Madeleine Beatty

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

Responsible and accountable AI models are essential as generative AI continues to revolutionize industries. The rise of this technology, however, presents challenges in understanding how these models make decisions. Mechanistic interpretability addresses this by going beyond traditional methods that track statistical relationships. A recent research work by a team from #Anthropic has involved visualizing which neurons activate in response to specific prompts, advancing monosemanticity—ensuring each neuron’s activation correlates to a single, clear function. This step is crucial for fostering more transparent and safer AI systems in various applications. This is a great progress toward interpreting the thinking mechanisms of large language models (LLMs). #GenerativeAI #AI #LLMs #GenerativeAI #Grainger #MachineLearning #DeepLearning #ExplainableAI #DataScience

45

2 Comments

Worth a read. I like this quote: “Regulation helps us avoid mistakes and enables new benefits as we move into an uncertain future.”

3

2024 is the year of AI value, and time for companies to get out of PoC purgatory. “I really want to start driving use of AI at scale,” Luke G., head of analytics and AI at TD told The Wall Street Journal. “I think now is the time.” https://gag.gl/w2TUIR In his recent RevX keynote, Domino Data Lab's Nick Elprin detailed how companies do this today. Watch and learn how to accelerate AI impact: https://gag.gl/oN6hJo #AI #datascience #GenerativeAI

1

The most effective data literacy programs focus on those who can immediately apply new skills in analytics, data science, and AI. Without that, many training efforts miss the mark, Kjell Carlsson of Domino Data Lab explains in Nathan Eddy's latest IT Pro Today article on how tailored upskiling is key to driving real impact. Read on to learn how closing the data skills gap can reduce productivity losses and improve decision-making. 🔗 https://gag.gl/wKR30r #DataSkills #AI #Upskilling #DominoDataLab #Productivity

1

It was exciting, challenging, and rewarding to write this O'Reilly book Graph-Powered Analytics and Machine Learning with TigerGraph with my coauthors Phuc Kien Nguyen and Alexander Thomas. You can still get it for free from TigerGraph, if you act soon. We're appreciate your feedback and maybe a review on Amazon. #graphanalytics #machinelearning #graphalgorithms

45

3 Comments

What does marketing in a post-cookies world look like? https://lnkd.in/gN7BPfjQ #datascientist #dataanalyst #cfo #cmo #productmanagers #founders #cofounders #cxo #ceo #cpo #bizops #revops #operations #b2bsaas #marketing #marketingmanagers

1

My team member Monwhea Jeng just posted this article on external validity/non-stationary effects. It grapples with the assumption we make in A/B testing that this week's experiment results are sufficient to decide what will do better in production next week. Real experiments aren't always as clean and stable as we might hope. This article uses thousands of historic Microsoft experiments to analyze how often that occurs, and to give an indication of why (is it novelty/primacy? is it a different source of noise that's not accounted for in our analyses?).

4

Our team at Databricks keeps pushing the bar higher on Data Intelligence. Data practitioners can now leverage AI in their data journey using the Databricks Assistant, which is now generally available to our customers. Very proud of the team's accomplishment 🎉 #Databricks #DataIntelligence #AI

8

2 Comments

🎥 Just tested out Snowflake's new Cortex AI Analyst and pitted it against my own LangGraph GPT-4o agent—interesting results! 👀 Three key areas where I think Cortex falls short: 1️⃣ Stateless design – No memory capabilities, you can't ask follow on questions or refine your prompt. 2️⃣ Lack of independent reasoning – It’s more reactive than proactive. 3️⃣ Metadata stuffing – Handling large data warehouses with static semantic stuffing could lead to scaling and maintenance issues. 💡 But I was impressed by Cortex's ability to transform vague, ambiguous questions into something concrete, tailored to the data. My take: Cortex could thrive in a framework like LangGraph. It would solve the memory issue, allow more flexible metadata handling, and function better as a query-checking or ambiguity-resolution node, potentially in parallel with GPT. I imagine Cortex is fine-tuned for Snowflake, which means it might not play as well outside of it—but if your data lives in Snowflake, integrating Cortex within LangGraph could be a powerful combination! What are your thoughts? #AI #Cortex #Snowflake #LangGraph #OpenAI #SQL

13

3 Comments

Question for the community: In an IC track, how do you define the difference between a Non-senior, Senior, Staff and Principal Engineer (ideally thinking towards Analytics Engineering)? How do you think it differs in a small team compared to a big team (when, say, mentoring opportunity is limited)? A lot of definitions that I've seen around are vague and don't provide actionable and measurable qualities that can be easily understood as reasons/targets for progressing or not.

3

7 Comments

Try this hands-on mini-app to explore the data preparation stage of training a #LLM: https://lnkd.in/gtU43Ghi Training an #LLM is like teaching someone to read. Before the actual "teaching," we break down language into pieces the model can understand. We process raw text (books, articles, etc.) into a training-suitable format by splitting it into tokens (words or word parts) and creating a comprehensive dictionary. This crucial step forms the foundation of how the model will perceive language. 🌟 What’s Inside? 1. 📚 Identify the Corpus of Data: Either upload a text file or write your own. 2. 🔍 Break Down the Data into Tokens: App will break down your text into smaller chunks called "tokens" * 📖 Generate Dictionary: The app then creates a comprehensive dictionary of all the unique tokens, all sorted and ranked. 🌟 Features: 📚 Input your data: Upload a text file or write your own 🔍 Tokenize: See your text broken into "tokens" 📖 Generate a dictionary: Get a sorted, ranked list of unique tokens -- In the next post, we will use BytePairEncoding (BPE) tokenizer; one of the most common tokenization technique that used to train the models such as GPT-3 released by OpenAI.

