In this paper we present an investigation of sequence-discriminative training of deep neural networks for automatic speech recognition. We evaluate different sequence-discriminative training criteria (MMI and MPE) and...
详细信息
ISBN:
(纸本)9781467369985
In this paper we present an investigation of sequence-discriminative training of deep neural networks for automatic speech recognition. We evaluate different sequence-discriminative training criteria (MMI and MPE) and optimization algorithms (including SGD and Rprop) using the RASR toolkit. Further, we compare the training of the whole network with that of the output layer only. Technical details necessary for a robust training are studied, since there is no consensus yet on the ultimate training recipe. The investigation extends our previous work on training linear bottleneck networks from scratch showing the consistently positive effect of sequence training.
The rate of convergence of the classical Thresholding Greedy Algorithm with respect to bases is studied in this paper. We bound the error of approximation by the product of both norms - the norm of f and the A1-norm o...
详细信息
The paper presents some experimental measurements carried out in an industrial scale prototype of Permanent Magnet Heater, PMH. The heater has been designed for the induction heating of aluminum billets with high effi...
详细信息
ISBN:
(纸本)9781479917631
The paper presents some experimental measurements carried out in an industrial scale prototype of Permanent Magnet Heater, PMH. The heater has been designed for the induction heating of aluminum billets with high efficiency. The experimental results are compared with the quantities calculated for the optimal design of the system and presented in a previous paper by the same authors. The design was carried out by means of transient magnetic and thermal 2D Finite Element Models coupled with multiobjective optimization algorithms. Experimental results have demonstrated the validity of the proposed design approach as well as the limits of 2D models in predicting the real performances of the device.
This paper studies the convergence of clipped stochastic gradient descent (SGD) algorithms with decision-dependent data distribution. Our setting is motivated by privacy preserving optimization algorithms that interac...
详细信息
One of the main purposes of Metabolic Engineering is the quantitative prediction of cell behaviour under selected genetic modifications. These methods can then be used to support adequate strain optimization algorithm...
详细信息
ISBN:
(纸本)9781479904532
One of the main purposes of Metabolic Engineering is the quantitative prediction of cell behaviour under selected genetic modifications. These methods can then be used to support adequate strain optimization algorithms in a outer layer. The purpose of the present study is to explore methods in which dynamical models provide for phenotype simulation methods, that will be used as a basis for strain optimization algorithms to indicate enzyme under/over expression or deletion of a few reactions as to maximize the production of compounds with industrial interest. This work details the developed optimization algorithms, based on Evolutionary Computation approaches, to enhance the production of a target metabolite by finding an adequate set of reaction deletions or by changing the levels of expression of a set of enzymes. To properly evaluate the strains, the ratio of the flux value associated with the target metabolite divided by the wild-type counterpart was employed as a fitness function. The devised algorithms were applied to the maximization of Serine production by Escherichia coli, using a dynamic kinetic model of the central carbon metabolism. In this case study, the proposed algorithms reached a set of solutions with higher quality, as compared to the ones described in the literature using distinct optimization techniques.
Future smart grid control demands delegation of liabilities to distributed, rather small energy resources in contrast to today's traditional large control power units. Distributed energy scheduling constitutes a c...
详细信息
ISBN:
(纸本)9781538623718
Future smart grid control demands delegation of liabilities to distributed, rather small energy resources in contrast to today's traditional large control power units. Distributed energy scheduling constitutes a complex task for optimization algorithms regarding the underlying high-dimensional, multimodal and non-linear problem structure. For predictive scheduling with high penetration of renewable energy resources, agent-based approaches using classifier-based decoders for modeling individual flexibilities have shown good performance. On the other hand, such decoder-based methods are currently designed for single entities and not able to cope with ensembles of energy resources. Aggregating training sets sampled from individually modeled energy units results in folded distributions with unfavorable properties for training a decoder. Nevertheless, this happens to be a quite frequent use case, e.g. when a hotel, a small business, a school or similar with an ensemble of co-generation, heat pump, solar power, and controllable consumers wants to take part in decentralized predictive scheduling. Recently, an extension to an established agent approach for scheduling individual single energy units has been proposed that is based on second level optimization. The agents' decision routine may be enhanced by a covariance matrix adaption evolution strategy that is hybridized with decoders. In this way, locally managed ensembles of energy units can be included. The applicability has already been demonstrated, but the effects of ensemble composition are so far unknown. Here, we give an widened view on the underlying power level distribution problem and extend the results by conducting a sensitivity analysis on the impact of ensemble size and penetration on communication overhead and residual error.
Learner Performance-based Behavior using Simulated Annealing (LPBSA) is an improvement of the Learner Performance-based Behavior (LPB) algorithm. LPBSA, like LPB, has been proven to deal with single and complex proble...
详细信息
Centralized architectures with fronthauls can be used to deal with some of the problems inherently associated with dense small cell deployments. This study examines a joint cell assignment and scheduling solution for ...
详细信息
ISBN:
(纸本)9781509016990
Centralized architectures with fronthauls can be used to deal with some of the problems inherently associated with dense small cell deployments. This study examines a joint cell assignment and scheduling solution for the downlink to increase the users' data rates, based on cell switching and a suboptimal optimization algorithm that nearly achieves the performance of the optimal Hungarian assignment. Moreover, an exhaustive sensitivity analysis with different network and traffic configurations is carried out in order to understand what conditions are more appropriate for the use of the proposed mechanism and solutions involving cell switching in general. Simulation results show that such solutions can greatly benefit from the use of receivers with interference suppression capabilities and a larger number of antennas, with a maximum data rate gain of 120%. High performance gains are observed with two different traffic models, and it is not necessary to be able to connect to a large number of cells in order to reap most of the benefits of the centralized dynamic cell selection.
This paper studies a non-stationary kernelized bandit (KB) problem, also called time-varying Bayesian optimization, where one seeks to minimize the regret under an unknown reward function that varies over time. In par...
详细信息
Deformable Image Registration is a complex optimization algorithm with the goal of modeling a non-rigid transformation between two images. A crucial issue in this field is guaranteeing the user a robust but computatio...
详细信息
ISBN:
(纸本)9781424441211
Deformable Image Registration is a complex optimization algorithm with the goal of modeling a non-rigid transformation between two images. A crucial issue in this field is guaranteeing the user a robust but computationally reasonable algorithm. We rank the performances of four stopping criteria and six stopping value computation strategies for a log domain deformable registration. The stopping criteria we test are: (a) velocity field update magnitude, (b) vector field Jacobian, (c) mean squared error, and (d) harmonic energy. Experiments demonstrate that comparing the metric value over the last three iterations with the metric minimum of between four and six previous iterations is a robust and appropriate strategy. The harmonic energy and vector field update magnitude metrics give the best results in terms of robustness and speed of convergence.
暂无评论