Search

Continual model learning for nonlinear dynamic systems, such as autonomous robots, presents several challenges. First, it tends to be computationally expensive as the amount of data collected by the robot quickly grows in time. Second, the model accuracy is impaired when data from repetitive motions prevail in the training set and outweigh scarcer samples that also capture interesting properties of the system. It is not known in advance which samples will be useful for model learning. Therefore, effective methods need to be employed to select informative training samples from the continuous data stream collected by the robot. Existing literature does not give any guidelines as to which of the available sample-selection methods are suitable for such a task. In this paper, we compare five sample-selection methods, including a novel method using the model prediction error. We integrate these methods into a model learning framework based on symbolic regression, which allows for learning accurate models in the form of analytic equations. Unlike the currently popular data-hungry deep learning methods, symbolic regression is able to build models even from very small training data sets. We demonstrate the approach on two real robots: the TurtleBot mobile robot and the Parrot Bebop drone. The results show that an accurate model can be constructed even from training sets as small as 24 samples. Informed sample-selection techniques based on prediction error and model variance clearly outperform uninformed methods, such as sequential or random selection.

A novel method for estimating team tactics in soccer videos based on a Deep Extreme Learning Machine (DELM) and unique characteristics of tactics is presented in this paper. The proposed method estimates the tactics of each team from players' formations and enables successful training from a limited amount of training data. Specifically, the estimation of tactics consists of two stages. First, by utilizing two DELMs corresponding to the two teams, the proposed method estimates the provisional tactics of each team. Second, the proposed method updates the team tactics based on unique characteristics of soccer tactics, the relationship between tactics of the two teams and information on ball possession. Consequently, since the proposed method estimates the team tactics that satisfy these characteristics, accurate estimation results can be obtained. In an experiment, the proposed method is applied to actual soccer videos to verify its effectiveness.

To detect contaminants accidentally included in packaged foods, food industries use an array of systems ranging from metal detectors to X-ray imagers. Low density plastic or glass contaminants, however, are not easily detected with standard methods. If the dielectric contrast between the packaged food and these contaminants in the microwave spectrum is sensible, Microwave Sensing (MWS) can be used as a contactless detection method, which is particularly useful when the food is already packaged. In this paper we propose using MWS combined with Machine Learning (ML). In particular, we report on experiments we did with packaged cocoa-hazelnut spread and show the accuracy of our approach. We also present an FPGA acceleration that runs the ML processing in real-time so as to keep up with the throughput of a production line.

Recently, Capsule Networks (CapsNets) have shown improved performance compared to the traditional Convolutional Neural Networks (CNNs), by encoding and preserving spatial relationships between the detected features in a better way. This is achieved through the so-called Capsules (i.e., groups of neurons) that encode both the instantiation probability and the spatial information. However, one of the major hurdles in the wide adoption of CapsNets is their gigantic training time, which is primarily due to the relatively higher complexity of their new constituting elements that are different from CNNs. In this paper, we implement different optimizations in the training loop of the CapsNets, and investigate how these optimizations affect their training speed and the accuracy. Towards this, we propose a novel framework FasTrCaps that integrates multiple lightweight optimizations and a novel learning rate policy called WarmAdaBatch (that jointly performs warm restarts and adaptive batch size), and steers them in an appropriate way to provide high training-loop speedup at minimal accuracy loss. We also propose weight sharing for capsule layers. The goal is to reduce the hardware requirements of CapsNets by removing unused/redundant connections and capsules, while keeping high accuracy through tests of different learning rate policies and batch sizes. We demonstrate that one of the solutions generated by the FasTrCaps framework can achieve 58.6% reduction in the training time, while preserving the accuracy (even 0.12% accuracy improvement for the MNIST dataset), compared to the CapsNet by Google Brain. The Pareto-optimal solutions generated by FasTrCaps can be leveraged to realize trade-offs between training time and achieved accuracy. We have open-sourced our framework on https://github.com/Alexei95/FasTrCaps., Comment: Accepted for publication at the 2020 International Joint Conference on Neural Networks (IJCNN)

