There are several AutoML packages that simplify the machine learning process significantly and produce good results (if we trust what their documentation states). What is wrong with them then? Why do people still use other libraries to build their own models if they can simply use AutoML?
The main reason is that people want something unique, non-standard.
But modern AutoML packages try to find the best combination of pre-processing methods, learning algorithms, hyper-parameter, etc. Why this is not enough?
Because they don't support the combination of several non-standard methods, tools, approaches, etc. They mainly just loop over the finite set of the most popular approaches and try them. It can be enough for many cases, actually for most cases. If you have a common task and don't have time to fine-tune everything then probably AutoML is what you need. But remember, that the cutting-edge technologies and products don't rely on AutoML. They create and use their own models and ML pipelines.
I just want to add that even if you want to create your own model, AutoML may be a good choice to test your ideas, understand what works better for the task, or build a POC.
I don't agree with @srSikh that large companies use only their own solutions. Take a look at Google Cloud AutoML - https://cloud.google.com/automl . Their customers are big companies, like Disney, and they say that AutoML is quite useful for them.
Ok, @CarBOn. I was not completely right. Disney uses AutoML Vision to classify images or perform object detection. Probably Google AutoML is now at the level of development that allows producing a very good result for such common computer vision tasks. However, we don't know, maybe Disney just uses AutoML for start (like @FlonetTus says), and then fine-tune the created model.
Just drop us an email to ...
Show more 
Log in with Google