How Transformers and Hugging Face boost your ML workflows

Transcript

Good morning, ladies and gentlemen, and thank you for allowing us to share our machine learning vision with you today. Extracting quality insights from complex documents is critical for many organizations. So let me start with a question. How long would it take to build a machine learning-powered application running search queries on a corpus of SEC filings? What if we added voice queries and built-in translation for 21 languages? What do you think? Weeks? Months? More than a year? We'll answer that question in a few minutes, but for now, your best guess is probably way too long for my business needs. But why is that? Let's face it, machine learning is still a challenging field. In 2019, VentureBeat wrote that 87% of ML projects don't make it into production. Gartner confirmed this failure rate and predicted that throughout 2022, only 20% of analytic insights would deliver business outcomes. This prediction was actually confirmed by PWC in 2021, reporting that only 25% of companies see widespread ML adoption. Indeed, even skilled teams have to fight their way through accessing data, turning it into datasets, maintaining tools and infrastructure to experiment and train models, and in general, clearing all the technical and organizational hurdles on the way to production. Is there anything we can do to shorten the ML cycle, accelerate innovation, foster collaboration, and deliver business value in production quicker? We at Hugging Face think so. Our machine learning libraries are the fastest-growing ever open-source project on GitHub. Our website, aka the Hugging Face Hub, is visited over 1 million times every month by data scientists and engineers. It gives them immediate access to over 20,000 models based on the state-of-the-art transformer architecture. These models are pre-trained for many text, vision, audio, and speech tasks. You can use them as is to predict. You can also fine-tune them on one of our 2000 datasets or your data. You can easily share your models publicly with the open-source community or privately within your organization. On top of our open-source tools, we've built a commercial machine learning platform that helps companies get to the next level of agility and performance. It includes features like private models and datasets, access control and versioning, no-code AutoML with AutoNLP, and accelerated single-digit millisecond predictions with the inference API and Infinity. Thanks to Hugging Face, you can now bring state-of-the-art models to production quicker than ever before and create business value in a matter of days, not months. In fact, here is the voice query application I mentioned earlier. It uses two off-the-shelf models that I downloaded from the Hugging Face Hub. I'm using them as is, without any additional training. The first model is a multilingual speech-to-text model. In other words, it can convert an audio clip in 21 languages into an English sentence. This sentence, our search query, is fed to the second model which uses semantic search to find the closest matches in the document corpus. Here, I'm using the 2020 10K filings for all S&P 500 companies. Let's run some examples. Let's start with an example in French. And we can see the translation and the matching documents. Let's do one in Spanish. Again, the translation and the matching documents. Now let's do a final example where I record my own voice. "Qui est le CFO de GAP?" And we get a good answer on who the CFO at GAP is. So how difficult was this? Less than 100 lines of Python code, user interface included. A trained engineer can do this in a day. Hugging Face brings software engineering agility and best practices to machine learning. Can you imagine how much innovation and competitive edge your own teams could deliver? My best guess is quite a lot, and we'd be happy to help you get started. Thank you again for your time. Enjoy the rest of your day, and I wish you a very happy 2022.