I'm working on AutoML since 2016, I've created AutoML cloud platform mljar and now working on open-source package mljar-supervised. Tldr autoML finds good defaults fast and is only getting better, keep an eye on it. By all means create your own but having an automatic baseline to compare your solution against is an invaluable asset in itself. They often use good default models and will often outperform hand crafted models. The main arguement I would give for why you should use them is as a sanity check or as a starting point. There is also some recent work in optimizing Nueral Architecture Search, making it more and more feasible to find suitable models for the wide range of tasks Nueral methods are capable of modelling. Many rely on Bayesian Optimization or some form of gradient descent but Tpot has found some competitive success with genetic algorithms. that were designed from a research perspective, taking part in open competitions and attempting to optimize algorithm selection and hyperparameter optimization in new ways. There is also another camp of tools such as auto-sklearn, auto-weka, Tpot etc. Ludwig by Uber is an open source library with a command line tool for quickly generating models, a great example of how quick and easy it is now to get a default model. This includes feature engineering and data preprocessing along with model selection and tuning, throwing in some other features such as model deployment and API's. Tools like H2o, Datarobot and MLjar aim to automate as much as the process as possible to enable a wider audience to use ML, opening up a field of new customers. My impression is that AutoML seems to be a very general box into which many very specific tools are shoved into. I've just begun a PhD in this field and one of my first goals is to go through as many AutoML libraries and tools, reading any relevant literature for them.
0 kommentar(er)
