One of the main issue in robotics is the lack of embedded computational power. Recently, state of the art algorithms providing a better understanding of the surroundings (Object detection, skeleton tracking, etc.) are requiring more and more computational power. The lack of embedded computational power is more significant in mass-produced robots because of the difficulties to follow the increasing computational requirements of state of the art algorithms. The integration of an additional GPU allows to overcome this lack of embedded computational power. We introduce in this paper a prototype of Pepper with an embedded GPU, but also with an additional 3D camera on the head of the robot and plugged to the late GPU. This prototype, called Adapted Pepper, was built for the European project called MuMMER (MultiModal Mall Entertainment Robot) in order to embed algorithms like OpenPose, YOLO or to process sensors information and, in all cases, avoid network dependency for deported computation.

Artificial neural networks (ANNs) have occupied the center stage in the realm of deep learning in the recent past. ANNs are made up of several hidden layers, and each hidden layer consists of several neurons. At each neuron, an affine linear map is composed with a nonlinear function known as activation function. During the training of an ANN, the linear map is optimized, however an activation function is usually fixed in the beginning along with the architecture of the ANN. There has been an increasing interest in developing a methodical understanding of activation functions, in particular with regards to the construction of novel activation functions and identifying mathematical properties leading to a better learning [1]. An activation function is considered good if it can increase the learning rate and leaning to better convergence which leads to more accurate results. At the early stage of deep learning research, researchers used shallow networks (fewer hidden layers), and tanh or sigmoid, were used as activation functions. As the research progressed and deeper networks (multiple hidden layers) came into fashion to achieve challenging tasks, Rectified Linear Unit (ReLU)([2], [3], [4]) emerged as the most popular activation function. Despite its simplicity, deep neural networks with ReLU have learned many complex and highly nonlinear functions with high accuracy.

Electroencephalography (EEG) measures the electrical brain activity in real-time by using sensors placed on the scalp. Artifacts, due to eye movements and blink, muscular/cardiac activity and generic electrical disturbances, have to be recognized and eliminated to allow a correct interpretation of the useful brain signals (UBS) of EEG. Independent Component Analysis (ICA) is effective to split the signal into independent components (ICs) whose re-projections on 2D scalp topographies (images), also called topoplots, allow to recognize/separate artifacts and by UBS. Until now, IC topoplot analysis, a gold standard in EEG, has been carried on visually by human experts and, hence, not usable in automatic, fast-response EEG. We present a completely automatic and effective framework for EEG artifact recognition by IC topoplots, based on 2D Convolutional Neural Networks (CNNs), capable to divide topoplots in 4 classes: 3 types of artifacts and UBS. The framework setup is described and results are presented, discussed and compared with those obtained by other competitive strategies. Experiments, carried on public EEG datasets, have shown an overall accuracy of above 98%, employing 1.4 sec on a standard PC to classify 32 topoplots, that is to drive an EEG system of 32 sensors. Though not real-time, the proposed framework is efficient enough to be used in fast-response EEG-based Brain-Computer Interfaces (BCI) and faster than other automatic methods based on ICs.

Continual learning from a sequential stream of data is a crucial challenge for machine learning research. Most studies have been conducted on this topic under the single-label classification setting along with an assumption of balanced label distribution. This work expands this research horizon towards multi-label classification. In doing so, we identify unanticipated adversity innately existent in many multi-label datasets, the long-tailed distribution. We jointly address the two independently solved problems, Catastropic Forgetting and the long-tailed label distribution by first empirically showing a new challenge of destructive forgetting of the minority concepts on the tail. Then, we curate two benchmark datasets, COCOseq and NUS-WIDEseq, that allow the study of both intra- and inter-task imbalances. Lastly, we propose a new sampling strategy for replay-based approach named Partitioning Reservoir Sampling (PRS), which allows the model to maintain a balanced knowledge of both head and tail classes. We publicly release the dataset and the code in our project page.

In the context of time-series forecasting, we propose a LSTM-based recurrent neural network architecture and loss function that enhance the stability of the predictions. In particular, the loss function penalizes the model, not only on the prediction error (mean-squared error), but also on the predicted variation error. We apply this idea to the prediction of future glucose values in diabetes, which is a delicate task as unstable predictions can leave the patient in doubt and make him/her take the wrong action, threatening his/her life. The study is conducted on type 1 and type 2 diabetic people, with a focus on predictions made 30-minutes ahead of time. First, we confirm the superiority, in the context of glucose prediction, of the LSTM model by comparing it to other state-of-the-art models (Extreme Learning Machine, Gaussian Process regressor, Support Vector Regressor). Then, we show the importance of making stable predictions by smoothing the predictions made by the models, resulting in an overall improvement of the clinical acceptability of the models at the cost in a slight loss in prediction accuracy. Finally, we show that the proposed approach, outperforms all baseline results. More precisely, it trades a loss of 4.3\% in the prediction accuracy for an improvement of the clinical acceptability of 27.1\%. When compared to the moving average post-processing method, we show that the trade-off is more efficient with our approach.

