Machine Learning in itself is a set of methodologies and techniques that take datasets and turn them into (smart) software called models.
The in-between of this process is dependent and performed by human machine learning experts.
Now, the entire idea of AutoML (Automated Machine Learning) is a set of methods and processes that are built such that the machine learning process (described above) can be accomplished without any human machine learning expert. And, well, it worked fine.
Google Cloud AutoML for instance, allows one to train their custom machine learning models for computer vision, natural language, and translation without writing any model code.
The whole AutoML idea and how exactly it worked confused me for a good period of my life, and one of the questions I used to ask was regarding how Google Cloud AutoML calculates it’s learning rate and decides the best calculations for cost/loss and optimization functions, and be able to perform on same level with neural networks designed by experts. Later on, I learnt about NASnet and founc this Google AI blog post about it.
We’ll cover how to get started with Google Cloud AutoML Vision, NLP, and Translation in the coming series.
The question then, may be “Why AutoML?” or “How much of good (or bad) has it been?”
The singular purpose of AutoML’s research was to enhance the capacities of Machine Learning as well as to speed up research in Machine Learning itself.
The successes that Machine Learning has achieved over the years are a product of direct critical reliance on human ML experts who carry out the manual piece of work, some of which include:
- Data pre-processing and cleaning,
- Constructing appropriate features,
- Selecting an appropriate model family,
- Optimizing hyper-parameters,
- Postprocessing machine learning models, and
- Analyzing the results obtained.
However, as non-ML experts continuously found that part of the work to be beyond them, and as the applications of machine learning in real life were becoming more and more demanding, the need for methods that could easily be used without expert know-how arose. This need brought about a research, and the results of that research that focuses on the progressive automation of ML were given the term Automated Machine Learning.
So, AutoML has turned out to be a great good, and apart from the insight it has provided in research, it’s really helped bring lots of machine learning ideas to reality — and without the expensive need of human experts.
C’est fini :)