 4,000, you have to first start from right to count ones, tens, and so on, to know what "4" is, whether it's hundreds or thousands. LLMs can't do that. Llama models nowadays tokenize the numbers by single number tokens, which is a step to the correct direction, although doesn't solve the reading order. OpenAI tokenizer tokenizes numbers in groups of 1-3. For example, the number 54204 is tokenized as: "542", "04" Imagine how difficult this makes reading numbers and doing arithmetic on them? It's harder than doing arithmetic with Roman numerals! It might make sense to have a better encoding for numbers for LLMs, not unlike we tokenize images with special tokens. Numbers are a different modality, they are read right-to-left, as a very special case not even shared with right-to-left languages.

5

the opposite argument here is with the design of the interface of these models. ex. ChatGPT, Gemini, and other leading models have a chat interface. this is a design flaw. use-cases like even basic research involve providing proper context for the model to produce better results. a chat interface doesn't influence or command that behavior. therefore we are practicing prompting techniques now. imo this can be easily solved by an appropriate design for the interface.

3

Hallucinations in LLMs: Are we making progress? The truth about RAG Retrieval-Augmented Generation (RAG) has been hailed as a solution to the problem of hallucinations in Large Language Models (LLMs). But is it really effective? To answer this question, researchers have created RAGTruth, a benchmark dataset designed to measure word-level hallucinations across various domains and tasks. The dataset contains nearly 18,000 responses generated by diverse LLMs using RAG, all carefully annotated at both the case and word levels. The status quo is that while RAG does help reduce hallucinations, LLMs can still present unsupported or contradictory claims even with the retrieved content. The researchers show that by using a high-quality dataset like RAGTruth, it's possible to finetune a relatively small LLM to achieve competitive performance in hallucination detection. This model is then used to filter the output of the main LLM. Both strategies that they tested - filtering for fewest detected hallucinations and only allowing responses with no hallucinations detected - reduced the hallucination rate effectively by up to 63.2%. The caveat is that they used quite large samples. For methods like this to be worthwhile, processes need to be further optimized to the point where small samples of 3-5 responsed will be sufficient. ↓ Liked this post? Join my newsletter with 25k+ readers that breaks down all you need to know about the latest LLM research: llmwatch.com 💡

316

7 Comments

Hallucinations in LLMs: Are we making progress? The truth about RAG Retrieval-Augmented Generation (RAG) has been hailed as a solution to the problem of hallucinations in Large Language Models (LLMs). But is it really effective? To answer this question, researchers have created RAGTruth, a benchmark dataset designed to measure word-level hallucinations across various domains and tasks. The dataset contains nearly 18,000 responses generated by diverse LLMs using RAG, all carefully annotated at both the case and word levels. The status quo is that while RAG does help reduce hallucinations, LLMs can still present unsupported or contradictory claims even with the retrieved content. The researchers show that by using a high-quality dataset like RAGTruth, it's possible to finetune a relatively small LLM to achieve competitive performance in hallucination detection. This model is then used to filter the output of the main LLM. Both strategies that they tested - filtering for fewest detected hallucinations and only allowing responses with no hallucinations detected - reduced the hallucination rate effectively by up to 63.2%. The caveat is that they used quite large samples. For methods like this to be worthwhile, processes need to be further optimized to the point where small samples of 3-5 responsed will be sufficient. ↓ Liked this post? Join my newsletter with 25k+ readers that breaks down all you need to know about the latest LLM research: llmwatch.com 💡

172

5 Comments

In general, there are three ways to get better outputs from LLMs: 1. Prompt engineering 2. Fine-tuning 3. Retrieval-Augmented Generation (RAG) There was a recent paper that pitted prompt engineering and fine-tuning up against each other in a controlled study on code review. The result? A fine-tuned GPT-3.5 was able to outperform a few-shot prompt with the base GPT-3.5 model. So does that mean fine-tuning is "better" than prompt engineering? Nope! If you look to research you'll be able to find data that slices either way. For example, in the fall of last year, Microsoft released a paper in which GPT-4 + their prompting framework, Medprompt, was able to beat a fine-tuned model, Med-PaLM 2, developed by Google. We did a deep dive on this paper which we've linked below. Fine-tuning and prompt engineering solve different parts of the equation and should be seen as additive. We always recommend teams start with prompt engineering, as it is the quickest to experiment with and will help you identify the limitations you're encountering and which of the other methods you can use as solutions. It's not an either/or question. Plus there are easy-to-use tools to help with both methods. PromptHub for prompt engineering and Entry Point AI for fine-tuning. For more on the question of prompt engineering vs fine-tuning (when to use each method, more of the research data etc), check out our rundown linked below.

30

9 Comments

It's no secret AI/ML X Biology is hot right now, with millions of dollars in funding going to companies with ML guided discovery strategies. This field is the next wave of startups, similar to CAR-T and CRISPR before them. Candidates looking to enter this field often ask about wetlab skills that pair well with ML guided strategies. Here are some skills requested by leading AI/ML X Bio companies: - Display Technologies: Yeast, Phage, mRNA, and Ribosome display are in high demand, with Phage being the most popular. Yeast is also gaining traction because of it's eukaryotic Expression benefits. - HTS Screening: With the evaluation of molecules derived by AI/ML, extensive screening is inevitable. - Automation: Developing assays and workflows for automated liquid handlers is crucial in this field. - CFPS / IVTT: Ideal for expressing multiple proteins simultaneously, despite challenges with folding and disulfide bonds. But some experts can help you solve these issues. ( We know them 😊) - SPR & BLI: Candidates skilled in SPR are highly sought after by companies. - Molecular Dynamics Data: Individuals who can integrate molecular dynamics data into ML algorithms for structure and binding prediction are in high demand. Upskilling in any of the mentioned skills can significantly benefit your career in AI/ML X Biology. While there may be other skills to consider, these are the most requested ones based on current trends in the industry. Unashamed Segue CoFactor is currently recruiting for display experts on both coasts. We are also in need of research associates with SPR. Please reach out for more information. #ML #biology

37

1 Comment