Search or filter publications

Filter by type:

Filter by publication type

Filter by year:



  • Showing results for:
  • Reset all filters

Search results

    Dutordoir V, Salimbeni HR, Hensman J, Deisenroth MPet al., 2018,

    Gaussian process conditional density estimation

    , Advances in Neural Information Processing Systems, Publisher: Neural Information Processing Systems Conference

    Conditional Density Estimation (CDE) models deal with estimating conditional distributions. The conditions imposed on the distribution are the inputs of the model. CDE is a challenging task as there is a fundamental trade-off between model complexity, representational capacity and overfitting. In this work, we propose to extend the model's input with latent variables and use Gaussian processes (GP) to map this augmented input onto samples from the conditional distribution. Our Bayesian approach allows for the modeling of small datasets, but we also provide the machinery for it to be applied to big data using stochastic variational inference. Our approach can be used to model densities even in sparse data regions, and allows for sharing learned structure between conditions. We illustrate the effectiveness and wide-reaching applicability of our model on a variety of real-world problems, such as spatio-temporal density estimation of taxi drop-offs, non-Gaussian noise modeling, and few-shot learning on omniglot images.

    Wang K, Shah A, Kormushev P, 2018,

    SLIDER: A Bipedal Robot with Knee-less Legs and Vertical Hip Sliding Motion

    , 21st International Conference on Climbing and Walking Robots and Support Technologies for Mobile Machines (CLAWAR 2018)
    Salimbeni HR, Cheng C-A, Boots B, Deisenroth MPet al.,

    Orthogonally decoupled variational Gaussian processes

    , Advances in Neural Information Processing Systems (NIPS) 2018, Publisher: Massachusetts Institute of Technology Press, ISSN: 1049-5258

    Gaussian processes (GPs) provide a powerful non-parametric framework for rea-soning over functions. Despite appealing theory, its superlinear computational andmemory complexities have presented a long-standing challenge. State-of-the-artsparse variational inference methods trade modeling accuracy against complexity.However, the complexities of these methods still scale superlinearly in the numberof basis functions, implying that that sparse GP methods are able to learn fromlarge datasets only when a small model is used. Recently, a decoupled approachwas proposed that removes the unnecessary coupling between the complexitiesof modeling the mean and the covariance functions of a GP. It achieves a linearcomplexity in the number of mean parameters, so an expressive posterior meanfunction can be modeled. While promising, this approach suffers from optimizationdifficulties due to ill-conditioning and non-convexity. In this work, we propose analternative decoupled parametrization. It adopts an orthogonal basis in the meanfunction to model the residues that cannot be learned by the standard coupled ap-proach. Therefore, our method extends, rather than replaces, the coupled approachto achieve strictly better performance. This construction admits a straightforwardnatural gradient update rule, so the structure of the information manifold that islost during decoupling can be leveraged to speed up learning. Empirically, ouralgorithm demonstrates significantly faster convergence in multiple experiments.

    Wilson J, Hutter F, Deisenroth MP,

    Maximizing acquisition functions for Bayesian optimization

    , Advances in Neural Information Processing Systems (NIPS) 2018, Publisher: Massachusetts Institute of Technology Press, ISSN: 1049-5258

    Bayesian optimization is a sample-efficient approach to global optimization that relies on theoretically motivated value heuristics (acquisition functions) to guide its search process. Fully maximizing acquisition functions produces the Bayes' decision rule, but this ideal is difficult to achieve since these functions are frequently non-trivial to optimize. This statement is especially true when evaluating queries in parallel, where acquisition functions are routinely non-convex, high-dimensional, and intractable. We first show that acquisition functions estimated via Monte Carlo integration are consistently amenable to gradient-based optimization. Subsequently, we identify a common family of acquisition functions, including EI and UCB, whose characteristics not only facilitate but justify use of greedy approaches for their maximization.

    Creswell A, Bharath AA, 2018,

    Denoising adversarial autoencoders

    , IEEE Transactions on Neural Networks and Learning Systems, ISSN: 2162-2388

    Unsupervised learning is of growing interest becauseit unlocks the potential held in vast amounts of unlabelled data tolearn useful representations for inference. Autoencoders, a formof generative model, may be trained by learning to reconstructunlabelled input data from a latent representation space. Morerobust representations may be produced by an autoencoderif it learns to recover clean input samples from corruptedones. Representations may be further improved by introducingregularisation during training to shape the distribution of theencoded data in the latent space. We suggestdenoising adversarialautoencoders, which combine denoising and regularisation, shap-ing the distribution of latent space using adversarial training.We introduce a novel analysis that shows how denoising maybe incorporated into the training and sampling of adversarialautoencoders. Experiments are performed to assess the contri-butions that denoising makes to the learning of representationsfor classification and sample synthesis. Our results suggest thatautoencoders trained using a denoising criterion achieve higherclassification performance, and can synthesise samples that aremore consistent with the input data than those trained withouta corruption process.

