In this work, we propose FedSSO, a server-side second-order optimization method for federated learning (FL). In contrast to previous works in this direction, we employ a server-side approximation for the Quasi-Newton method without requiring any training data from the clients. In this way, we not only shift the computation burden from clients to server, but also eliminate the additional communication for second-order updates between clients and server entirely. We provide theoretical guarantee for convergence of our novel method, and empirically demonstrate our fast convergence and communication savings in both convex and non-convex settings.

Breakthroughs in transformer-based models have revolutionized not only the NLP field, but also vision and multimodal systems. However, although visualization and interpretability tools have become available for NLP models, internal mechanisms of vision and multimodal transformers remain largely opaque. With the success of these transformers, it is increasingly critical to understand their inner workings, as unraveling these black-boxes will lead to more capable and trustworthy models. To contribute to this quest, we propose VL-InterpreT, which provides novel interactive visualizations for interpreting the attentions and hidden representations in multimodal transformers. VL-InterpreT is a task agnostic and integrated tool that (1) tracks a variety of statistics in attention heads throughout all layers for both vision and language components, (2) visualizes cross-modal and intra-modal attentions through easily readable heatmaps, and (3) plots the hidden representations of vision and language tokens as they pass through the transformer layers. In this paper, we demonstrate the functionalities of VL-InterpreT through the analysis of KD-VLP, an end-to-end pretraining vision-language multimodal transformer-based model, in the tasks of Visual Commonsense Reasoning (VCR) and WebQA, two visual question answering benchmarks. Furthermore, we also present a few interesting findings about multimodal transformer behaviors that were learned through our tool.

Although most cultures, particularly in the west, provide a great many commemorations of distant ancestors – statues, portraits, buildings – we are much less willing to consider our far-off descendants. We might invoke grandchildren, at a push great-grandchildren, but after that, it all becomes a bit vague and, well, unimaginable. And while we look with awe and fascination at the Egyptian pyramids, built 5,000 years ago, we seem incapable of thinking, or even contemplating, 5,000 years in the future. That lies in the realm of science fiction, which is tantamount to fantasy. But the chances are, barring a global catastrophe, humanity will still be very much around in 5,000 years, and going by the average existence of mammal species, should still be thriving in 500,000 years. If we play our cards right, we could even be here in 5m or 500m years, which means that there may be thousands or even millions times more human beings to come than have already existed.

The robustness of deep neural networks is crucial to modern AI-enabled systems and should be formally verified. Sigmoid-like neural networks have been adopted in a wide range of applications. Due to their non-linearity, Sigmoid-like activation functions are usually over-approximated for efficient verification, which inevitably introduces imprecision. Considerable efforts have been devoted to finding the so-called tighter approximations to obtain more precise verification results. However, existing tightness definitions are heuristic and lack theoretical foundations. We conduct a thorough empirical analysis of existing neuron-wise characterizations of tightness and reveal that they are superior only on specific neural networks. We then introduce the notion of network-wise tightness as a unified tightness definition and show that computing network-wise tightness is a complex non-convex optimization problem. We bypass the complexity from different perspectives via two efficient, provably tightest approximations. The results demonstrate the promising performance achievement of our approaches over state of the art: (i) achieving up to 251.28% improvement to certified lower robustness bounds; and (ii) exhibiting notably more precise verification results on convolutional networks.

The data revolution has generated a huge demand for data-driven solutions. This demand propels a growing number of easy-to-use tools and training for aspiring data scientists that enable the rapid building of predictive models. Today, weapons of math destruction can be easily built and deployed without detailed planning and validation. This rapidly extends the list of AI failures, i.e. deployments that lead to financial losses or even violate democratic values such as equality, freedom and justice. The lack of planning, rules and standards around the model development leads to the ,,anarchisation of AI". This problem is reported under different names such as validation debt, reproducibility crisis, and lack of explainability. Post-mortem analysis of AI failures often reveals mistakes made in the early phase of model development or data acquisition. Thus, instead of curing the consequences of deploying harmful models, we shall prevent them as early as possible by putting more attention to the initial planning stage. In this paper, we propose a quick and simple framework to support planning of AI solutions. The POCA framework is based on four pillars: Performance, Opaqueness, Consequences, and Assumptions. It helps to set the expectations and plan the constraints for the AI solution before any model is built and any data is collected. With the help of the POCA method, preliminary requirements can be defined for the model-building process, so that costly model misspecification errors can be identified as soon as possible or even avoided. AI researchers, product owners and business analysts can use this framework in the initial stages of building AI solutions.

