The length of the geodesic between two data points along a Riemannian manifold, induced by a deep generative model, yields a principled measure of similarity. Current approaches are limited to low-dimensional latent spaces, due to the computational complexity of solving a non-convex optimisation problem. We propose finding shortest paths in a finite graph of samples from the aggregate approximate posterior, that can be solved exactly, at greatly reduced runtime, and without a notable loss in quality. Our approach, therefore, is hence applicable to high-dimensional problems, e.g., in the visual domain. We validate our approach empirically on a series of experiments using variational autoencoders applied to image data, including the Chair, FashionMNIST, and human movement data sets.

Abstract. Long Short-Term Memory (LSTM) is one of the best recent supervised sequence learning methods. Using gradient descent, it trains memory cells represented as differentiable computational graph structures. Interestingly, LSTM's cell structure seems somewhat arbitrary. In this paper we optimize its computational structure using a multiobjective evolutionary algorithm. The fitness function reflects the structure's usefulness for learning various formal languages. The evolved cells help to understand crucial features that aid sequence learning.

No abstract

We introduce a methodology for efficiently computing a lower bound to empowerment, allowing it to be used as an unsupervised cost function for policy learning in real-time control. Empowerment, being the channel capacity between actions and states, maximises the influence of an agent on its near future. It has been shown to be a good model of biological behaviour in the absence of an extrinsic goal. But empowerment is also prohibitively hard to compute, especially in nonlinear continuous spaces. We introduce an efficient, amortised method for learning empowermentmaximising policies. We demonstrate that our algorithm can reliably handle continuous dynamical systems using system dynamics learned from raw data. The resulting policies consistently drive the agents into states where they can use their full potential.

Abstract-The development of new, light robotic systems has opened up a wealth of human-robot interaction applications. In particular, the use of robot manipulators as personal assistant for the disabled is realistic and affordable, but still requires research as to the brain-computer interface.Based on our previous work with tetraplegic individuals, we investigate the use of low-cost yet stable surface Electromyography (sEMG) interfaces for individuals with Spinal Muscular Atrophy (SMA), a disease leading to the death of neuronal cells in the anterior horn of the spinal cord; with sEMG, we can record remaining active muscle fibers. We show the ability of two individuals with SMA to actively control a robot in 3.5D continuously decoded through sEMG after a few minutes of training, allowing them to regain some independence in daily life. Although movement is not nearly as fast as natural, unimpaired movement, reach and grasp success rates are near 100% after 50 s of movement.

Abstract. We investigate if a deep Convolutional Neural Network can learn representations of local image patches that are usable in the important task of keypoint matching. We examine several possible loss functions for this correspondance task and show emprically that a newly suggested loss formulation allows a Convolutional Neural Network to find compact local image descriptors that perform comparably to stateof-the-art approaches.

We investigate the relation between grip force and grip stiffness for the human hand with and without voluntary cocontraction. Apart from gaining biomechanical insight, this issue is particularly relevant for variable-stiffness robotic systems, which can independently control the two parameters, but for which no clear methods exist to design or efficiently exploit them. Subjects were asked in one task to produce different levels of force, and stiffness was measured. As expected, this task reveals a linear coupling between force and stiffness. In a second task, subjects were then asked to additionally decouple stiffness from force at these force levels by using cocontraction. We measured the electromyogram from relevant groups of muscles and analyzed the possibility to predict stiffness and force. Optical tracking was used for avoiding wrist movements. We found that subjects were able to decouple grip stiffness from force when using cocontraction on average by about 20% of the maximum measured stiffness over all force levels, while this ability increased with the applied force. This result contradicts the force–stiffness behavior of most variable-stiffness actuators. Moreover, we found the thumb to be on average twice as stiff as the index finger and discovered that intrinsic hand muscles predominate our prediction of stiffness, but not of force. EMG activity and grip force allowed to explain 72 ± 12% of the measured variance in stiffness by simple linear regression, while only 33 ± 18% variance in force. Conclusively the high signal-to-noise ratio and the high correlation to stiffness of these muscles allow for a robust and reliable regression of stiffness, which can be used to continuously teleoperate compliance of modern robotic hands.

Recurrent Neural Networks (RNNs) are rich models for the processing of sequential data. Recent work on advancing the state of the art has been focused on the optimization or modelling of RNNs, mostly motivated by adressing the problems of the vanishing and exploding gradients. The control of overfitting has seen considerably less attention. This paper contributes to that by analyzing fast dropout, a recent regularization method for generalized linear models and neural networks from a back-propagation inspired perspective. We show that fast dropout implements a quadratic form of an adaptive, per-parameter regularizer, which rewards large weights in the light of underfitting, penalizes them for overconfident predictions and vanishes at minima of an unregularized training loss. The derivatives of that regularizer are exclusively based on the training error signal. One consequence of this is the absence of a global weight attractor, which is particularly appealing for RNNs, since the dynamics are not biased towards a certain regime. We positively test the hypothesis that this improves the performance of RNNs on four musical data sets.