Sign In

Kirk Mettler

Chief Data Scientist and R guy at IBM

Kirk Mettler is a Data Scientist with a strong background in classical statistical approaches and modern machine learning methods. With experience in industrial companies as both an in-house developer and senior management consultant to Fortune 500 clients, he excels at bridging communication between technical experts and non-technical users.

His areas of expertise include Consultative Business Development, R programming, Team Architecture & Deployment, Linear Models, Predictive Modeling/Analytics, Decision Trees, and involvement in Start-Ups/Venture Capital.

Kirk Mettler pursued a Graduate Study in IEOR at Columbia Engineering and holds a BS in Operations Research and Industrial Engineering from Cornell University. He completed his High School education at Choate Rosemary Hall, Middle School at The Hopkins School, and Elementary School at The Foote School.

Throughout his career, Kirk Mettler has held significant roles in various organizations. He served as the Chief Data Scientist at IBM, Director of Content Partnerships at DataCamp, Co-Founder at Scry Health, Partner at Floop and Bigcomputing, and CHIEF OPERATING OFFICER/CHIEF FINANCIAL OFFICER at Revolution Analytics (now Microsoft Advanced Analytics). He was also the GENERAL MANAGER at Mettler Realty, LLC and CHIEF OPERATING OFFICER at Circuit Wise.

Highlights

Aug 14 · linkedin

Coming into Wood's Hole yesterday from Martha's Vineyard, I saw this boat on the Oceanographic Institute's dock. It is Mayflower Autonomous Ship (MAS) project. I wondered why as it had completed is mission to travel unmanned across the Atlantic two years ago. Turns out the project continues and the Mayflower is getting some upgrades. Ahh the benefits of modular design.

https://lnkd.in/eaaUbCtF

May 19 · linkedin

I have now spent the last nine months working with LLMs and Foundational models in general. Overall, I find them very interesting tools that do a good to exceptional job at some tasks that I have struggled to get good results in the past no matter how much effort I put into building models and approaches for those tasks. We are moving forward and iterating at such a breakneck speed both in terms of the models themselves and the infrastructure to support those models. I believe one area that is noticeably lagging behind the other workstreams in measuring and monitoring model performance. That is not to say there are not people working in this area and doing good work. It is just that I find the approaches so far unsatisfying. Even the LLM I pinged on the topic seems a little disappointed in the current state: There are several ways to measure the accuracy of a large language model (LLM) like myself. Here are some common approaches: Perplexity: This is a measure of how well the model can predict the next word in a sequence. Lower perplexity scores indicate better predictive performance. Held-out Test Sets: A portion of the training data is held out and used to evaluate the model's performance on that unseen data. Metrics like accuracy, F1 score, BLEU score etc. are calculated. Human Evaluation: Humans evaluate the model's outputs for qualities like coherence, grammaticality, factual accuracy etc. and provide ratings. Task-Specific Benchmarks: For specific tasks like question-answering, translation, summarization etc., there are benchmark datasets where the model's performance is measured against human references or prior systems. Adversarial Testing: The model is tested on carefully constructed inputs designed to expose failures or inconsistencies. Probing: Internal representations of the model are analyzed to evaluate what linguistic capabilities it has acquired. No single metric fully captures all aspects of an LLM's performance. It's common to evaluate across multiple dimensions like factual accuracy, consistency, coherence, bias, toxicity etc. Rigorous evaluation remains an open challenge in the field of large language models.

Well I guess that just means we have more work to do. Happy Sunday

Kirk Mettler
Add to my network

Location

Milford, Connecticut, United States