Atmospheric pollution remains one of the major public health threat worldwide with an estimated 7 millions deaths annually. In Africa, rapid urbanization and poor transport infrastructure are worsening the problem. In this paper, we have analysed spatio-temporal variations of PM2.5 across different geographical regions in Africa. The West African region remains the most affected by the high levels of pollution with a daily average of 40.856 $\mu g/m^3$ in some cities like Lagos, Abuja and Bamako. In East Africa, Uganda is reporting the highest pollution level with a daily average concentration of 56.14 $\mu g/m^3$ and 38.65 $\mu g/m^3$ for Kigali. In countries located in the central region of Africa, the highest daily average concentration of PM2.5 of 90.075 $\mu g/m^3$ was recorded in N'Djamena. We compare three data driven models in predicting future trends of pollution levels. Neural network is outperforming Gaussian processes and ARIMA models.

Graph pre-training strategies have been attracting a surge of attention in the graph mining community, due to their flexibility in parameterizing graph neural networks (GNNs) without any label information. The key idea lies in encoding valuable information into the backbone GNNs, by predicting the masked graph signals extracted from the input graphs. In order to balance the importance of diverse graph signals (e.g., nodes, edges, subgraphs), the existing approaches are mostly hand-engineered by introducing hyperparameters to re-weight the importance of graph signals. However, human interventions with sub-optimal hyperparameters often inject additional bias and deteriorate the generalization performance in the downstream applications. This paper addresses these limitations from a new perspective, i.e., deriving curriculum for pre-training GNNs. We propose an end-to-end model named MentorGNN that aims to supervise the pre-training process of GNNs across graphs with diverse structures and disparate feature spaces. To comprehend heterogeneous graph signals at different granularities, we propose a curriculum learning paradigm that automatically re-weighs graph signals in order to ensure a good generalization in the target domain. Moreover, we shed new light on the problem of domain adaption on relational data (i.e., graphs) by deriving a natural and interpretable upper bound on the generalization error of the pre-trained GNNs. Extensive experiments on a wealth of real graphs validate and verify the performance of MentorGNN.

Recent advancements in AI, especially deep learning, have contributed to a significant increase in the creation of new realistic-looking synthetic media (video, image, and audio) and manipulation of existing media, which has led to the creation of the new term ``deepfake''. Based on both the research literature and resources in English and in Chinese, this paper gives a comprehensive overview of deepfake, covering multiple important aspects of this emerging concept, including 1) different definitions, 2) commonly used performance metrics and standards, and 3) deepfake-related datasets, challenges, competitions and benchmarks. In addition, the paper also reports a meta-review of 12 selected deepfake-related survey papers published in 2020 and 2021, focusing not only on the mentioned aspects, but also on the analysis of key challenges and recommendations. We believe that this paper is the most comprehensive review of deepfake in terms of aspects covered, and the first one covering both the English and Chinese literature and sources.

Picking up transparent objects is still a challenging task for robots. The visual properties of transparent objects such as reflection and refraction make the current grasping methods that rely on camera sensing fail to detect and localise them. However, humans can handle the transparent object well by first observing its coarse profile and then poking an area of interest to get a fine profile for grasping. Inspired by this, we propose a novel framework of vision-guided tactile poking for transparent objects grasping. In the proposed framework, a segmentation network is first used to predict the horizontal upper regions named as poking regions, where the robot can poke the object to obtain a good tactile reading while leading to minimal disturbance to the object's state. A poke is then performed with a high-resolution GelSight tactile sensor. Given the local profiles improved with the tactile reading, a heuristic grasp is planned for grasping the transparent object. To mitigate the limitations of real-world data collection and labelling for transparent objects, a large-scale realistic synthetic dataset was constructed. Extensive experiments demonstrate that our proposed segmentation network can predict the potential poking region with a high mean Average Precision (mAP) of 0.360, and the vision-guided tactile poking can enhance the grasping success rate significantly from 38.9% to 85.2%. Thanks to its simplicity, our proposed approach could also be adopted by other force or tactile sensors and could be used for grasping of other challenging objects. All the materials used in this paper are available at https://sites.google.com/view/tactilepoking.

The ability to continuously expand knowledge over time and utilize it to rapidly generalize to new tasks is a key feature of human linguistic intelligence. Existing models that pursue rapid generalization to new tasks (e.g., few-shot learning methods), however, are mostly trained in a single shot on fixed datasets, unable to dynamically expand their knowledge; while continual learning algorithms are not specifically designed for rapid generalization. We present a new learning setup, Continual Learning of Few-Shot Learners (CLIF), to address the challenges of both learning settings in a unified setup. CLIF assumes a model learns from a sequence of diverse NLP tasks arriving sequentially, accumulating knowledge for improved generalization to new tasks, while also retaining performance on the tasks learned earlier. We examine how the generalization ability is affected in the continual learning setup, evaluate a number of continual learning algorithms, and propose a novel regularized adapter generation approach. We find that catastrophic forgetting affects generalization ability to a less degree than performance on seen tasks; while continual learning algorithms can still bring considerable benefit to the generalization ability.