As one of the most popular sequence-to-sequence modeling approaches for speech recognition, the RNN-Transducer has achieved evolving performance with more and more sophisticated neural network models of growing size and increasing training epochs. While strong computation resources seem to be the prerequisite of training superior models, we try to overcome it by carefully designing a more efficient training pipeline. In this work, we propose an efficient 3-stage progressive training pipeline to build highly-performing neural transducer models from scratch with very limited computation resources in a reasonable short time period. The effectiveness of each stage is experimentally verified on both Librispeech and Switchboard corpora. The proposed pipeline is able to train transducer models approaching state-of-the-art performance with a single GPU in just 2-3 weeks. Our best conformer transducer achieves 4.1% WER on Librispeech test-other with only 35 epochs of training.
An efficient data-driven prediction strategy for multi-antenna frequency-selective channels must operate based on a small number of pilot symbols. This paper proposes novel channel prediction algorithms that address this goal by integrating transfer and meta-learning with a reduced-rank parametrization of the channel. The proposed methods optimize linear predictors by utilizing data from previous frames, which are generally characterized by distinct propagation characteristics, in order to enable fast training on the time slots of the current frame. The proposed predictors rely on a novel long-short-term decomposition (LSTD) of the linear prediction model that leverages the disaggregation of the channel into long-term space-time signatures and fading amplitudes. We first develop predictors for single-antenna frequency-flat channels based on transfer/meta-learned quadratic regularization. Then, we introduce transfer and meta-learning algorithms for LSTD-based prediction models that build on equilibrium propagation (EP) and alternating least squares (ALS). Numerical results under the 3GPP 5G standard channel model demonstrate the impact of transfer and meta-learning on reducing the number of pilots for channel prediction, as well as the merits of the proposed LSTD parametrization.
This paper approaches the unsupervised learning problem by gradient descent in the space of probability density functions. Our main result shows that along the gradient flow induced by a distribution-dependent ordinary differential equation (ODE), the unknown data distribution emerges as the long-time limit of this flow of densities. That is, one can uncover the data distribution by simulating the distribution-dependent ODE. Intriguingly, we find that the simulation of the ODE is equivalent to the training of generative adversarial networks (GANs). The GAN framework, by definition a non-cooperative game between a generator and a discriminator, can therefore be viewed alternatively as a cooperative game between a navigator and a calibrator (in collaboration to simulate the ODE). At the theoretic level, this new perspective simplifies the analysis of GANs and gives new insight into their performance. To construct a solution to the distribution-dependent ODE, we first show that the associated nonlinear Fokker-Planck equation has a unique weak solution, using the Crandall-Liggett theorem for differential equations in Banach spaces. From this solution to the Fokker-Planck equation, we construct a unique solution to the ODE, relying on Trevisan's superposition principle. The convergence of the induced gradient flow to the data distribution is obtained by analyzing the Fokker-Planck equation.
We propose a simple model of an idealized online cultural market in which $N$ items, endowed with a hidden quality metric, are recommended to users by a ranking algorithm possibly biased by the current items' popularity. Our goal is to better understand the underlying mechanisms of the well-known fact that popularity bias can prevent higher-quality items from becoming more popular than lower-quality items, producing an undesirable misalignment between quality and popularity rankings. We do so under the assumption that users, having limited time/attention, are able to discriminate the best-quality only within a random subset of the items. We discover the existence of a harmful regime in which improper use of popularity can seriously compromise the emergence of quality, and a benign regime in which wise use of popularity, coupled with a small discrimination effort on behalf of users, guarantees the perfect alignment of quality and popularity ranking. Our findings clarify the effects of algorithmic popularity bias on quality outcomes, and may inform the design of more principled mechanisms for techno-social cultural markets.
Semantic segmentation is a well-addressed topic in the computer vision literature, but the design of fast and accurate video processing networks remains challenging. In addition, to run on embedded hardware, computer vision models often have to make compromises on accuracy to run at the required speed, so that a latency/accuracy trade-off is usually at the heart of these real-time systems' design. For the specific case of videos, models have the additional possibility to make use of computations made for previous frames to mitigate the accuracy loss while being real-time. In this work, we propose to tackle the task of fast future video segmentation prediction through the use of convolutional layers with time-dependent channel masking. This technique only updates a chosen subset of the feature maps at each time-step, bringing simultaneously less computation and latency, and allowing the network to leverage previously computed features. We apply this technique to several fast architectures and experimentally confirm its benefits for the future prediction subtask.
