Benefiting from convenient cycling and flexible parking locations, the Dockless Public Bicycle-sharing (DL-PBS) network becomes increasingly popular in many countries. However, redundant and low-utility stations waste public urban space and maintenance costs of DL-PBS vendors. In this paper, we propose a Bicycle Station Dynamic Planning (BSDP) system to dynamically provide the optimal bicycle station layout for the DL-PBS network. The BSDP system contains four modules: bicycle drop-off location clustering, bicycle-station graph modeling, bicycle-station location prediction, and bicycle-station layout recommendation. In the bicycle drop-off location clustering module, candidate bicycle stations are clustered from each spatio-temporal subset of the large-scale cycling trajectory records. In the bicycle-station graph modeling module, a weighted digraph model is built based on the clustering results and inferior stations with low station revenue and utility are filtered. Then, graph models across time periods are combined to create a graph sequence model. In the bicycle-station location prediction module, the GGNN model is used to train the graph sequence data and dynamically predict bicycle stations in the next period. In the bicycle-station layout recommendation module, the predicted bicycle stations are fine-tuned according to the government urban management plan, which ensures that the recommended station layout is conducive to city management, vendor revenue, and user convenience. Experiments on actual DL-PBS networks verify the effectiveness, accuracy and feasibility of the proposed BSDP system.

Billions of users create a large amount of data every day, which in a sense comes from various types of sources. This data is in most cases unorganized and unclassified and is presented in various formats such as text, video, audio, or images. Processing and analyzing this data is a major challenge that we face every day. The problem of unstructured and unorganized text dates back to ancient times, but Text Classification as a discipline first appeared in the early 60s, where 30 years later the interest in various spheres for it increased [1], and began to be applied in various types of domains and applications such as for movie review [2], document classification [3], ecommerce [4], social media [5], online courses [6, 7], etc. As interest has grown more in the upcoming years, the uses start solving the problems with higher accurate results in more flexible ways. Knowledge Engineering (KE) was one of the applications of text classification in the late 80s, where the process took place by manually defining rules based on expert knowledge in terms of categorization of the document for a particular category [1]. After this time, there was a great wave of use of various modern and advanced methods for text classification, which all improved this discipline and made it more interesting for scientists and researchers, more specifically the use of machine learning techniques. These techniques bring a lot of advantages, as they are now in very large numbers, where they provide solutions to almost every problem we may encounter. The need for education and learning dates back to ancient times, where people are constantly improving and trying to gain as much knowledge as possible.

Beyond the complexity of CNNs that require training on large annotated datasets, the domain shift between design and operational data has limited the adoption of CNNs in many real-world applications. For instance, in person re-identification, videos are captured over a distributed set of cameras with non-overlapping viewpoints. The shift between the source (e.g. lab setting) and target (e.g. cameras) domains may lead to a significant decline in recognition accuracy. Additionally, state-of-the-art CNNs may not be suitable for such real-time applications given their computational requirements. Although several techniques have recently been proposed to address domain shift problems through unsupervised domain adaptation (UDA), or to accelerate/compress CNNs through knowledge distillation (KD), we seek to simultaneously adapt and compress CNNs to generalize well across multiple target domains. In this paper, we propose a progressive KD approach for unsupervised single-target DA (STDA) and multi-target DA (MTDA) of CNNs. Our method for KD-STDA adapts a CNN to a single target domain by distilling from a larger teacher CNN, trained on both target and source domain data in order to maintain its consistency with a common representation. Our proposed approach is compared against state-of-the-art methods for compression and STDA of CNNs on the Office31 and ImageClef-DA image classification datasets. It is also compared against state-of-the-art methods for MTDA on Digits, Office31, and OfficeHome. In both settings -- KD-STDA and KD-MTDA -- results indicate that our approach can achieve the highest level of accuracy across target domains, while requiring a comparable or lower CNN complexity.

Policy specification is a process by which a human can initialize a robot's behaviour and, in turn, warm-start policy optimization via Reinforcement Learning (RL). While policy specification/design is inherently a collaborative process, modern methods based on Learning from Demonstration or Deep RL lack the model interpretability and accessibility to be classified as such. Current state-of-the-art methods for policy specification rely on black-box models, which are an insufficient means of collaboration for non-expert users: These models provide no means of inspecting policies learnt by the agent and are not focused on creating a usable modality for teaching robot behaviour. In this paper, we propose a novel machine learning framework that enables humans to 1) specify, through natural language, interpretable policies in the form of easy-to-understand decision trees, 2) leverage these policies to warm-start reinforcement learning and 3) outperform baselines that lack our natural language initialization mechanism. We train our approach by collecting a first-of-its-kind corpus mapping free-form natural language policy descriptions to decision tree-based policies. We show that our novel framework translates natural language to decision trees with a 96% and 97% accuracy on a held-out corpus across two domains, respectively. Finally, we validate that policies initialized with natural language commands are able to significantly outperform relevant baselines (p < 0.001) that do not benefit from our natural language-based warm-start technique.

Eye gaze estimation has become increasingly significant in computer vision.In this paper,we systematically study the mainstream of eye gaze estimation methods,propose a novel methodology to estimate eye gaze points and eye gaze directions simultaneously.First,we construct a local sharing network for feature extraction of gaze points and gaze directions estimation,which can reduce network computational parameters and converge quickly;Second,we propose a Multiview Multitask Learning (MTL) framework,for gaze directions,a coplanar constraint is proposed for the left and right eyes,for gaze points,three views data input indirectly introduces eye position information,a cross-view pooling module is designed, propose joint loss which handle both gaze points and gaze directions estimation.Eventually,we collect a dataset to use of gaze points,which have three views to exist public dataset.The experiment show our method is state-of-the-art the current mainstream methods on two indicators of gaze points and gaze directions.

