Etienne Bernard, Co-Founder & CEO of NuMind – Interview Series

Etienne Bernard, is the Co-Founder & CEO of NuMind a software company founded in June 2022 specializing in developing machine learning tools. Etienne is an expert in AI & machine learning. After a PhD (ENS) & postdoc (MIT) in statistical physics, Etienne joined Wolfram Research where he became the head of machine learning for 7 years. During this time, Etienne led the development of automatic learning tools, a user-friendly deep learning framework, and various machine learning applications.

What initially attracted you to machine learning?

The first time I heard the term “machine learning” was in 2009 I believe, thanks to the Netflix prize. I found the idea that machines can learn fascinating and powerful. It was already clear to me that this would lead to plenty of important applications – including the exciting possibility of creating AIs. I immediately decided to dive into it, and never came back.

After getting a PhD (ENS) & postdoc (MIT) in statistical physics, you joined Wolfram Research where you became the head of machine learning for 7 years. What were some of the more interesting projects that you worked on?

My favorite kind of projects at Wolfram was developing automatic machine learning functions for the Wolfram Language (a.k.a. Mathematica). The first one was Classify, where you just give it the data and it returns a classifier. To me, machine learning has always been about being automatic. You don't tune the hyper-parameters of your human student, and you shouldn’t for your machine either! It was quite challenging from a scientific and software engineering perspective to create truly robust and efficient automatic machine learning functions.

Creating a high-level neural network framework was also a very interesting project. Lots of difficult design decisions about how to represent neural networks symbolically, how to visualize them, and how to manipulate them (i.e. being able to cut some pieces, glue others together, replace layers, etc.) I think we did a decent job by the way, and if it was open source, I’m pretty sure it would be heavily used 😉

During this period of time you also wrote a seminal book titled “Introduction to Machine Learning”, what were some of the challenges behind writing such a comprehensive book?

Oh, there were many! It took two years in total to write. I could have decided to just write a “how-to” book, which would have been easier, but part of my journey at Wolfram has been about learning machine learning, and I felt the need to transmit that. So the main difficulty was to figure out what to talk about exactly, and in what order, in order to make it interesting and easy to understand. Then there was the pedagogical details: should I use a math formula for this concept? Or some code? Or just a visualization? I wanted to make this book as accessible as possible and this gave me a lot of headaches. Overall I am happy with the result. I hope it will be useful to many!

Could you share the genesis story behind NuMind?

Okay. I wanted to create a startup for a while, originally in 2012 to create an auto ML tool, but the work at Wolfram was too much fun. Then around 2019-2020, the first large language models (LLMs) started to appear, like GPT-2 and then GPT-3. It was a shock to me how well they could understand and generate text. At the same time, I could see how painful it was to create NLP models: you needed to deal with an annotation team, to have experts running plenty of experiments, etc. I thought that there should be a way to use these LLMs through a tool to dramatically improve the experience of creating NLP models. My co-founder, Samuel (who happens to be my cousin), shared the same vision, and so we decided to create this tool.

The goal of NuMind is to spread the use of machine learning – and artificial intelligence in general – by creating simple yet powerful tools. What are some of the tools that are currently available?

Indeed. Our first tool is for creating custom NLP models. For example, let's say that you want to analyze the sentiment of your users from their feedback. Using an off-the-shelf model is generally not great, because it has been trained on a different kind of data, and for a slightly different task (sentiment analysis tasks are surprisingly different from each other!). Instead, you want to train a custom model that works well on your data. Our tool allows to do just that, in an extremely simple and efficient manner. Basically you load your data, perform a small amount of annotation, and get a model that you can deploy through an API. This is possible thanks to the use of LLMs, but also this new learning paradigm that we call Interactive AI Development.

What are some of the custom models that you are seeing developed from the first round of NuMind customers?

There have been a few sentiment analyzers. For example one client is monitoring the sentiment of group chats where people are helping each other fight their addictions. This analysis is needed in order to intervene in the rare case where the sentiment is declining. Another client uses us to find which job openings are best for a given resume – and by the way, I believe there is a lot of potential in these sorts of matchmaking AIs. We also have customers that are extracting information from medical and legal documents.

How much time savings can companies see by using NuMind tools?

It is application dependent of course, but compared to traditional solutions (labeling data and training a model separately), we see up to a 10x speed improvement to obtain a model and put it into production. I expect this number to improve as we continue developing the product. Eventually, I believe projects that would have taken months will be completed in days, and with better performance.

Could you explain how NuMind’s Interactive AI Development works?

The idea of Interactive AI Development comes from how humans teach each other. For example, let’s say that you hire an intern to classify your emails. You would first describe the task and its purpose. Then you might give a few good examples, some corner cases maybe. Then your intern would start labeling emails, and a conversation would begin. Your intern would come back with questions such as “How should I label this one?” or “I think we should create a new label for this one”, or even asking you “why” we should label a certain way. Similarly you might ask questions to your intern to identify and correct their knowledge gaps. This way of teaching is very natural and extremely efficient in terms of exchange of information. We are trying to mimic this workflow in order for humans to efficiently teach machines.

In technical terms, this workflow is a low-latency, high-bandwidth, multimodal, and bidirectional communication between the human and the machine, and we decided to call it Interactive AI Development to stress the bi-directionality and low-latency aspects. I see this as a third paradigm to teach machines, after classic programming, and classic machine learning (where you just give a bunch of examples of the task for the computer to figure out what to do).

This new paradigm is unlocked by LLMs. Indeed, you need to have something that is already somehow smart in the machine in order to efficiently interact with it. I believe this paradigm will become common place in the near future, and we can already see glimpses of it with chat-based LLMs, and with our tool of course.

We are applying this paradigm to teach NLP tasks, but this can – and will – be used for so much more, including developing software.

Is there anything else that you would like to share about NuMind?

Perhaps that it is a tool that can be used by both expert and non-experts in machine learning, that it is multilingual, that you own your models, and that the data can stay on your machine!

Otherwise we are in a private beta phase, so if you have any NLP needs, we would be glad to talk and figure out if/how we can help you!

Thank you for the great interview, readers who wish to learn more should visit NuMind.