In this work, we propose \texttt{TimeGrad}, an autoregressive model for multivariate probabilistic time series forecasting which samples from the data distribution at each time step by estimating its gradient. To this end, we use diffusion probabilistic models, a class of latent variable models closely connected to score matching and energy-based methods. Our model learns gradients by optimizing a variational bound on the data likelihood and at inference time converts white noise into a sample of the distribution of interest through a Markov chain using Langevin sampling. We demonstrate experimentally that the proposed autoregressive denoising diffusion model is the new state-of-the-art multivariate probabilistic forecasting method on real-world data sets with thousands of correlated dimensions. We hope that this method is a useful tool for practitioners and lays the foundation for future research in this area.
The question of how to produce a smooth interpolating curve from a stream of data points is addressed in this paper. To this end, we formalize the concept of real-time interpolator (RTI): a trainable unit that recovers smooth signals that are consistent with the received input samples in an online manner. Specifically, an RTI works under the requirement of producing a function section immediately after a sample is received (zero delay), without changing the reconstructed signal in past time sections. This work formulates the design of spline-based RTIs as a bi-level optimization problem. Their training consists in minimizing the average curvature of the interpolated signals over a set of example sequences. The latter are representative of the nature of the data sequence to be interpolated, allowing to tailor the RTI to a specific signal source. Our overall design allows for different possible schemes. In this work, we present two approaches, namely, the parametrized RTI and the recurrent neural network (RNN)-based RTI, including their architecture and properties. Experimental results show that the two proposed RTIs can be trained in a data-driven fashion to achieve improved performance (in terms of the curvature loss metric) with respect to a myopic-type RTI that only exploits the local information at each time sample, while maintaining smooth, zero-delay, and consistency requirements.
Nosie is an important cause of low quality Optical coherence tomography (OCT) image. The neural network model based on Convolutional neural networks(CNNs) has demonstrated its excellent performance in image denoising. However, OCT image denoising still faces great challenges because many previous neural network algorithms required a large number of labeled data, which might cost much time or is expensive. Besides, these CNN-based algorithms need numerous parameters and good tuning techniques, which is hardware resources consuming. To solved above problems, We proposed a new Cycle-Consistent Generative Adversarial Nets called Dual-Merged Cycle-WGAN for retinal OCT image denoiseing, which has remarkable performance with less unlabeled traning data. Our model consists of two Cycle-GAN networks with imporved generator, descriminator and wasserstein loss to achieve good training stability and better performance. Using image merge technique between two Cycle-GAN networks, our model could obtain more detailed information and hence better training effect. The effectiveness and generality of our proposed network has been proved via ablation experiments and comparative experiments. Compared with other state-of-the-art methods, our unsupervised method obtains best subjective visual effect and higher evaluation objective indicators.
We introduce HyperMorph, a framework that facilitates efficient hyperparameter tuning in learning-based deformable image registration. Classical registration algorithms perform an iterative pair-wise optimization to compute a deformation field that aligns two images. Recent learning-based approaches leverage large image datasets to learn a function that rapidly estimates a deformation for a given image pair. In both strategies, the accuracy of the resulting spatial correspondences is strongly influenced by the choice of certain hyperparameter values. However, an effective hyperparameter search consumes substantial time and human effort as it often involves training multiple models for different fixed hyperparameter values and may lead to suboptimal registration. We propose an amortized hyperparameter learning strategy to alleviate this burden by learning the impact of hyperparameters on deformation fields. We design a meta network, or hypernetwork, that predicts the parameters of a registration network for input hyperparameters, thereby comprising a single model that generates the optimal deformation field corresponding to given hyperparameter values. This strategy enables fast, high-resolution hyperparameter search at test-time, reducing the inefficiency of traditional approaches while increasing flexibility. We also demonstrate additional benefits of HyperMorph, including enhanced robustness to model initialization and the ability to rapidly identify optimal hyperparameter values specific to a dataset, image contrast, task, or even anatomical region, all without the need to retrain models. We make our code publicly available at http://hypermorph.voxelmorph.net.
The classification of time-series data is pivotal for streaming data and comes with many challenges. Although the amount of publicly available datasets increases rapidly, deep neural models are only exploited in a few areas. Traditional methods are still used very often compared to deep neural models. These methods get preferred in safety-critical, financial, or medical fields because of their interpretable results. However, their performance and scale-ability are limited, and finding suitable explanations for time-series classification tasks is challenging due to the concepts hidden in the numerical time-series data. Visualizing complete time-series results in a cognitive overload concerning our perception and leads to confusion. Therefore, we believe that patch-wise processing of the data results in a more interpretable representation. We propose a novel hybrid approach that utilizes deep neural networks and traditional machine learning algorithms to introduce an interpretable and scale-able time-series classification approach. Our method first performs a fine-grained classification for the patches followed by sample level classification.