The recently concluded Toronto Machine Learning Summit 2023 (TMLS 2023) brought together researchers, academics, and practitioners in the machine learning (ML) space. With an agenda including talks, roundtable discussions, and poster presentations, there was much to soak in on the latest trends and advancements in ML and MLOps. Scribble Data was a sponsor of the conference this year, and it gave us an opportunity to meet with some excellent fellow practitioners and enthusiasts in this space. Here are some of my takeaways from the two-day event:
LLMs take center stage
The buzz at TMLS2023 revolved around the remarkable rise of LLMs. Whereas last year’s conference had probably a couple of talks on LLMs, this year close to 40% of the talks were on this subject These advanced models, characterized by their massive size and complexity, have the potential to revolutionize the field of natural language processing (NLP). A number of topics of interest were discussed. Speakers emphasized that while training foundation LLMs is a challenging and risk-prone process, the growing availability of open-source variants encourages their widespread adoption. The talks highlighted various applications of LLMs, including information extraction, summarization, and even transforming natural language into code.
Prompt engineering and fine-tuning LLMs , crucial techniques for achieving the desired output from these models, were hot topics at the conference as well with a number of talks on best practices being shared by speakers.
Traditional-ML and MLOps remain vital
Amidst all the excitement surrounding LLMs, there was still a respectable roster of presentations at TMLS 2023 that underlined the enduring importance of traditional ML and MLOps. Speakers emphasized that feature stores, an essential component of ML pipelines, continue to play a crucial role in effective ML operations. Additionally, the conference delved into critical aspects of trustworthy ML, such as fairness, bias mitigation, and ML observability. The discussions reiterated the need for robust frameworks to ensure responsible and accountable machine learning practices.
LLMs products require careful thought before they can be rolled out
Despite the potential of LLMs, several challenges persist before their widespread deployment in production settings. One major obstacle is the limited use of LLMs beyond demonstrations and pilot projects due to concerns regarding cost, latency, accuracy, and trust. Also, there was general consensus at the conference, among speakers and attendees, that prompt engineering is currently more of an art than a science and is likely to become less significant as LLMs continue to evolve and improve. Where LLMs would find extensive deployment in the near future are systems that involve a human-in-the-loop approach where the LLM is used to assist a human in a decision task, but the human makes the final decision.
The talks at TMLS 2023 and the conversations we had reinforced Scribble’s thinking on two aspects.
First, traditional ML will continue to play an important role in organizations as they strive to derive value from data, and there is now tacit acceptance of the importance of the discipline and structure required when building production-ready ML systems. Second, Generative AI, in general, is a space to be plugged into. There is tremendous potential in using this technology to unlock fantastic opportunities, but before real production-grade use cases proliferate, there are challenges to be solved.
What are we building at Scribble?
At Scribble, we are building a data product platform that combines the power of generative AI and machine learning to solve advanced analytics use cases. If you are a business leader interested in addressing specific challenges, we would love to have a conversation with you. Additionally, if you’re seeking your next adventure and want to play a pivotal role in helping organizations streamline their data consumption, we invite you to explore the open roles we are currently hiring for. Visit our careers page at https://www.scribbledata.io/careers and drop us a line! We look forward to connecting with you.
We learned a lot from fellow attendees:
As always, it was an absolute pleasure making new friends and reconnecting with old ones including:
- Jim Dowling
- Simba Khadder
- Raj Vijayaraj
- Royal Sequeira
- David Scharbach
- Faraz Thambi
- Richard Zuroff
- Muzammil Hussain
- San Lee