We introduce the first end-to-end Deep Reinforcement Learning based framework for active high frequency trading. We train DRL agents to to trade one unit of Intel Corporation stocks by employing the Proximal Policy Optimization algorithm. The training is performed on three contiguous months of high frequency Limit Order Book data. In order to maximise the signal to noise ratio in the training data, we compose the latter by only selecting training samples with largest price changes. The test is then carried out on the following month of data. Hyperparameters are tuned using the Sequential Model Based Optimization technique. We consider three different state characterizations, which differ in the LOB-based meta-features they include. Agents learn trading strategies able to produce stable positive returns in spite of the highly stochastic and non-stationary environment, which is remarkable itself. Analysing the agents' performances on the test data, we argue that the agents are able to create a dynamic representation of the underlying environment highlighting the occasional regularities present in the data and exploiting them to create long-term profitable trading strategies.

Recent improvements in the predictive quality of natural language processing systems are often dependent on a substantial increase in the number of model parameters. This has led to various attempts of compressing such models, but existing methods have not considered the differences in the predictive power of various model components or in the generalizability of the compressed models. To understand the connection between model compression and out-of-distribution generalization, we define the task of compressing language representation models such that they perform best in a domain adaptation setting. We choose to address this problem from a causal perspective, attempting to estimate the \textit{average treatment effect} (ATE) of a model component, such as a single layer, on the model's predictions. Our proposed ATE-guided Model Compression scheme (AMoC), generates many model candidates, differing by the model components that were removed. Then, we select the best candidate through a stepwise regression model that utilizes the ATE to predict the expected performance on the target domain. AMoC outperforms strong baselines on 46 of 60 domain pairs across two text classification tasks, with an average improvement of more than 3\% in F1 above the strongest baseline.

Behavioural symptoms and urinary tract infections (UTI) are among the most common problems faced by people with dementia. One of the key challenges in the management of these conditions is early detection and timely intervention in order to reduce distress and avoid unplanned hospital admissions. Using in-home sensing technologies and machine learning models for sensor data integration and analysis provides opportunities to detect and predict clinically significant events and changes in health status. We have developed an integrated platform to collect in-home sensor data and performed an observational study to apply machine learning models for agitation and UTI risk analysis. We collected a large dataset from 88 participants with a mean age of 82 and a standard deviation of 6.5 (47 females and 41 males) to evaluate a new deep learning model that utilises attention and rational mechanism. The proposed solution can process a large volume of data over a period of time and extract significant patterns in a time-series data (i.e. attention) and use the extracted features and patterns to train risk analysis models (i.e. rational). The proposed model can explain the predictions by indicating which time-steps and features are used in a long series of time-series data. The model provides a recall of 91\% and precision of 83\% in detecting the risk of agitation and UTIs. This model can be used for early detection of conditions such as UTIs and managing of neuropsychiatric symptoms such as agitation in association with initial treatment and early intervention approaches. In our study we have developed a set of clinical pathways for early interventions using the alerts generated by the proposed model and a clinical monitoring team has been set up to use the platform and respond to the alerts according to the created intervention plans.

With the wide use of deep neural networks (DNN), model interpretability has become a critical concern, since explainable decisions are preferred in high-stake scenarios. Current interpretation techniques mainly focus on the feature attribution perspective, which are limited in indicating why and how particular explanations are related to the prediction. To this end, an intriguing class of explanations, named counterfactuals, has been developed to further explore the "what-if" circumstances for interpretation, and enables the reasoning capability on black-box models. However, generating counterfactuals for raw data instances (i.e., text and image) is still in the early stage due to its challenges on high data dimensionality and unsemantic raw features. In this paper, we design a framework to generate counterfactuals specifically for raw data instances with the proposed Attribute-Informed Perturbation (AIP). By utilizing generative models conditioned with different attributes, counterfactuals with desired labels can be obtained effectively and efficiently. Instead of directly modifying instances in the data space, we iteratively optimize the constructed attribute-informed latent space, where features are more robust and semantic. Experimental results on real-world texts and images demonstrate the effectiveness, sample quality as well as efficiency of our designed framework, and show the superiority over other alternatives. Besides, we also introduce some practical applications based on our framework, indicating its potential beyond the model interpretability aspect.

Visual steel surface defect detection is an essential step in steel sheet manufacturing. Several machine learning-based automated visual inspection (AVI) methods have been studied in recent years. However, most steel manufacturing industries still use manual visual inspection due to training time and inaccuracies involved with AVI methods. Automatic steel defect detection methods could be useful in less expensive and faster quality control and feedback. But preparing the annotated training data for segmentation and classification could be a costly process. In this work, we propose to use the Transfer Learning-based U-Net (TLU-Net) framework for steel surface defect detection. We use a U-Net architecture as the base and explore two kinds of encoders: ResNet and DenseNet. We compare these nets' performance using random initialization and the pre-trained networks trained using the ImageNet data set. The experiments are performed using Severstal data. The results demonstrate that the transfer learning performs 5% (absolute) better than that of the random initialization in defect classification. We found that the transfer learning performs 26% (relative) better than that of the random initialization in defect segmentation. We also found the gain of transfer learning increases as the training data decreases, and the convergence rate with transfer learning is better than that of the random initialization.