Bevan Emma Huang

TL;DR: Decouple data science from data/platform engineering, especially with LLM workloads that require significant compute resources. Use off-the-shelf platform resources as much as possible so you can unleash data scientists quickly. Data scientists deliver greatest value to the organization through analysis, so use them that way. My only addition is to make sure you have the best available data science talent for your domain. I'll take a small team of high-caliber data scientists any day over a large team of so-so talent. Quality will always win in data science, because the right answer to the right question at the right moment is the only thing that matters. A wrong answer is worthless, obviously, but so too is the right answer to the wrong question. https://lnkd.in/eJ3A7Sg6

11

We make decisions weighing out uncertainties and risks all the time… we need models to present their outputs with uncertainties to make fully informed decisions as we use them

7

MIT's Data Provenance Initiative reports a significant decrease in available AI training data. Key points: - 5% of all data and 25% of high-quality sources in training sets are now restricted. - 45% of data in the C4 set is restricted by terms of service. Publishers and data holders are setting up paywalls, changing terms of service, blocking web crawlers, and taking legal actions, impacting AI development and public data use. Read more: - MIT Study: ["Consent in Crisis"](https://lnkd.in/eV-Xgk4Q) - New York Times Article: ["Data Disappearing Fast"](https://lnkd.in/exQWyEHz)

9

2 Comments

Another insightful study from James Zou on the importance of continuous monitoring and assessment of Language Models (LLMs) to ensure their reliability and effectiveness across various applications. Zou and his team have shared their curated collection of questions and answers from GPT-3.5 and GPT-4, encouraging further research in this field. Check out the paper to learn more about the significance of LLMs and their impact on the future of natural language processing. #LLM #LanguageModels #NaturalLanguageProcessing #GPT3.5 #GPT4 https://lnkd.in/evy4V6FD

24

6 Comments

One week until my book R for the Rest of Us: A Statistics-Free Introduction is out. For the next week, I'm posting summaries of each chapter. Here's Chapter 2: Principles of Data Visualization! This chapter delves into the principles of high-quality data visualization, using a drought visualization created by Cédric Scherer and Georgios Karamanis as an example. It introduces the grammar of graphics, a theory that underlies the ggplot2 package in R, and demonstrates how to use ggplot2 to recreate the drought visualization step-by-step. Key Takeaways • Minimalist aesthetics, such as removing grid lines and unnecessary text labels, can make a data visualization more effective by keeping the focus on the data. • The grammar of graphics provides a consistent way to describe and create all types of graphs, rather than thinking of them as distinct chart types. • ggplot2 implements the grammar of graphics, allowing users to build up a visualization in layers by mapping data to aesthetic properties, choosing geometric objects, adjusting scales, and setting the overall theme. Read it online!

65

1 Comment

**Introducing GraphRAG by Microsoft Research!** GraphRAG, now available on GitHub, is a tool for complex data discovery. It leverages a large language model to create rich knowledge graphs from text documents, enabling precise question-answering over private datasets. This approach outperforms naive retrieval-augmented generation (RAG) by providing structured information retrieval and comprehensive response generation. Ideal for answering global questions, GraphRAG supports informed decision-making with higher comprehensiveness and diversity. 🔗 [Read more](https://lnkd.in/eEcukH_A)

56

I've added four new tutorials to the "Fundamentals" section of "Practical Cheminformatics Tutorials", bringing the total number of Jupyter notebooks to 28. The additional tutorials include:  Recursive SMARTS Reaction Enumeration - the basics Reaction Enumeration - advanced topics Enumerating Stereoisomers and Tautomers  All the tutorials can be run on Google Colab without installing any software locally.   https://lnkd.in/eQ-RbdTx