Machine learning (ML) applications are increasingly prevalent. Protecting the confidentiality of ML models becomes paramount for two reasons: (a) a model can be a business advantage to its owner, and (b) an adversary may use a stolen model to find transferable adversarial examples that can evade classification by the original model. Access to the model can be restricted to be only via well-defined prediction APIs. Nevertheless, prediction APIs still provide enough information to allow an adversary to mount model extraction attacks by sending repeated queries via the prediction API. In this paper, we describe new model extraction attacks using novel approaches for generating synthetic queries, and optimizing training hyperparameters. Our attacks outperform state-of-the-art model extraction in terms of transferability of both targeted and non-targeted adversarial examples (up to +29-44 percentage points, pp), and prediction accuracy (up to +46 pp) on two datasets. We provide take-aways on how to perform effective model extraction attacks. We then propose PRADA, the first step towards generic and effective detection of DNN model extraction attacks. It analyzes the distribution of consecutive API queries and raises an alarm when this distribution deviates from benign behavior. We show that PRADA can detect all prior model extraction attacks with no false positives., Comment: 17 pages, 7 figures, 9 tables. Accepted for publication in the 4th IEEE European Symposium on Security and Privacy (EuroS&P 2019)

In this paper, we discuss the relevance of training data on modern object detectors used on onboard applications. Whereas modern deep learning techniques require large amounts of data, datasets with typical scenarios for autonomous vehicles are scarce and have a reduced number of samples. We conduct a comprehensive set of experiments to understand the effect of using a combination of two relatively small datasets to train an end-to-end object detector, based on the popular Faster R-CNN and enhanced with orientation estimation capabilities. We also test the adequacy of training models using partially available ground-truth labels, as a consequence of combining datasets aimed at different applications. Data augmentation is also introduced into the training pipeline. Results show a significant performance improvement in our exemplary case as a result of the higher variability of the training samples, thus opening a new way to improve the detection performance independently from the detector architecture. This work was supported by the Spanish Government through the CICYT projects TRA2015-63708-R and TRA2016-78886-C3-1-R, and the Comunidad de Madrid through SEGVAUTO-TRIES (S2013/MIT-2713). Publicado

In this paper, we propose an encoder-decoder based deep convolutional network for semantic segmentation in autonomous driving scenarios. The architecture of the proposed model is based on VGG16 [1]. Residual learning is introduced to preserve the context while decreasing the size of feature maps between the stacks of convolutional layers. Also, the resolution is preserved through shortcuts from the encoder stage to the decoder stage. Experiments are conducted on popular benchmark datasets CamVid and CityScapes to demonstrate the efficacy of the proposed model. The experiments are corroborated with comparative analysis with popular encoder-decoder networks such as SegNet and Enet architectures demonstrating that the proposed approach outperforms existing methods despite having fewer trainable parameters.

Using Programming by Demonstration to teach robot learners generalisable skills relies on having effective human teachers. This paper aims to address two problems commonly observed in demonstration data sets that arise due to poor teaching strategies; undemonstrated states and ambiguous demonstrations. Overcoming these issues through the use of visual feedback and simple heuristic rules is investigated as a potential way of guiding novice users to more effectively teach robot learners to generalise a task. The proposed method intends to offer the user a more transparent understanding of the robot learner's model state during the teaching phase, to create a more interactive and robust teaching process. Results from a single-factor, three-phase repeated measures study with $\mathbf{n}=\mathbf{30}$ participants, comparing the proposed feedback and heuristic rules set against an unguided condition, show a statistically significant $(\mathbf{F}(\mathbf{2},\mathbf{58})=\mathbf{7.952},\mathbf{p}=\mathbf{0.001})$ improvement of user teaching efficiency of approximately 180% when using the proposed feedback visualisation.

