Adaptive Automated Machine Learning

The ever-growing demand for machine learning has led to the development of automated machine learning (AutoML) systems that can be used off the shelf by non-experts. Further, the demand for ML applications with high predictive performance exceeds the number of machine learning experts and makes the development of AutoML systems necessary. Automated Machine Learning tackles the problem of finding machine learning models with high predictive performance. Existing approaches incorporating deep learning techniques assume that all data is available at the beginning of the training process (offline learning). They configure and optimise a pipeline of preprocessing, feature engineering, and model selection by choosing suitable hyperparameters in each model pipeline step. Furthermore, they assume that the user is fully aware of the choice and, thus, the consequences of the underlying metric (such as precision, recall, or F1-measure). By variation of this metric, the search for suitable configurations and thus the adaptation of algorithms can be tailored to the user’s needs. With the creation of a vast amount of data from all kinds of sources every day, our capability to process and understand these data sets in a single batch is no longer viable. ... mehrBy training machine learning models incrementally (i.ex. online learning), the flood of data can be processed sequentially within data streams. However, if one assumes an online learning scenario, where an AutoML instance executes on evolving data streams, the question of the best model and its configuration remains open.
In this work, we address the adaptation of AutoML in an offline learning scenario toward a certain utility an end-user might pursue as well as the adaptation of AutoML towards evolving data streams in an online learning scenario with three main contributions:
1. We propose a System that allows the adaptation of AutoML and the search for neural architectures towards a particular utility an end-user might pursue.
2. We introduce an online deep learning framework that fosters the research of deep learning models under the online learning assumption and enables the automated search for neural architectures.
3. We introduce an online AutoML framework that allows the incremental adaptation of ML models.
We evaluate the contributions individually, in accordance with predefined requirements and to state-of-the- art evaluation setups. The outcomes lead us to conclude that (i) AutoML, as well as systems for neural architecture search, can be steered towards individual utilities by learning a designated ranking model from pairwise preferences and using the latter as the target function for the offline learning scenario; (ii) architectual small neural networks are in general suitable assuming an online learning scenario; (iii) the configuration of machine learning pipelines can be automatically be adapted to ever-evolving data streams and lead to better performances.