579

8 Comments

Highly recommend watching this interview for a fact-based, clear-eyed view of how the latest generation of AI can (and will) create value for enterprises. A few highlights that resonate with my experience: 1. Internally within enterprises the smart AI investments leverage automation to accelerate knowledge worker productivity, such as using internal data to train AI "helpers" that improve speed, quality, and throughput. Not writing faster emails or building decks with a few clicks... think fulfilling client orders, curating/annotating saleable data, or anything else that creates measurable competitive advantage. 2. Once you build these helpers, you also need to invest time and energy to support knowledge workers as they learn how to harness these new superpowers. If you got a bionic arm, you'd need to get used to it before you're comfortable enough to use it effectively. Same goes for AI. 3. You have to nail the use case, because AI systems do some things really well, some things terribly, and a lot of things so-so. Value creation comes from the first bucket. 4. Mostly AI will show up as features in existing products and platforms, reducing friction and widening existing capabilities. Bringing AI into platforms and workflows is a smart way to think about investment, and it makes it very easy to evaluate ROI. 5. Alas, too many non-technical business leaders fail to understand the massive difference between small-scale PoC work -- where costs are negligible and infrastructure data+compute infrastructure are non-issues -- and production-grade AI applications. There's a lot of sticker shock. https://lnkd.in/e862ye5f

6

🚨New Publication Alert : Qwen2 Technial Report The Qwen2 Technical Report introduces a new series of large language models, with sizes ranging from 0.5 to 72 billion parameters. The flagship model, Qwen2-72B, achieves high scores across various benchmarks, demonstrating strong capabilities in language understanding, generation, multilingual proficiency, coding, mathematics, and reasoning. The instruction-tuned variant, Qwen2-72B-Instruct, also excels in specific benchmarks. Qwen2 supports about 30 languages and is openly available for community use, with resources for quantization, fine-tuning, and deployment on Hugging Face and ModelScope. For more details, you can view the [full report](https://lnkd.in/dhiXhE_r).

4

1 Comment

Week Two Helix Biogen Institute: From Newbie to Budding Bioinformatician! Can you believe it's only been two weeks at Helix Biogen Institute? It feels like I've learned a year's worth already! So, drumroll please... You're now looking at the newest (albeit very green) Bioinformatician in the building. Okay, maybe "in training" is more accurate, but let's not sweat the details! This week, I dove headfirst into Bioinformatics Module 1. Let me tell you, it's like being a detective, but instead of solving crimes, we're decoding life itself! Imagine mixing biology with super-smart computers, throw in some mind-bending algorithms, and voila – you've got Bioinformatics! We covered so much ground – from the types of data we can crunch, to all the cool branches and applications. Did you know Bioinformatics helps in designing new vaccines? My mind is officially blown. Of course, it's not all smooth sailing. We talked about the limitations too, but honestly, that just makes me more excited to be part of solving those challenges. Right now, I'm knee-deep in my first assignment. It's tough, but in that "I-can't-wait-to-figure-this-out" kind of way. Before I disappear back into my Bioinformatics bubble, let me just say: Science is AMAZING. If you're even a tiny bit curious, come check out Helix Biogen Institute. Who knows? You might find yourself geeking out over genetic sequences like me! Alright, duty calls – this assignment won't solve itself. Catch you on the flip side, fellow science enthusiasts!

10

Good initial answer. Rather than companies saying they do "AI" would like to see specific figures - i.e., our ML models are X% accurate at X and concrete information on what they are doing to further improve that accuracy. What kind of data are you using? What kind of compute? How are you preprocessing the data? How do you avoid overfitting as the data size increases?

1

Join me for an interesting cross industry discussion about how to think about managing data. 

8

For a weekend listen or viewing, highly recommend this talk from Jonathan Keeney who introduces Biocompute and its integration with Nextflow from recent Nextflow Summit 2024, Boston. Some of you who are in the business of building NGS pipelines in trial setting to meet primary or secondary endpoints might find this useful. In case of a submission, regulatory standards around FAIR principles can be be very strict as it should be, so containerized workflow need to capture data provenance, usuability, parameters, timestamp, user credentials, compute environment, error capture other than just final output results. Good to learn how it started around 2012 and where it is now with IEEE recognition.