    Sæmundsson S, Hofmann K, Deisenroth MP, 2018,

    Meta reinforcement learning with latent variable Gaussian processes

    , Uncertainty in Artificial Intelligence (UAI) 2018, Publisher: Association for Uncertainty in Artificial Intelligence (AUAI)

    Learning from small data sets is critical inmany practical applications where data col-lection is time consuming or expensive, e.g.,robotics, animal experiments or drug design.Meta learning is one way to increase the dataefficiency of learning algorithms by general-izing learned concepts from a set of trainingtasks to unseen, but related, tasks. Often, thisrelationship between tasks is hard coded or re-lies in some other way on human expertise.In this paper, we frame meta learning as a hi-erarchical latent variable model and infer therelationship between tasks automatically fromdata. We apply our framework in a model-based reinforcement learning setting and showthat our meta-learning model effectively gen-eralizes to novel tasks by identifying how newtasks relate to prior ones from minimal data.This results in up to a60%reduction in theaverage interaction time needed to solve taskscompared to strong baselines.

    Saputra RP, Kormushev P, 2018,

    Casualty detection for mobile rescue robots via ground-projected point clouds

    , Towards Autonomous Robotic Systems (TAROS) 2018, Publisher: Springer, Cham, Pages: 473-475, ISSN: 0302-9743

    In order to operate autonomously, mobile rescue robots needto be able to detect human casualties in disaster situations. In this paper,we propose a novel method for autonomous detection of casualties lyingdown on the ground based on point-cloud data. This data can be obtainedfrom different sensors, such as an RGB-D camera or a 3D LIDAR sensor.The method is based on a ground-projected point-cloud (GPPC) imageto achieve human body shape detection. A preliminary experiment hasbeen conducted using the RANSAC method for floor detection and, theHOG feature and the SVM classifier to detect human body shape. Theresults show that the proposed method succeeds to identify a casualtyfrom point-cloud data in a wide range of viewing angles.

    Pardo F, Tavakoli A, Levdik V, Kormushev Pet al., 2018,

    Time limits in reinforcement learning

    , International Conference on Machine Learning, Pages: 4042-4051

    In reinforcement learning, it is common to let anagent interact for a fixed amount of time with itsenvironment before resetting it and repeating theprocess in a series of episodes. The task that theagent has to learn can either be to maximize itsperformance over (i) that fixed period, or (ii) anindefinite period where time limits are only usedduring training to diversify experience. In thispaper, we provide a formal account for how timelimits could effectively be handled in each of thetwo cases and explain why not doing so can causestate-aliasing and invalidation of experience re-play, leading to suboptimal policies and traininginstability. In case (i), we argue that the termi-nations due to time limits are in fact part of theenvironment, and thus a notion of the remainingtime should be included as part of the agent’s in-put to avoid violation of the Markov property. Incase (ii), the time limits are not part of the envi-ronment and are only used to facilitate learning.We argue that this insight should be incorporatedby bootstrapping from the value of the state atthe end of each partial episode. For both cases,we illustrate empirically the significance of ourconsiderations in improving the performance andstability of existing reinforcement learning algo-rithms, showing state-of-the-art results on severalcontrol tasks.

    Olofsson S, Deisenroth M, Misener R, 2018,

    Design of experiments for model discrimination hybridising analytical and data-driven approaches

    , 35th International Conference on Machine Learning (ICML), Publisher: ICML

    Healthcare companies must submit pharmaceuti-cal drugs or medical devices to regulatory bodiesbefore marketing new technology. Regulatorybodies frequently require transparent and inter-pretable computational modelling to justify a newhealthcare technology, but researchers may haveseveral competing models for a biological sys-tem and too little data to discriminate betweenthe models. In design of experiments for modeldiscrimination, the goal is to design maximallyinformative physical experiments in order to dis-criminate between rival predictive models. Priorwork has focused either on analytical approaches,which cannot manage all functions, or on data-driven approaches, which may have computa-tional difficulties or lack interpretable marginalpredictive distributions. We develop a method-ology introducing Gaussian process surrogatesin lieu of the original mechanistic models. Wethereby extend existing design and model discrim-ination methods developed for analytical modelsto cases of non-analytical models in a computa-tionally efficient manner.

    Olofsson S, Deisenroth MP, Misener R, 2018,

    Design of Experiments for Model Discrimination using Gaussian Process Surrogate Models

    , Computer Aided Chemical Engineering, Vol: 44, Pages: 847-852, ISSN: 1570-7946

    © 2018 Elsevier B.V. Given rival mathematical models and an initial experimental data set, optimal design of experiments for model discrimination discards inaccurate models. Model discrimination is fundamentally about finding out how systems work. Not knowing how a particular system works, or having several rivalling models to predict the behaviour of the system, makes controlling and optimising the system more difficult. The most common way to perform model discrimination is by maximising the pairwise squared difference between model predictions, weighted by measurement noise and model uncertainty resulting from uncertainty in the fitted model parameters. The model uncertainty for analytical model functions is computed using gradient information. We develop a novel method where we replace the black-box models with Gaussian process surrogate models. Using the surrogate models, we are able to approximately marginalise out the model parameters, yielding the model uncertainty. Results show the surrogate model method working for model discrimination for classical test instances.

This data is extracted from the Web of Science and reproduced under a licence from Thomson Reuters. You may not copy or re-distribute this data in whole or in part without the written consent of the Science business of Thomson Reuters.

Request URL: Request URI: /respub/WEB-INF/jsp/search-t4-html.jsp Query String: id=954&limit=10&respub-action=search.html Current Millis: 1560819275150 Current Time: Tue Jun 18 01:54:35 BST 2019