The 2021 edition of the State of AI Report got here out last week. So did the Kaggle State of Machine Learning and Data Science Survey. There’s much to be learned and discussed in these reports, and a few takeaways caught my attention.
“AI is increasingly being applied to mission critical infrastructure like national electric grids and automatic supermarket warehousing calculations during pandemics. Nevertheless, there are questions on whether the maturity of the industry has caught up with the enormity of its growing deployment.”
There is no denying that Machine Learning-powered applications are reaching into every corner of IT. But what does that mean for corporations and organizations? How can we construct rock-solid Machine Learning workflows? Should all of us hire 100 Data Scientists ? Or 100 DevOps engineers?
“Transformers have emerged as a general purpose architecture for ML. Not only for Natural Language Processing, but in addition Speech, Computer Vision and even protein structure prediction.”
Old timers have learned the hard way that there’s no silver bullet in IT. Yet, the Transformer architecture is indeed very efficient on a wide range of Machine Learning tasks. But how can all of us sustain with the frantic pace of innovation in Machine Learning? Will we really want expert skills to leverage these cutting-edge models? Or is there a shorter path to creating business value in less time?
Well, here’s what I feel.
Machine Learning For The Masses!
Machine Learning is in all places, or at the least it’s attempting to be. A couple of years ago, Forbes wrote that “Software ate the world, now AI is eating Software“, but what does this really mean? If it implies that Machine Learning models should replace hundreds of lines of fossilized legacy code, then I’m all for it. Die, evil business rules, die!
Now, does it mean that Machine Learning will actually replace Software Engineering? There is definitely quite a lot of fantasizing at once about AI-generated code, and a few techniques are definitely interesting, akin to finding bugs and performance issues. Nevertheless, not only shouldn’t we even consider eliminating developers, we should always work on empowering as many as we are able to in order that Machine Learning becomes just one other boring IT workload (and boring technology is great). In other words, what we really want is for Software to eat Machine Learning!
Things should not different this time
For years, I’ve argued and swashbuckled that decade-old best practices for Software Engineering also apply to Data Science and Machine Learning: versioning, reusability, testability, automation, deployment, monitoring, performance, optimization, etc. I felt alone for some time, after which the Google cavalry unexpectedly showed up:
“Do machine learning like the nice engineer you’re, not like the nice machine learning expert you are not.” - Rules of Machine Learning, Google
There is no have to reinvent the wheel either. The DevOps movement solved these problems over 10 years ago. Now, the Data Science and Machine Learning community should adopt and adapt these proven tools and processes immediately. That is the one way we’ll ever manage to construct robust, scalable and repeatable Machine Learning systems in production. If calling it MLOps helps, fantastic: I won’t argue about one other buzzword.
It’s really high time we stopped considering proof of concepts and sandbox A/B tests as notable achievements. They’re merely a small stepping stone toward production, which is the one place where assumptions and business impact will be validated. Every Data Scientist and Machine Learning Engineer should obsess about getting their models in production, as quickly and as often as possible. An okay production model beats an awesome sandbox model every time.
Infrastructure? So what?
It’s 2021. IT infrastructure should not stand in the way in which. Software has devoured it some time ago, abstracting it away with cloud APIs, infrastructure as code, Kubeflow and so forth. Yes, even on premises.
The identical is quickly happening for Machine Learning infrastructure. In response to the Kaggle survey, 75% of respondents use cloud services, and over 45% use an Enterprise ML platform, with Amazon SageMaker, Databricks and Azure ML Studio taking the highest 3 spots.
With MLOps, software-defined infrastructure and platforms, it’s never been easier to tug all these great ideas out of the sandbox, and to maneuver them to production. To reply my original query, I’m pretty sure you could hire more ML-savvy Software and DevOps engineers, no more Data Scientists. But deep down inside, you type of knew that, right?
Now, let’s discuss Transformers.
Transformers! Transformers! Transformers! (Ballmer style)
Says the State of AI report: “The Transformer architecture has expanded far beyond NLP and is emerging as a general purpose architecture for ML”. For instance, recent models like Google’s Vision Transformer, a convolution-free transformer architecture, and CoAtNet, which mixes transformers and convolution, have set recent benchmarks for image classification on ImageNet, while requiring fewer compute resources for training.
Transformers also do thoroughly on audio (say, speech recognition), in addition to on point clouds, a way used to model 3D environments like autonomous driving scenes.
The Kaggle survey echoes this rise of Transformers. Their usage keeps growing 12 months over 12 months, while RNNs, CNNs and Gradient Boosting algorithms are receding.
On top of increased accuracy, Transformers also keep fulfilling the transfer learning promise, allowing teams to avoid wasting on training time and compute costs, and to deliver business value quicker.
With Transformers, the Machine Learning world is progressively moving from “Yeehaa!! Let’s construct and train our own Deep Learning model from scratch” to “Let’s pick a proven off the shelf model, fine-tune it on our own data, and be home early for dinner.“
It is a Good Thing in so some ways. State-of-the-art is continually advancing, and hardly anyone can sustain with its relentless pace. Keep in mind that Google Vision Transformer model I discussed earlier? Would you prefer to test it here and now? With Hugging Face, it’s the best thing.
How concerning the latest zero-shot text generation models from the Big Science project?
You’ll be able to do the identical with one other 16,000+ models and 1,600+ datasets, with additional tools for inference, AutoNLP, latency optimization, and hardware acceleration. We may assist you to get your project off the bottom, from modeling to production.
Our mission at Hugging Face is to make Machine Learning as friendly and as productive as possible, for beginners and experts alike.
We consider in writing as little code as possible to coach, optimize, and deploy models.
We consider in built-in best practices.
We consider in making infrastructure as transparent as possible.
We consider that nothing beats top quality models in production, fast.
Machine Learning as Code, right here, right now!
A number of you appear to agree. We now have over 52,000 stars on Github. For the primary 12 months, Hugging Face can be featured within the Kaggle survey, with usage already over 10%.
Thanks all. And yeah, we’re just getting began.
Interested by how Hugging Face might help your organization construct and deploy production-grade Machine Learning solutions? Get in contact at julsimon@huggingface.co (no recruiters, no sales pitches, please).