23

#Ontologies, #KnowledgeGraph, #LLM & #RAG = all the hottest 🔥 buzzwords today 😎 Tony Seale provides one of the clearest explanation (in the post below) of their relationship and how the "reinforce" each-other.

15

2 Comments

🌟 Exciting blog post from NHS England on Reproducible Analytical Pipelines  🥁 Keep banging the drum Sam Hollings 📈 NHS England are sharing their journey on Reproducible Analytical Pipelines (RAP) and how it's enhancing the quality of their data processes. 🔍 What is RAP? RAP aims to automate statistical and analytical processes, making data analysis more efficient, transparent, and robust. This means less manual work and more time for data analysts to focus on what truly matters - delivering insights that improve healthcare services. 🌐 Why It Matters for the Culture of Data Services? Adopting RAP aligns the commitment to enhancing the quality and reliability of data services. It's not just about technology; it's about fostering a culture where continuous improvement and collaboration are at the forefront. 🛠️ Progress NHS England Data teams are integrating RAP across the organisation and It's still early days, but the impact is tangible and transferable. 📈 Looking Ahead They are committed to expanding the use of RAP, enhancing data capabilities, and continuing to share their progress. Hoping to hear more of this type of work. #NHS #DataScience #Innovation #RAP #HealthcareTechnology #Teamwork #Data #Health #Collaboration Link to the source in the comments. Image created by DALLE-3

17

1 Comment

This blog post is a long time coming, Applied Scientist Ryne Carbone dives into how he used Thompson Sampling to improve the bank linking experience at Ramp. Well worth a read.

9

When working with Generalized Linear Models (GLM) and Wald's approach to inference (confidence intervals, testing hypotheses) we often want to work on the response scale due to its easier interpretation. For example, when using #logisticregression to test hypotheses about probability of some event (like clinical success) working on the probability scale allows us to report difference in probabilities expressed in percentage points and test these differences. ⚠ But Wald's inference is symmetric, i.e. the interval estimates are symmetric about the point estimates. And since GLM employs a non-linear transformation of the conditional expectation (except the linear regression, where the link is identity), and since Wald's inference is not transformation invariant, this will make problems near to the boundaries, e.g. [0] and [1] probabilities, making the algorithm to work on the boundary of parameter space. 😵 Consequences? Wald's CIs easily can exceed 100% or be negative, especially in smaller samples (< 10), but this can happen even for several tens. While impact of this issue is usually smaller on the inference about their differences (differences often go between the 0.2 - 0.8 range), it can also be problematic for 2 reasons: ⚡ the difference itself may be close to the + or - boundary (0.9 - 0.6 = 0.3, which is far from 1, but 0.01 - 0.95 = -0.94, which is close to -1), so their CIs may exceed the boundary. ⚡ the 2nd order parameters - (co)variances, which are much more difficult to estimate, especially if you fit a repeated-observation model with unstructured covariance pattern, may be strongly affected. And since they are used to all further contrast testing... see? 💡 The logistic "S" curve resembles a straight segment in the 0.25-0.75 range - more or less. In this range it doesn't really matter on which scale you work. I prefer the probability scale for consistency with classic Wald's (or Rao's) "z" tests for 2 proportions, often requested by reviewers. / 💡 It's not surprising: Wald's p-value and CI of the average marginal effect (AME) over the simple logistic regression with a single 2-level categorical predictor, in absence of covariates, is 1:1 the Wald's z (or Pearson's chi2; whatever) test of 2 proportions with non-pooled standard errors. It's precisely the Wald's test of two % difference on the probability scale. / 👨‍🏫 Then, if we go with probabilities beyond the range we need to switch to the linear predictor scale (here - log odds), do the analysis, and transform the conditional expectations and their functions back to the response (probability) scale. This will give us asymmetric confidence intervals and much more stable (co)variance estimators. 😲 But the result of testing on the linear predictor scale may be different from those obtained on the probability scale. Then it's safer 👉 to report the result obtained on the linear predictor scale. #statistics #datascience #biostatistics #abtesting #abtest #data #research

107

15 Comments