With a fixed model structure, knowledge distillation and filter grafting are two effective ways to boost single model accuracy. However, the working mechanism and the differences between distillation and grafting have not been fully unveiled. In this paper, we evaluate the effect of distillation and grafting in the filter level, and find that the impacts of the two techniques are surprisingly complementary: distillation mostly enhances the knowledge of valid filters while grafting mostly reactivates invalid filters. This observation guides us to design a unified training framework called DGD, where distillation and grafting are naturally combined to increase the knowledge density inside the filters given a fixed model structure. Through extensive experiments, we show that the knowledge densified network in DGD shares both advantages of distillation and grafting, lifting the model accuracy to a higher level.

While deep learning makes significant achievements in Artificial Intelligence (AI), the lack of transparency has limited its broad application in various vertical domains. Explainability is not only a gateway between AI and real world, but also a powerful feature to detect flaw of the models and bias of the data. Local Interpretable Model-agnostic Explanation (LIME) is a widely-accepted technique that explains the prediction of any classifier faithfully by learning an interpretable model locally around the predicted instance. As an extension of LIME, this paper proposes an high-interpretability and high-fidelity local explanation method, known as Local Explanation using feature Dependency Sampling and Nonlinear Approximation (LEDSNA). Given an instance being explained, LEDSNA enhances interpretability by feature sampling with intrinsic dependency. Besides, LEDSNA improves the local explanation fidelity by approximating nonlinear boundary of local decision. We evaluate our method with classification tasks in both image domain and text domain. Experiments show that LEDSNA's explanation of the back-box model achieves much better performance than original LIME in terms of interpretability and fidelity.

This work presents a Long Short-Term Memory (LSTM) network for forecasting a monthly electricity demand time series with a one-year horizon. The novelty of this work is the use of pattern representation of the seasonal time series as an alternative to decomposition. Pattern representation simplifies the complex nonlinear and nonstationary time series, filtering out the trend and equalizing variance. Two types of patterns are defined: x-pattern and y-pattern. The former requires additional forecasting for the coding variables. The latter determines the coding variables from the process history. A hybrid approach based on x-patterns turned out to be more accurate than the standard LSTM approach based on a raw time series. In this combined approach an x-pattern is forecasted using a sequence-to-sequence LSTM network and the coding variables are forecasted using exponential smoothing. A simulation study performed on the monthly electricity demand time series for 35 European countries confirmed the high performance of the proposed model and its competitiveness to classical models such as ARIMA and exponential smoothing as well as the MLP neural network model.

We introduce LiPopt, a polynomial optimization framework for computing increasingly tighter upper bounds on the Lipschitz constant of neural networks. The underlying optimization problems boil down to either linear (LP) or semidefinite (SDP) programming. We show how to use the sparse connectivity of a network, to significantly reduce the complexity of computation. This is specially useful for convolutional as well as pruned neural networks. We conduct experiments on networks with random weights as well as networks trained on MNIST, showing that in the particular case of the $\ell_\infty$-Lipschitz constant, our approach yields superior estimates, compared to baselines available in the literature.

In this work, we propose a new generative model that is capable of automatically decoupling global and local representations of images in an entirely unsupervised setting. The proposed model utilizes the variational auto-encoding framework to learn a (low-dimensional) vector of latent variables to capture the global information of an image, which is fed as a conditional input to a flow-based invertible decoder with architecture borrowed from style transfer literature. Experimental results on standard image benchmarks demonstrate the effectiveness of our model in terms of density estimation, image generation and unsupervised representation learning. Importantly, this work demonstrates that with only architectural inductive biases, a generative model with a plain log-likelihood objective is capable of learning decoupled representations, requiring no explicit supervision. The code for our model is available at https://github.com/XuezheMax/wolf.

Selective manipulation of data attributes using deep generative models is an active area of research. In this paper, we present a novel method to structure the latent space of a Variational Auto-Encoder (VAE) to encode different continuous-valued attributes explicitly. This is accomplished by using an attribute regularization loss which enforces a monotonic relationship between the attribute values and the latent code of the dimension along which the attribute is to be encoded. Consequently, post-training, the model can be used to manipulate the attribute by simply changing the latent code of the corresponding regularized dimension. The results obtained from several quantitative and qualitative experiments show that the proposed method leads to disentangled and interpretable latent spaces that can be used to effectively manipulate a wide range of data attributes spanning image and symbolic music domains.

