The task of multi-step ahead prediction in language models is challenging considering the discrepancy between training and testing. At test time, a language model is required to make predictions given past predictions as input, instead of the past targets that are provided during training. This difference, known as exposure bias, can lead to the compounding of errors along a generated sequence at test time. In order to improve generalization in neural language models and address compounding errors, we propose a curriculum learning based method that gradually changes an initially deterministic teacher policy to a gradually more stochastic policy, which we refer to as \textit{Nearest-Neighbor Replacement Sampling}. A chosen input at a given timestep is replaced with a sampled nearest neighbor of the past target with a truncated probability proportional to the cosine similarity between the original word and its top $k$ most similar words. This allows the teacher to explore alternatives when the teacher provides a sub-optimal policy or when the initial policy is difficult for the learner to model. The proposed strategy is straightforward, online and requires little additional memory requirements. We report our main findings on two language modelling benchmarks and find that the proposed approach performs particularly well when used in conjunction with scheduled sampling, that too attempts to mitigate compounding errors in language models.

Abstract--Extended Kalman Filtering (EKF) can be used to propagate and quantify input uncertainty through a Deep Neural Network (DNN) assuming mild hypotheses on the input distribution. This methodology yields results comparable to existing methods of uncertainty propagation for DNNs while lowering the computational overhead considerably. Additionally, EKF allows model error to be naturally incorporated into the output uncertainty. This question tends to come up during confidence scoring in areas such as automatic speech recognition where things like background noise can distort the input signal [1]. However, it can be approximated by a Gaussian and modified later if necessary [2].

Recently, product images have gained increasing attention in clothing recommendation since the visual appearance of clothing products has a significant impact on consumers' decision. Most existing methods rely on conventional features to represent an image, such as the visual features extracted by convolutional neural networks (CNN features) and the scale-invariant feature transform algorithm (SIFT features), color histograms, and so on. Nevertheless, one important type of features, the \emph{aesthetic features}, is seldom considered. It plays a vital role in clothing recommendation since a users' decision depends largely on whether the clothing is in line with her aesthetics, however the conventional image features cannot portray this directly. To bridge this gap, we propose to introduce the aesthetic information, which is highly relevant with user preference, into clothing recommender systems. To achieve this, we first present the aesthetic features extracted by a pre-trained neural network, which is a brain-inspired deep structure trained for the aesthetic assessment task. Considering that the aesthetic preference varies significantly from user to user and by time, we then propose a new tensor factorization model to incorporate the aesthetic features in a personalized manner. We conduct extensive experiments on real-world datasets, which demonstrate that our approach can capture the aesthetic preference of users and significantly outperform several state-of-the-art recommendation methods.

Process Mining consists of techniques where logs created by operative systems are transformed into process models. In process mining tools it is often desired to be able to classify ongoing process instances, e.g., to predict how long the process will still require to complete, or to classify process instances to different classes based only on the activities that have occurred in the process instance thus far. Recurrent neural networks and its subclasses, such as Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM), have been demonstrated to be able to learn relevant temporal features for subsequent classification tasks. In this paper we apply recurrent neural networks to classifying process instances. The proposed model is trained in a supervised fashion using labeled process instances extracted from event log traces. This is the first time we know of GRU having been used in classifying business process instances. Our main experimental results shows that GRU outperforms LSTM remarkably in training time while giving almost identical accuracies to LSTM models. Additional contributions of our paper are improving the classification model training time by filtering infrequent activities, which is a technique commonly used, e.g., in Natural Language Processing (NLP).

Class labels are often imperfectly observed, due to mistakes and to genuine ambiguity among classes. We propose a new semi-supervised deep generative model that explicitly models noisy labels, called the Mislabeled VAE (M-VAE). The M-VAE can perform better than existing deep generative models which do not account for label noise. Additionally, the derivation of M-VAE gives new theoretical insights into the popular M1+M2 semi-supervised model.

In this paper, we integrate VAEs and flow-based generative models successfully and get f-VAEs. Compared with VAEs, f-VAEs generate more vivid images, solved the blurred-image problem of VAEs. Compared with flow-based models such as Glow, f-VAE is more lightweight and converges faster, achieving the same performance under smaller-size architecture. Recently, deep generative models has been widely studied and developed. Outside of Generative Adversarial Networks (GANs) (Goodfellow et al. 2014), Variational Autoencoders (VAEs) (Kingma and Welling 2013) and flow-based models (Dinh, Krueger, and Bengio 2014; Dinh, Sohldickstein, and Bengio 2016) are two distinct kinds of competitive generative models. They have their own advantages and disadvantages, and we try to integrate them to a new model.

