Deep neural network (DNN) quantization converting floating-point (FP) data in the network to integers (INT) is an effective way to shrink the model size for memory saving and simplify the operations for compute acceleration. Recently, researches on DNN quantization develop from inference to training, laying a foundation for the online training on accelerators. However, existing schemes leaving batch normalization (BN) untouched during training are mostly incomplete quantization that still adopts high precision FP in some parts of the data paths. Currently, there is no solution that can use only low bit-width INT data during the whole training process of large-scale DNNs with acceptable accuracy. In this work, through decomposing all the computation steps in DNNs and fusing three special quantization functions to satisfy the different precision requirements, we propose a unified complete quantization framework termed as "WAGEUBN" to quantize DNNs involving all data paths including W (Weights), A (Activation), G (Gradient), E (Error), U (Update), and BN. Moreover, the Momentum optimizer is also quantized to realize a completely quantized framework. Experiments on ResNet18/34/50 models demonstrate that WAGEUBN can achieve competitive accuracy on ImageNet dataset. For the first time, the study of quantization in large-scale DNNs is advanced to the full 8-bit INT level. In this way, all the operations in the training and inference can be bit-wise operations, pushing towards faster processing speed, decreased memory cost, and higher energy efficiency. Our throughout quantization framework has great potential for future efficient portable devices with online learning ability.

--We study collaborative machine learning (ML) across wireless devices, each with its own local dataset. Offloading these datasets to a cloud or an edge server to implement powerful ML solutions is often not feasible due to latency, bandwidth and privacy constraints. Instead, we consider federated edge learning (FEEL), where the devices share local updates on the model parameters rather than their datasets. We consider a heterogeneous cellular network (HCN), where small cell base stations (SBSs) orchestrate FL among the mobile users (MUs) within their cells, and periodically exchange model updates with the macro base station (MBS) for global consensus. We employ gradient sparsification and periodic averaging to increase the communication efficiency of this hierarchical federated learning (FL) framework. We then show using CIF AR-10 dataset that the proposed hierarchical learning solution can significantly reduce the communication latency without sacrificing the model accuracy. V ast amount of data is generated today by mobile devices, from smart phones to autonomous vehicles, drones, and various Internet-of-things (IoT) devices, such as wearable sensors, smart meters, and surveillance cameras. Machine learning (ML) is key to exploit these massive datasets to make intelligent inferences and predictions. Most ML solutions are centralized; that is, they assume that all the data collected from numerous devices in a distributed manner is available at a central server, where a powerful model is trained on the data.

Automated skin lesion analysis is very crucial in clinical practice, as skin cancer is among the most common human malignancy. Existing approaches with deep learning have achieved remarkable performance on this challenging task, however, heavily relying on large-scale labelled datasets. In this paper, we present a novel active learning framework for cost-effective skin lesion analysis. The goal is to effectively select and utilize much fewer labelled samples, while the network can still achieve state-of-the-art performance. Our sample selection criteria complementarily consider both informativeness and representativeness, derived from decoupled aspects of measuring model certainty and covering sample diversity. To make wise use of the selected samples, we further design a simple yet effective strategy to aggregate intra-class images in pixel space, as a new form of data augmentation. We validate our proposed method on data of ISIC 2017 Skin Lesion Classification Challenge for two tasks. Using only up to 50% of samples, our approach can achieve state-of-the-art performances on both tasks, which are comparable or exceeding the accuracies with full-data training, and outperform other well-known active learning methods by a large margin.

Regularized approaches have been successfully applied to linear system identification in recent years. Many of them model unknown impulse responses exploiting the so called Reproducing Kernel Hilbert spaces (RKHSs) that enjoy the notable property of being in one-to-one correspondence with the class of positive semidefinite kernels. The necessary and sufficient condition for a RKHS to be stable, i.e. to contain only BIBO stable linear dynamic systems, has been known in the literature at least since 2006. However, an open question still persists and concerns the equivalence of such condition with the absolute summability of the kernel. This paper provides a definite answer to this matter by proving that such correspondence does not hold. A counterexample is introduced that illustrates the existence of stable RKHSs that are induced by non-absolutely summable kernels.

We address the problem of tuning word embeddings for specific use cases and domains. We propose a new method that automatically combines multiple domain-specific embeddings, selected from a wide range of pre-trained domain-specific embeddings, to improve their combined expressive power. Our approach relies on two key components: 1) a ranking function, based on a new embedding similarity measure, that selects the most relevant embeddings to use given a domain and 2) a dimensionality reduction method that combines the selected embeddings to produce a more compact and efficient encoding that preserves the expressiveness. We empirically show that our method produces effective domain-specific embeddings that consistently improve the performance of state-of-the-art machine learning algorithms on multiple tasks, compared to generic embeddings trained on large text corpora.

