Nowadays, it is growing interest to make Machine Learning (ML) systems more understandable and trusting to general users. Thus, generating explanations for ML system behaviours that are understandable to human beings is a central scientific and technological issue addressed by the rapidly growing research area of eXplainable Artificial Intelligence (XAI). Recently, it is becoming more and more evident that new directions to create better explanations should take into account what a good explanation is to a human user, and consequently, develop XAI solutions able to provide user-centred explanations. This paper suggests taking advantage of developing an XAI general approach that allows producing explanations for an ML system behaviour in terms of different and user-selected input features, i.e., explanations composed of input properties that the human user can select according to his background knowledge and goals. To this end, we propose an XAI general approach which is able: 1) to construct explanations in terms of input features that represent more salient and understandable input properties for a user, which we call here Middle-Level input Features (MLFs), 2) to be applied to different types of MLFs. We experimentally tested our approach on two different datasets and using three different types of MLFs. The results seem encouraging.

Background: Keyword extraction is a popular research topic in the field of natural language processing. Keywords are terms that describe the most relevant information in a document. The main problem that researchers are facing is how to efficiently and accurately extract the core keywords from a document. However, previous keyword extraction approaches have utilized the text and graph features, there is the lack of models that can properly learn and combine these features in a best way. Methods: In this paper, we develop a multimodal Key-phrase extraction approach, namely Phraseformer, using transformer and graph embedding techniques. In Phraseformer, each keyword candidate is presented by a vector which is the concatenation of the text and structure learning representations. Phraseformer takes the advantages of recent researches such as BERT and ExEm to preserve both representations. Also, the Phraseformer treats the key-phrase extraction task as a sequence labeling problem solved using classification task. Results: We analyze the performance of Phraseformer on three datasets including Inspec, SemEval2010 and SemEval 2017 by F1-score. Also, we investigate the performance of different classifiers on Phraseformer method over Inspec dataset. Experimental results demonstrate the effectiveness of Phraseformer method over the three datasets used. Additionally, the Random Forest classifier gain the highest F1-score among all classifiers. Conclusions: Due to the fact that the combination of BERT and ExEm is more meaningful and can better represent the semantic of words. Hence, Phraseformer significantly outperforms single-modality methods.

Deep learning promises performant anomaly detection on time-variant datasets, but greatly suffers from low availability of suitable training datasets and frequently changing tasks. Deep transfer learning offers mitigation by letting algorithms built upon previous knowledge from different tasks or locations. In this article, a modular deep learning algorithm for anomaly detection on time series datasets is presented that allows for an easy integration of such transfer learning capabilities. It is thoroughly tested on a dataset from a discrete manufacturing process in order to prove its fundamental adequacy towards deep industrial transfer learning - the transfer of knowledge in industrial applications' special environment.

Among different ML models, Deep Neural Networks (DNNs) [130] are well-known and widely used for their powerful representation learning from high-dimensional data such as images, texts, and speech. However, as ML algorithms enter sensitive real-world domains with trustworthiness, safety, and fairness prerequisites, the need for corresponding techniques and metrics for high-stake domains is more noticeable than before. Hence, researchers in different fields propose guidelines for Trustworthy AI [208], Safe AI [5], and Explainable AI [155] as stepping stones for next generation Responsible AI [6, 247]. Furthermore, government reports and regulations on AI accountability [75], trustworthiness [216], and safety [31] are gradually creating mandating laws to protect citizens' data privacy, fair data processing, and upholding safety for AI-based products. The development and deployment of ML algorithms for open-world tasks come with reliability and dependability limitations rooting from model performance, robustness, and uncertainty limitations [156]. Unlike traditional code-based software, ML models have fundamental safety drawbacks, including performance limitations on their training set and run-time robustness in their operational domain.

Vision transformers (ViTs) have recently received explosive popularity, but their enormous model sizes and training costs remain daunting. Conventional post-training pruning often incurs higher training budgets. In contrast, this paper aims to trim down both the training memory overhead and the inference complexity, without sacrificing the achievable accuracy. We launch and report the first-of-its-kind comprehensive exploration, on taking a unified approach of integrating sparsity in ViTs "from end to end". Specifically, instead of training full ViTs, we dynamically extract and train sparse subnetworks, while sticking to a fixed small parameter budget. Our approach jointly optimizes model parameters and explores connectivity throughout training, ending up with one sparse network as the final output. The approach is seamlessly extended from unstructured to structured sparsity, the latter by considering to guide the prune-and-grow of self-attention heads inside ViTs. For additional efficiency gains, we further co-explore data and architecture sparsity, by plugging in a novel learnable token selector to adaptively determine the currently most vital patches. Extensive results on ImageNet with diverse ViT backbones validate the effectiveness of our proposals which obtain significantly reduced computational cost and almost unimpaired generalization. Perhaps most surprisingly, we find that the proposed sparse (co-)training can even improve the ViT accuracy rather than compromising it, making sparsity a tantalizing "free lunch". For example, our sparsified DeiT-Small at (5%, 50%) sparsity for (data, architecture), improves 0.28% top-1 accuracy, and meanwhile enjoys 49.32% FLOPs and 4.40% running time savings. Our codes are available at https://github.com/VITA-Group/SViTE.