Most recently, with the proliferation of IoT devices, computational nodes in manufacturing systems IIoT(Industrial-Internet-of-things) and the lunch of 5G networks, there will be millions of connected devices generating a massive amount of data. In such an environment, the controlling systems need to be intelligent enough to deal with a vast amount of data to detect defects in a real-time process. Driven by such a need, artificial intelligence models such as deep learning have to be deployed into IIoT systems. However, learning and using deep learning models are computationally expensive, so an IoT device with limited computational power could not run such models. To tackle this issue, edge intelligence had emerged as a new paradigm towards running Artificial Intelligence models on edge devices. Although a considerable amount of studies have been proposed in this area, the research is still in the early stages. In this paper, we propose a novel edge-based multi-phase pruning pipelines to ensemble learning on IIoT devices. In the first phase, we generate a diverse ensemble of pruned models, then we apply integer quantisation, next we prune the generated ensemble using a clustering-based technique. Finally, we choose the best representative from each generated cluster to be deployed to a distributed IoT environment. On CIFAR-100 and CIFAR-10, our proposed approach was able to outperform the predictability levels of a baseline model (up to 7%), more importantly, the generated learners have small sizes (up to 90% reduction in the model size) that minimise the required computational capabilities to make an inference on the resource-constraint devices.

The task of Knowledge Graph Completion (KGC) aims to automatically infer the missing fact information in Knowledge Graph (KG). In this paper, we take a new perspective that aims to leverage rich user-item interaction data (user interaction data for short) for improving the KGC task. Our work is inspired by the observation that many KG entities correspond to online items in application systems. However, the two kinds of data sources have very different intrinsic characteristics, and it is likely to hurt the original performance using simple fusion strategy. To address this challenge, we propose a novel adversarial learning approach by leveraging user interaction data for the KGC task. Our generator is isolated from user interaction data, and serves to improve the performance of the discriminator. The discriminator takes the learned useful information from user interaction data as input, and gradually enhances the evaluation capacity in order to identify the fake samples generated by the generator. To discover implicit entity preference of users, we design an elaborate collaborative learning algorithms based on graph neural networks, which will be jointly optimized with the discriminator. Such an approach is effective to alleviate the issues about data heterogeneity and semantic complexity for the KGC task. Extensive experiments on three real-world datasets have demonstrated the effectiveness of our approach on the KGC task.

This paper proposes a thorough theoretical analysis of Stochastic Gradient Descent (SGD) with decreasing step sizes. First, we show that the recursion defining SGD can be provably approximated by solutions of a time inhomogeneous Stochastic Differential Equation (SDE) in a weak and strong sense. Then, motivated by recent analyses of deterministic and stochastic optimization methods by their continuous counterpart, we study the long-time convergence of the continuous processes at hand and establish non-asymptotic bounds. To that purpose, we develop new comparison techniques which we think are of independent interest. This continuous analysis allows us to develop an intuition on the convergence of SGD and, adapting the technique to the discrete setting, we show that the same results hold to the corresponding sequences. In our analysis, we notably obtain non-asymptotic bounds in the convex setting for SGD under weaker assumptions than the ones considered in previous works. Finally, we also establish finite time convergence results under various conditions, including relaxations of the famous {\L}ojasiewicz inequality, which can be applied to a class of non-convex functions.

High-capacity models require vast amounts of data, and data augmentation is a common remedy when this resource is limited. Standard augmentation techniques apply small hand-tuned transformations to existing data, which is a brittle process that realistically only allows for simple transformations. We propose a Bayesian interpretation of data augmentation where the transformations are modelled as latent variables to be marginalized, and show how these can be inferred variationally in an end-to-end fashion. This allows for significantly more complex transformations than manual tuning, and the marginalization implies a form of test-time data augmentation. The resulting model can be interpreted as a probabilistic extension of spatial transformer networks. Experimentally, we demonstrate improvements in accuracy and uncertainty quantification in image and time series classification tasks.