Who needs fastai anymore? We have already been replaced by machines. Just kidding.
Here is a link to what Google released today. Sounds like they had mission statement similar to Jeremy and Rachel. In short, they are turning on image recognition to the masses. Just supply data and be done! Maybe over simplified, but who will try it first?
This post is similar, more of a follow on, to this post here in the fastai forums.
To share some industry perspective :
(1) This is a nice UX on top of transfer learning. From what I can understand, the AutoML Research work and ‘Cloud AutoML’ don’t have much in common. For applying the learnings from ‘Neural Architecture Search with Reinforcement Learning’, it would take a lot of waiting time in production with not as much gains.
(2) Microsoft has had a similar service called CustomVision.ai under its Cognitive Services for a year now. Drag photos, train a classifier, get ROC curves, deploy to Cloud or to mobile. https://azure.microsoft.com/en-us/services/cognitive-services/custom-vision-service/ , without a single line of code. One interesting aspect is that it uses active learning, so you can quickly put it in a production setting, monitor incoming active photos, label any incorrect ones and it will learn on the fly in 2-3 minutes. The cool part is the optimizations it does to make mobile friendly models to be able to process as many frames per second as possible.
Like @coderama says, other companies have been doing something similar for a while. Clarifai is another example.
Apple recently open sourced Turi Create (from a company they bought last year for $$$). It’s a Python-based tool, like the fastai library, but again with the purpose of making it easier for non-ML practitioners to train their own models (mostly using transfer learning).
Today, Rachel written at length a blog post titled “An Opinionated Introduction to AutoML and Neural Architecture Search” to demystify Google’s AutoML and Neural Architecture Search that is part of a broader field known as AutoML hype.
The key takeaways for me are (quoting the blog post):
- How useful is AutoML? … For many machine learning projects, choosing a model is just one piece of the complex process of building machine learning products.
- … So does this mean that data scientists can be replaced? Not yet …
how can humans and computers work together to make machine learning more effective? The focus of augmented ML is on figuring out how a human and machine can best work together to take advantage of their different strengths.