A major impediment to the application of deep learning to real-world problems is the scarcity of labeled data. Small training sets are in fact of no use to deep networks as, due to the large number of trainable parameters, they will very likely be subject to overfitting phenomena. On the other hand, the increment of the training set size through further manual or semi-automatic labellings can be costly, if not possible at times. Thus, the standard techniques to address this issue are transfer learning and data augmentation, which consists of applying some sort of "transformation" to existing labeled instances to let the training set grow in size. Although this approach works well in applications such as image classification, where it is relatively simple to design suitable transformation operators, it is not obvious how to apply it in more structured scenarios. Motivated by the observation that in virtually all application domains it is easy to obtain unlabeled data, in this paper we take a different perspective and propose a \emph{label augmentation} approach. We start from a small, curated labeled dataset and let the labels propagate through a larger set of unlabeled data using graph transduction techniques. This allows us to naturally use (second-order) similarity information which resides in the data, a source of information which is typically neglected by standard augmentation techniques. In particular, we show that by using known game theoretic transductive processes we can create larger and accurate enough labeled datasets which use results in better trained neural networks. Preliminary experiments are reported which demonstrate a consistent improvement over standard image classification datasets., Accepted on IEEE International Conference on Pattern Recognition

The proliferation and rapid diffusion of fake news on the Internet highlight the need of automatic hoax detection systems. In the context of social networks, machine learning (ML) methods can be used for this purpose. Fake news detection strategies are traditionally either based on content analysis (i.e. analyzing the content of the news) or - more recently - on social context models, such as mapping the news' diffusion pattern., In this paper, we first propose a novel ML, fake news detection method which, by combining news content and social context features, outperforms existing methods in the literature, increasing their already high accuracy by up to 4.8%. Second, we implement our method within a Facebook Messenger chatbot and validate it with a real-world application, obtaining a fake news detection accuracy of 81.7%.

Source at https://doi.org/10.1109/CBMS.2018.00073 Video analysis including classification, segmentation or tagging is one of the most challenging but also interesting topics multimedia research currently try to tackle. This is often related to videos from surveillance cameras or social media. In the last years, also medical institutions produce more and more video and image content. Some areas of medical image analysis, like radiology or brain scans, are well covered, but there is a much broader potential of medical multimedia content analysis. For example, in colonoscopy, 20% of polyps are missed or incompletely removed on average. Thus, automatic detection to support medical experts can be useful. In this paper, we present and evaluate several machine learning-based approaches for real-time polyp detection for live colonoscopy. We propose pixel-wise localization and frame-wise detection methods which include both handcrafted and deep learning based approaches. The experimental results demonstrate the capability of analyzing multimedia content in real clinical settings, the possible improvements in the work flow and the potential improved detection rates for medical experts.

The use of computer-based and online education systems has made new data available that can describe the temporal and process-level progression of learning. To date, machine learning research has not considered the impacts of these properties on the machine learning prediction task in educational settings. Machine learning algorithms may have applications in supporting targeted intervention approaches. The goals of this paper are to: 1) determine the impact of process-level information on machine learning prediction results and 2) establish the effect of type of machine learning algorithm used on prediction results. Data were collected from a university level course in human factors engineering (n=35), which included both traditional classroom assessment and computer-based assessment methods. A set of common regression and classification algorithms were applied to the data to predict final course score. The overall prediction accuracy as well as the chronological progression of prediction accuracy was analyzed for each algorithm. Simple machine learning algorithms (linear regression, logistic regression) had comparable performance with more complex methods (support vector machines, artificial neural networks). Process-level information was not useful in post-hoc predictions, but contributed significantly to allowing for accurate predictions to be made earlier in the course. Process level information provides useful prediction features for development of targeted intervention techniques, as it allows more accurate predictions to be made earlier in the course. For small course data sets, the prediction accuracy and simplicity of linear regression and logistic regression make these methods preferable to more complex algorithms.

In this paper, a novel approach for human gesture classification on skeletal data is proposed for the application of exergaming in physiotherapy. Unlike existing methods, we propose to use a general classifier like Random Forests to recognize dynamic gestures. The temporal dimension is handled afterwards by majority voting in a sliding window over the consecutive predictions of the classifier. The gestures can have partially similar postures, such that the classifier will decide on the dissimilar postures. This brute-force classification strategy is permitted, because dynamic human gestures show sufficient dissimilar postures. Online continuous human gesture recognition can classify dynamic gestures in an early stage, which is a crucial advantage when controlling a game by automatic gesture recognition. Also, ground truth can be easily obtained, since all postures in a gesture get the same label, without any discretization into consecutive postures. This way, new gestures can be easily added, which is advantageous in adaptive game development. We evaluate our strategy by a leave-one-subject-out cross-validation on a self-captured stealth game gesture dataset and the publicly available Microsoft Research Cambridge-12 Kinect (MSRC-12) dataset. On the first dataset we achieve an excellent accuracy rate of 96.72%. Furthermore, we show that Random Forests perform better than Support Vector Machines. On the second dataset we achieve an accuracy rate of 98.37%, which is on average 3.57% better then existing methods.

