Three papers from the Machine Learning Section accepted at NIPS
The thirty-second Conference on Neural Information Processing Systems (NIPS) has accepted three papers from researchers in DIKU's Machine Learning Section. The conference will take place in Montreal, Canada in the beginning of December.
In one paper, the researchers introduce a new framework and algorithm to make decisions under limited feedback in high-dimensional action spaces. Another paper introduces and analyses a new algorithm for online learning that adaptively exploits limited feedback to improve performance for easy data while guaranteeing robustness for worst-case data. And the third paper presents a new type of convolutional neural network for 3D data.
Below you will find extended descriptions of each paper.
Factored bandits: Operating with high-dimensional structured actions
The paper identifies the minimal structural assumptions that are required to learn efficiently in problems, where high-dimensional actions can be factored into a product of atomic choices. The paper unifies and generalizes a number of problems from the literature and provides a single all-purpose algorithm that is competitive with state-of-the-art problem-tailored solutions.
We introduce the factored bandits model, which is a framework for learning with limited (bandit) feedback, where actions can be decomposed into a Cartesian product of atomic actions. Factored bandits incorporate rank-1 bandits as a special case, but significantly relax the assumptions on the form of the reward function. We provide an anytime algorithm for stochastic factored bandits and up to constants matching upper and lower regret bounds for the problem. Furthermore, we show that with a slight modification the proposed algorithm can be applied to utility based dueling bandits. We obtain an improvement in the additive terms of the regret bound compared to state of the art algorithms (the additive terms are dominating up to time horizons which are exponential in the number of arms).
Adaptation to Easy Data in Prediction with Limited Advice
The paper identifies the minimal amount of extra information required to take advantage of two types of easy data in learning with limited feedback. The two types of easiness considered are i.i.d. (independent identically distributed) data and data with limited effective loss range (which is data, where the losses can vary freely, but are restricted to stay close together within each round - see the illustration below). A novel algorithm is constructed that can exploit easiness without compromising on performance in the worst case and requiring no prior information about the nature of the learning problem (easy or hard). The algorithm has potential applications in domains where feedback is expensive to acquire and it is essential to achieve the best with minimally possible amount of feedback.
3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data
Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, Taco Cohen
A collaborative effort between researchers in Amsterdam, Lausanne, and the Department of Computer Science in Copenhagen has resulted in a new type of convolutional neural network for 3D data. This network is "steerable", in the sense that it responds predictably to rotations or translations of the input (it is SE(3)-equivariant). The resulting model automatically learns features in any orientation, making it both more accurate and more data efficient than standard convolutional neural networks. The advantages of the approach are highlighted on two applications from molecular modelling: predicting amino acid probabilities within a protein and protein structure classification.
The 3D steerable convolutional neural network naturally works with geometric data such as vector fields, both as input and internal features. Below is an example of a protein structure encoded as a vector field. The model will ensure that any rotations of the input will result in corresponding rotations of the feature maps (see this video on 3D Steerable CNNs for examples).