With the rise of self-driving vehicles comes the risk of accidents and the need for higher safety, and protection for pedestrian detection in the following scenarios: imminent crashes, thus the car should crash into an object and avoid the pedestrian, and in the case of road intersections, where it is important for the car to stop when pedestrians are crossing. Currently, a special topology of deep neural networks called Fused Deep Neural Network (F-DNN) is considered to be the state of the art in pedestrian detection, as it has the lowest miss rate, yet it is very slow. Therefore, acceleration is needed to speed up the performance. This project proposes two contributions to address this problem, by using a deep neural network used for object detection, called Single Shot Multi-Box Detector (SSD). The first contribution is training and tuning the hyperparameters of SSD to improve pedestrian detection. The second contribution is a new FPGA design for accelerating the model on the Altera Arria 10 platform. The final system will be used in self-driving vehicles in real-time. Preliminary results of the improved SSD shows 3% higher miss-rate than F-DNN on Caltech pedestrian detection benchmark, but 4x performance improvement. The acceleration design is expected to achieve an additional performance improvement significantly outweighing the minimal difference in accuracy.

While the machine learning literature dedicated to fully automated reasoning algorithms is abundant, the number of methods enabling the inference process on the basis of previously defined knowledge structures is scanter. Fuzzy Cognitive Maps (FCMs) are neural networks that can be exploited towards this goal because of their flexibility to handle external knowledge. However, FCMs suffer from a number of issues that range from the limited prediction horizon to the absence of theoretically sound learning algorithms able to produce accurate predictions. In this paper, we propose a neural network system named Short-term Cognitive Networks that tackle some of these limitations. In our model weights are not constricted and may have a causal nature or not. As a second contribution, we present a nonsynaptic learning algorithm to improve the network performance without modifying the previously defined weights. Moreover, we derive a stop condition to prevent the learning algorithm from iterating without decreasing the simulation error.

The tremendous potential exhibited by deep learning is often offset by architectural and computational complexity, making widespread deployment a challenge for edge scenarios such as mobile and other consumer devices. To tackle this challenge, we explore the following idea: Can we learn generative machines to automatically generate deep neural networks with efficient network architectures? In this study, we introduce the idea of generative synthesis, which is premised on the intricate interplay between a generator-inquisitor pair that work in tandem to garner insights and learn to generate highly efficient deep neural networks that best satisfies operational requirements. What is most interesting is that, once a generator has been learned through generative synthesis, it can be used to generate not just one but a large variety of different, unique highly efficient deep neural networks that satisfy operational requirements. Experimental results for image classification, semantic segmentation, and object detection tasks illustrate the efficacy of generative synthesis in producing generators that automatically generate highly efficient deep neural networks (which we nickname FermiNets) with higher model efficiency and lower computational costs (reaching >10x more efficient and fewer multiply-accumulate operations than several tested state-of-the-art networks), as well as higher energy efficiency (reaching >4x improvements in image inferences per joule consumed on a Nvidia Tegra X2 mobile processor). As such, generative synthesis can be a powerful, generalized approach for accelerating and improving the building of deep neural networks for on-device edge scenarios.

Cloud Robotics is a paradigm where distributed robots are connected to cloud services via networks to access unlimited computation power, at the cost of network communication. However, due to limitations such as network latency and variability, it is difficult to control dynamic, human compliant service robots directly from the cloud. In this work, by leveraging asynchronous protocol with a heartbeat signal, we combine cloud robotics with a smart edge device to build a Fog Robotic system. We use the system to enable robust teleoperation of a dynamic self-balancing robot from the cloud. We first use the system to pick up boxes from static locations, a task commonly performed in warehouse logistics. To make cloud teleoperation more efficient, we deploy image based visual servoing (IBVS) to perform box pickups automatically. Visual feedbacks, including apriltag recognition and tracking, are performed in the cloud to emulate a Fog Robotic object recognition system for IBVS. We demonstrate the feasibility of real-time dynamic automation system using this cloud-edge hybrid, which opens up possibilities of deploying dynamic robotic control with deep-learning recognition systems in Fog Robotics. Finally, we show that Fog Robotics enables the self-balancing service robot to pick up a box automatically from a person under unstructured environments.