We present a novel class of convolutional neural networks (CNNs) for set functions, i.e., data indexed with the powerset of a finite set. The convolutions are derived as linear, shift-equivariant functions for various notions of shifts on set functions. The framework is fundamentally different from graph convolutions based on the Laplacian, as it provides not one but several basic shifts, one for each element in the ground set. Prototypical experiments with several set function classification tasks on synthetic datasets and on datasets derived from real-world hypergraphs demonstrate the potential of our new powerset CNNs.

In multi-temporal SAR interferometry (MT-InSAR), persistent scatterer (PS) pixels are used to estimate geophysical parameters, essentially deformation. Conventionally, PS pixels are selected on the basis of the estimated noise present in the spatially uncorrelated phase component along with look-angle error in a temporal interferometric stack. In this study, two deep learning architectures, namely convolutional neural network for interferometric semantic segmentation (CNN-ISS) and convolutional long short term memory network for interferometric semantic segmentation (CLSTM-ISS), based on learning spatial and spatio-temporal behaviour respectively, were proposed for selection of PS pixels. These networks were trained to relate the interferometric phase history to its classification into phase stable (PS) and phase unstable (non-PS) measurement pixels using ~10,000 real world interferometric images of different study sites containing man-made objects, forests, vegetation, uncropped land, water bodies, and areas affected by lengthening, foreshortening, layover and shadowing. The networks were trained using training labels obtained from the Stanford method for Persistent Scatterer Interferometry (StaMPS) algorithm. However, pixel selection results, when compared to a combination of R-index and a classified image of the test dataset, reveal that CLSTM-ISS estimates improved the classification of PS and non-PS pixels compared to those of StaMPS and CNN-ISS. The predicted results show that CLSTM-ISS reached an accuracy of 93.50%, higher than that of CNN-ISS (89.21%). CLSTM-ISS also improved the density of reliable PS pixels compared to StaMPS and CNN-ISS and outperformed StaMPS and other conventional MT-InSAR methods in terms of computational efficiency.

Causal discovery witnessed significant progress over the past decades. In particular, many recent causal discovery methods make use of independent, non-Gaussian noise to achieve identifiability of the causal models. Existence of hidden direct common causes, or confounders, generally makes causal discovery more difficult; whenever they are present, the corresponding causal discovery algorithms can be seen as extensions of overcomplete independent component analysis (OICA). However, existing OICA algorithms usually make strong parametric assumptions on the distribution of independent components, which may be violated on real data, leading to sub-optimal or even wrong solutions. In addition, existing OICA algorithms rely on the Expectation Maximization (EM) procedure that requires computationally expensive inference of the posterior distribution of independent components. To tackle these problems, we present a Likelihood-Free Overcomplete ICA algorithm (LFOICA) that estimates the mixing matrix directly by back-propagation without any explicit assumptions on the density function of independent components. Thanks to its computational efficiency, the proposed method makes a number of causal discovery procedures much more practically feasible. For illustrative purposes, we demonstrate the computational efficiency and efficacy of our method in two causal discovery tasks on both synthetic and real data.

Dimensionality reduction on Riemannian manifolds is challenging due to the complex nonlinear data structures. While probabilistic principal geodesic analysis~(PPGA) has been proposed to generalize conventional principal component analysis (PCA) onto manifolds, its effectiveness is limited to data with a single modality. In this paper, we present a novel Gaussian latent variable model that provides a unique way to integrate multiple PGA models into a maximum-likelihood framework. This leads to a well-defined mixture model of probabilistic principal geodesic analysis (MPPGA) on sub-populations, where parameters of the principal subspaces are automatically estimated by employing an Expectation Maximization algorithm. We further develop a mixture Bayesian PGA (MBPGA) model that automatically reduces data dimensionality by suppressing irrelevant principal geodesics. We demonstrate the advantages of our model in the contexts of clustering and statistical shape analysis, using synthetic sphere data, real corpus callosum, and mandible data from human brain magnetic resonance~(MR) and CT images.

We study the multi-channel sparse blind deconvolution (MCS-BD) problem, whose task is to simultaneously recover a kernel $\mathbf a$ and multiple sparse inputs $\{\mathbf x_i\}_{i=1}^p$ from their circulant convolution $\mathbf y_i = \mathbf a \circledast \mathbf x_i $ ($i=1,\cdots,p$). We formulate the task as a nonconvex optimization problem over the sphere. Under mild statistical assumptions of the data, we prove that the vanilla Riemannian gradient descent (RGD) method, with random initializations, provably recovers both the kernel $\mathbf a$ and the signals $\{\mathbf x_i\}_{i=1}^p$ up to a signed shift ambiguity. In comparison with state-of-the-art results, our work shows significant improvements in terms of sample complexity and computational efficiency. Our theoretical results are corroborated by numerical experiments, which demonstrate superior performance of the proposed approach over the previous methods on both synthetic and real datasets.