Latent space geometry has shown itself to provide a rich and rigorous framework for interacting with the latent variables of deep generative models. The existing theory, however, relies on the decoder being a Gaussian distribution as its simple reparametrization allows us to interpret the generating process as a random projection of a deterministic manifold. Consequently, this approach breaks down when applied to decoders that are not as easily reparametrized. We here propose to use the Fisher-Rao metric associated with the space of decoder distributions as a reference metric, which we pull back to the latent space. We show that we can achieve meaningful latent geometries for a wide range of decoder distributions for which the previous theory was not applicable, opening the door to `black box' latent geometries.

Attention-based neural networks have achieved state-of-the-art results on a wide range of tasks. Most such models use deterministic attention while stochastic attention is less explored due to the optimization difficulties or complicated model design. This paper introduces Bayesian attention belief networks, which construct a decoder network by modeling unnormalized attention weights with a hierarchy of gamma distributions, and an encoder network by stacking Weibull distributions with a deterministic-upward-stochastic-downward structure to approximate the posterior. The resulting auto-encoding networks can be optimized in a differentiable way with a variational lower bound. It is simple to convert any models with deterministic attention, including pretrained ones, to the proposed Bayesian attention belief networks. On a variety of language understanding tasks, we show that our method outperforms deterministic attention and state-of-the-art stochastic attention in accuracy, uncertainty estimation, generalization across domains, and robustness to adversarial attacks. We further demonstrate the general applicability of our method on neural machine translation and visual question answering, showing great potential of incorporating our method into various attention-related tasks.

In this work we introduce a new approach for identifiable non-linear ICA models. Recently there has been a renaissance in identifiability results in deep generative models, not least for non-linear ICA. These prior works, however, have assumed access to a sufficiently-informative auxiliary set of observations, denoted $\mathbf{u}$. We show here how identifiability can be obtained in the absence of this side-information, rendering possible fully-unsupervised identifiable non-linear ICA. While previous theoretical results have established the impossibility of identifiable non-linear ICA in the presence of infinitely-flexible universal function approximators, here we rely on the intrinsically-finite modelling capacity of any particular chosen parameterisation of a deep generative model. In particular, we focus on generative models which perform clustering in their latent space -- a model structure which matches previous identifiable models, but with the learnt clustering providing a synthetic form of auxiliary information. We evaluate our proposals using VAEs, on synthetic and image datasets, and find that the learned clusterings function effectively: deep generative models with latent clusterings are empirically identifiable, to the same degree as models which rely on side information.

A central challenge in training classification models in the real-world federated system is learning with non-IID data. To cope with this, most of the existing works involve enforcing regularization in local optimization or improving the model aggregation scheme at the server. Other works also share public datasets or synthesized samples to supplement the training of under-represented classes or introduce a certain level of personalization. Though effective, they lack a deep understanding of how the data heterogeneity affects each layer of a deep classification model. In this paper, we bridge this gap by performing an experimental analysis of the representations learned by different layers. Our observations are surprising: (1) there exists a greater bias in the classifier than other layers, and (2) the classification performance can be significantly improved by post-calibrating the classifier after federated training. Motivated by the above findings, we propose a novel and simple algorithm called Classifier Calibration with Virtual Representations (CCVR), which adjusts the classifier using virtual representations sampled from an approximated gaussian mixture model. Experimental results demonstrate that CCVR achieves state-of-the-art performance on popular federated learning benchmarks including CIFAR-10, CIFAR-100, and CINIC-10. We hope that our simple yet effective method can shed some light on the future research of federated learning with non-IID data.

Understanding generalization in deep learning has been one of the major challenges in statistical learning theory over the last decade. While recent work has illustrated that the dataset and the training algorithm must be taken into account in order to obtain meaningful generalization bounds, it is still theoretically not clear which properties of the data and the algorithm determine the generalization performance. In this study, we approach this problem from a dynamical systems theory perspective and represent stochastic optimization algorithms as random iterated function systems (IFS). Well studied in the dynamical systems literature, under mild assumptions, such IFSs can be shown to be ergodic with an invariant measure that is often supported on sets with a fractal structure. As our main contribution, we prove that the generalization error of a stochastic optimization algorithm can be bounded based on the `complexity' of the fractal structure that underlies its invariant measure. Leveraging results from dynamical systems theory, we show that the generalization error can be explicitly linked to the choice of the algorithm (e.g., stochastic gradient descent -- SGD), algorithm hyperparameters (e.g., step-size, batch-size), and the geometry of the problem (e.g., Hessian of the loss). We further specialize our results to specific problems (e.g., linear/logistic regression, one hidden-layered neural networks) and algorithms (e.g., SGD and preconditioned variants), and obtain analytical estimates for our bound.For modern neural networks, we develop an efficient algorithm to compute the developed bound and support our theory with various experiments on neural networks.