Large Scale Kernel Methods for Fun and Profit

Kernel methods are among the most flexible classes of machine learning models with strong theoretical guarantees. Wide classes of functions can be approximated arbitrarily well with kernels, while fast convergence and learning rates have been formally shown to hold. Exact kernel methods are known to scale poorly with increasing dataset size, and we believe that one of the factors limiting their usage in modern machine learning is the lack of scalable and easy to use algorithms and software. The main goal of this thesis is to study kernel methods from the point of view of efficient learning, with particular emphasis on large-scale data, but also on low-latency training, and user efficiency. We improve the state-of-the-art for scaling kernel solvers to datasets with billions of points using the Falkon algorithm, which combines random projections with fast optimization. Running it on GPUs, we show how to fully utilize available computing power for training kernel machines. To boost the ease-of-use of approximate kernel solvers, we propose an algorithm for automated hyperparameter tuning. By minimizing a penalized loss function, a model can be learned together with its hyperparameters, reducing the time needed for user-driven experimentation. In the setting of multi-class learning, we show that – under stringent but realistic assumptions on the separation between classes – a wide set of algorithms needs much fewer data points than in the more general setting (without assumptions on class separation) to reach the same accuracy. The first part of the thesis develops a framework for efficient and scalable kernel machines. This raises the question of whether our approaches can be used successfully in real-world applications, especially compared to alternatives based on deep learning which are often deemed hard to beat. The second part aims to investigate this question on two main applications, chosen because of the paramount importance of having an efficient algorithm. First, we consider the problem of instance segmentation of images taken from the iCub robot. Here Falkon is used as part of a larger pipeline, but the efficiency afforded by our solver is essential to ensure smooth human-robot interactions. In the second instance, we consider time-series forecasting of wind speed, analysing the relevance of different physical variables on the predictions themselves. We investigate different schemes to adapt i.i.d. learning to the time-series setting. Overall, this work aims to demonstrate, through novel algorithms and examples, that kernel methods are up to computationally demanding tasks, and that there are concrete applications in which their use is warranted and more efficient than that of other, more complex, and less theoretically grounded models.