Conventional metric learning methods usually assume that the training and test samples are captured in similar scenarios so that their distributions are assumed to be the same. This assumption doesn't hold in many real visual recognition applications, especially when samples are captured across different datasets. In this paper, we propose a new deep transfer metric learning (DTML) method to learn a set of hierarchical nonlinear transformations for cross-domain visual recognition by transferring discriminative knowledge from the labeled source domain to the unlabeled target domain. Specifically, our DTML learns a deep metric network by maximizing the inter-class variations and minimizing the intra-class variations, and minimizing the distribution divergence between the source domain and the target domain at the top layer of the network. To better exploit the discriminative information from the source domain, we further develop a deeply supervised transfer metric learning (DSTML) method by including an additional objective on DTML where the output of both the hidden layers and the top layer are optimized jointly. Experimental results on cross-dataset face verification and person re-identification validate the effectiveness of the proposed methods. Accepted version

Humans often seek a second or third opinion about an important matter. Then, a final decision is reached after weighing and combining these opinions. This idea is the base of the ensemble based systems. Ensembles of classifiers are well established as a method for obtaining highly accurate classifiers by combining less accurate ones. On the other hand, evolving classifiers are inspired by the idea of evolve their structure in order to adapt to the changes of the environment. In this paper, we present a proof-of-concept method for constructing an ensemble system based on Evolving Fuzzy Systems. The main contribution of this approach is that the base-classifiers are self-developing (evolving) Fuzzy-rule-based (FRB) classifiers. Thus, we present an ensemble system which is based on evolving classifiers and keeps the properties of the evolving approach classification of streaming data. It is important to clarify that the evolving classifiers are gradually developing but they are not genetic or evolutionary. This work has been supported by the Spanish Government under i-Support (Intelligent Agent Based Driver Decision Support) Project (TRA2011-29454-C03-03).

The ever-growing need for gene-expression data analysis motivates studies in sample generation due to the lack of enough gene-expression data. It is common that there are thousands of genes but only tens or rarely hundreds of samples available. In this paper, we attempt to formulate the sample generation task as follows: first, building alternative Gene Regulatory Network (GRN) models; second, sampling data from each of them; and then filtering the generated samples using metrics that measure compatibility, diversity and coverage with respect to the original dataset. We constructed two alternative GRN models using Probabilistic Boolean Networks and Ordinary Differential Equations. We developed a multi-objective filtering mechanism based on the three metrics to assess the quality of the newly generated data. We presented a number of experiments to show effectiveness and applicability of the proposed multi-model framework.

ImageNet is a large-scale database of object classes with millions of images. Unfortunately only a small fraction of them is manually annotated with bounding-boxes. This prevents useful developments, such as learning reliable object detectors for thousands of classes. In this paper we propose to automatically populate ImageNet with many more bounding-boxes, by leveraging existing manual annotations. The key idea is to localize objects of a target class for which annotations are not available, by transferring knowledge from related source classes with available annotations. We distinguish two kinds of source classes: ancestors and siblings. Each source provides knowledge about the plausible location, appearance and context of the target objects, which induces a probability distribution over windows in images of the target class. We learn to combine these distributions so as to maximize the location accuracy of the most probable window. Finally, we employ the combined distribution in a procedure to jointly localize objects in all images of the target class. Through experiments on 0.5 million images from 219 classes we show that our technique (i) annotates a wide range of classes with bounding-boxes; (ii) effectively exploits the hierarchical structure of ImageNet, since all sources and types of knowledge we propose contribute to the results; (iii) scales efficiently.