While deep learning has been successfully applied to many real-world computer vision tasks, training robust classifiers usually requires a large amount of well-labeled data. However, the annotation is often expensive and time-consuming. Few-shot image classification has thus been proposed to effectively use only a limited number of labeled examples to train models for new classes. Recent works based on transferable metric learning methods have achieved promising classification performance through learning the similarity between the features of samples from the query and support sets. However, rare of them explicitly considers the model interpretability, which can actually be revealed during the training phase. For that, in this work, we propose a metric learning based method named Region Comparison Network (RCN), which is able to reveal how few-shot learning works as in a neural network as well as to find out specific regions that are related to each other in images coming from the query and support sets. Moreover, we also present a visualization strategy named Region Activation Mapping (RAM) to intuitively explain what our method has learned by visualizing intermediate variables in our network. We also present a new way to generalize the interpretability from the level of tasks to categories, which can also be viewed as a method to find the prototypical parts for supporting the final decision of our RCN. Extensive experiments on four benchmark datasets clearly show the effectiveness of our method over existing baselines.

Detecting changed regions in paired satellite images plays a key role in many remote sensing applications. The evolution of recent techniques could provide satellite images with very high spatial resolution (VHR) and made it challenging to apply image coregistration whose accuracy is the basis of many change detection methods.Due to the advantage in deep feature representation, deep learning is introduced to detect changes on unregistered images. However, the absence of ground truth makes the performance of deep learning models in unsupervised task hard to be evaluated or be guaranteed.To alleviate the effect of unregistered pairs and make better use of deep learning structures, we propose a novel change detection procedure based on a special neural network architecture---Generative Adversarial Network (GAN).GAN features generating realistic images rather than giving hypervectors that contain visual features, so it is easy to evaluate the GAN model by judging the generated images. In this paper, we show that GAN model can be trained upon a pair of images through utilizing the proposed expanding strategy to create a training set and optimising designed objective functions. The optimised GAN model would produce many coregistered images where changes can be easily spotted and then the change map can be presented through a comparison strategy using these generated images explicitly.Compared to other deep learning-based methods, our method is less sensitive to the problem of unregistered images and makes most of the deep learning structure.Experimental results on synthetic images and real data with many different scenes could demonstrate the effectiveness of the proposed approach.

Deep active inference has been proposed as a scalable approach to perception and action that deals with large policy and state spaces. However, current models are limited to fully observable domains. In this paper, we describe a deep active inference model that can learn successful policies directly from high-dimensional sensory inputs. The deep learning architecture optimizes a variant of the expected free energy and encodes the continuous state representation by means of a variational autoencoder. We show, in the OpenAI benchmark, that our approach has comparable or better performance than deep Q-learning, a state-of-the-art deep reinforcement learning algorithm.

Data often are formed of multiple modalities, which jointly describe the observed phenomena. Modeling the joint distribution of multimodal data requires larger expressive power to capture high-level concepts and provide better data representations. However, multimodal generative models based on variational inference are limited due to the lack of flexibility of the approximate posterior, which is obtained by searching within a known parametric family of distributions. We introduce a method that improves the representational capacity of multimodal variational methods using normalizing flows. It approximates the joint posterior with a simple parametric distribution and subsequently transforms into a more complex one. Through several experiments, we demonstrate that the model improves on state-of-the-art multimodal methods based on variational inference on various computer vision tasks such as colorization, edge and mask detection, and weakly supervised learning. We also show that learning more powerful approximate joint distributions improves the quality of the generated samples.

Abstract--In this paper, several Collaborative Filtering (CF) approaches with latent variable methods were studied using user-item interactions to capture important hidden variations of the sparse customer purchasing behaviors. The latent factors are used to generalize the purchasing pattern of the customers and to provide product recommendations. CF with Neural Collaborative Filtering (NCF) was shown to produce the highest Normalized Discounted Cumulative Gain (NDCG) performance on the real-world proprietary dataset provided by a large parts supply company. Different hyperparameters were tested using Bayesian Optimization (BO) for applicability in the CF framework. External data sources like click-data and metrics like Clickthrough Rate (CTR) were reviewed for potential extensions to the work presented. The work shown in this paper provides techniques the Company can use to provide product recommendations to enhance revenues, attract new customers, and gain advantages over competitors. With today's ever-increasing ease of access to the internet more advertisements, attract new clients, and retain existing and information, we have reached a point of information clients [6].