Motivated by recent theoretical advances in the approximation theory of DNNs, this work takes an adaptive basis viewpoint of neural network training and initialization. This perspective naturally leads to a hybrid least squares/gradient descent trai…

This work is supported by a grant from the MIT-IBM Watson AI Lab. We also thank Jonas Mueller for helpful feedback.

Motivated by the Wald test statistic, we derived the M&U pruning criterion for neural networks. Our M&U pruning criterion takes both magnitude and uncertainty into account and can balance between them through the hyperparameter λ. It can be consider…

In this work, we show that unsupervised learning approaches (hierarchical clustering using K-Means, clustering using SOM, and deep learning) can be used to qualitatively classify the different intensities of chaos in household environments thus redu…

We presented a large-scale multimodal camera+LiDAR dataset, which is significantly larger, higher quality, more geographically diverse than any existing similar dataset. It covers 76km2 by union the diluted the ego poses by 150 meters. We showed dom…

In this paper, we studied the problem of password leaking on smartwatches via the on-board motion sensors. Although side-channel attacks based on motion data have been widely investigated on smartphones, the problem has so far been overlooked on sma…

In this paper we presented the adaption of the feature relevance bounds approach to ordinal regression data using the explicit order variant. The optimization problem was phrased by approximating the generalization ability of the model with a bound …