8

1 Comment

I went a bit over-the-top setting up my dev environment for #AdventofCode this year! I'll use it to learn Palantir Technologies Foundry and Polars. You can see how I setup the project on Foundry here: https://lnkd.in/eV3eN_Tm #coding #automation

4

My first dbt Coalesce conference was an amazing time. It was good to catch up with TJ Tripp, Zoe (Yi) Liang, and Zachary Lancaster, finally meet Kevin Hu, PhD and Jacob Mulligan in real life, and meet lots of new cool people in the data world along the way. My top 3 Learnings from the conference are: 📈 The semantic layer has come a long way in the past year and it's the gateway to the future of analytics (such as powering AI analytics assistants). 📖 Data literacy isn't a one-and-done training but its an ongoing mission. 🎤 Extensive slide notes are really helpful when 100 people are starting at you and your mind goes blank. What did y'all enjoy at Coalesce this year?

31

2 Comments

Who is my boss Open Internal Catalog (Polaris) & Snowflake RBAC: We think: (for easy maintenance) 1. create a reader catalog role. Read-only privilege on all Polaris Internal Catalog tables 2. Grant it to sf_read_principal 3. Create a Snowflake REST Catalog Integration 4. Create an Iceberg Table 5. Create an Analyst role that has read privilege on SF tables. 6. Grant Analyst role to a user. Snowflake Dremio

11

An oldie but a goodie. "Data Products vs Data Mesh...and a surprise guest!" Catch it @ https://lnkd.in/gAi7giyC *** I'm starting a newsletter 📰 (I've reached limits on LinkedIn). To stay in touch, join @ https://lnkd.in/g53ReSzM (Subscription is free and will come with benefits like free books signed by authors and the opportunity to be a guest on the #CarCast). *** #Data #Analytics #DataMesh #DataProducts #ArtificialIntelligence #AI #GenAI

17

3 Comments

Can LLM's reason without prompting? We have all worked on "prompt engineering", trying to convince our recalcitrant LLM on the right way to proceed in providing solutions to the problem at hand. What if that were not required? What if LLM's actually could reason on their own? This is what the paper "Chain of Thought reasoning without prompting - https://lnkd.in/gpmwk3uW" attempts to do. They find that " CoT reasoning paths can be elicited from pre-trained LLMs by simply altering the decoding process". Apparently, the top-k decoding process has the CoT paths built in. " Extensive empirical studies on various reasoning benchmarks show that the proposed CoT-decoding effectively elicits reasoning capabilities from language models, which were previously obscured by standard greedy decoding." A very interesting read and one that implies that maybe, the days of "prompt engineers" earning their pay, which, great while it lasted, might be coming to an imminent end.

25

2 Comments

https://lnkd.in/gyWUagxw This is a great piece. It doesn’t only just apply to data science but really to any new endeavor we set out on. I am about halfway through my MIT Data Science AI/ML certification. I have NEVER academically had to absorb this amount of material and new content in such a short calendar period. OTOH I have never learned more in such a short period of time. I have had to rewrite my book on how I study. I have to daily find multiple new sources for concepts I thought I had down a month ago. EVERY new source gives me new insights that I TOTALLY did not connect, prior. But I guess that is a good sign. Booz Allen Hamilton North Dakota State University Great Learning Global MIT Professional Education Massachusetts Institute of Technology Robert Ulichney, PhD Robert Bunnett University of Iowa Iowa State University Society of Actuaries Data Science University of Dayton Rio Salado College United States Air Force Academy Southern New Hampshire University

5

Excited to share that I had the honor of co-authoring a new publication: Function-Level Software Metrics for Predicting Vulnerable Code. Huge congratulations to Md Rayhan Amin on the publication, and shout-out to colleagues from Mississippi State University (The Department of Computer Science & Engineering at Mississippi State University)—Daniel Tanner, YASHITA SHARMA, and Tanmay Bhowmik! The paper was published in the 2024 IEEE International Conference on Information Reuse and Integration for Data Science (IRI). Abstract: "Limited experience or hard deadlines often lead developers to technical Q&A websites, e.g., Stack Overflow, to find quick answers to their coding problems. Research, however, suggests that a large share of the answers on Stack Overflow include vulnerable code and their use in a software system may cause major security implications. Researchers have suggested several approaches, e.g., leveraging software metrics, to predict vulnerability in the source code. The majority of these approaches require the whole code base or a big chunk of it to perform the prediction. The answers on those Q&A websites, however, are code snippets typically with a single function, or, at most a couple of related functions. Predicting vulnerabilities in such code snippets has received limited attention. Few examples present in the literature suggest leveraging function-level software metrics to predict vulnerable code. Such work, however, provides incomplete analysis with heavily imbalanced datasets. This paper examines whether function-level code metrics can differentiate the vulnerable and fixed versions of a given function. In particular, we develop a "near-balanced" dataset with a total of 132,754 function instances, calculate 18 function-level code metrics for both vulnerable (49.6% of the total) and fixed versions of the functions, and conduct binary logistic regression with the Wald test. The results suggest that function-level code metrics may not be useful in predicting if a given function contains vulnerability." Read more here: https://lnkd.in/gJyBapgD #Research #Publication #IEEE #IRI #IRI2024 #SoftwareMetrics #VulnerableCode #Vulnerability #SourceCode #DataScience #Cybersecurity #IRI24

16

3 Comments