Compare MATCH results to auto-sklearn

[bruffridge]

https://automl.github.io/auto-sklearn/master/examples/20_basic/example_multilabel_classification.html#sphx-glr-examples-20-basic-example-multilabel-classification-py

[bruffridge]

Though they have weights entirely learned automatically through various gradient descent methods, most state-of-the-art language models still bear the touch of a human designer through their architectures. AutoML is a developing paradigm in machine learning which attempts to automate the design of the model architecture. In this paradigm, many machine learning models are trained to accomplish the same task. The architectures of these models are constructed of components which are connected to one another in arbitrary configurations. Some models will prove themselves to be more accurate than others; their votes will be weighted more heavily in the ensemble’s predictions. The training process finds not only the model weights, but the best-performing models.
In the genetic algorithm variation of this concept, the worst-performing models are winnowed out, replaced by mutated versions of the better-performing models. This allows a form of natural selection to take place, and ideally encourages increasingly better-performing model architectures to evolve naturally.
We have explored the usage of AutoML libraries such as auto-sklearn. After three hours of training the PeTaL dataset, our multilabel auto-sklearn implementation was unable to learn anything at all. However, we have yet to explore many other AutoML configurations and implementations.