In recent years Deep Neural Networks (DNNs) have been rapidly developed in various applications, together with increasingly complex architectures. The performance gain of these DNNs generally comes with high computational costs and large memory consumption, which may not be affordable for mobile platforms. Deep model quantization can be used for reducing the computation and memory costs of DNNs, and deploying complex DNNs on mobile equipment. In this work, we propose an optimization framework for deep model quantization. First, we propose a measurement to estimate the effect of parameter quantization errors in individual layers on the overall model prediction accuracy. Then, we propose an optimization process based on this measurement for finding optimal quantization bit-width for each layer. This is the first work that theoretically analyse the relationship between parameter quantization errors of individual layers and model accuracy. Our new quantization algorithm outperforms previous quantization optimization methods, and achieves 20-40% higher compression rate compared to equal bit-width quantization at the same model prediction accuracy.

Many computationally-efficient methods for Bayesian deep learning rely on continuous optimization algorithms, but the implementation of these methods requires significant changes to existing code-bases. In this paper, we propose Vprop, a method for Gaussian variational inference that can be implemented with two minor changes to the off-the-shelf RMSprop optimizer. Vprop also reduces the memory requirements of Black-Box Variational Inference by half. We derive Vprop using the conjugate-computation variational inference method, and establish its connections to Newton's method, natural-gradient methods, and extended Kalman filters. Overall, this paper presents Vprop as a principled, computationally-efficient, and easy-to-implement method for Bayesian deep learning.

Machine learning approaches hold great potential for the automated detection of lung nodules in chest radiographs, but training the algorithms requires vary large amounts of manually annotated images, which are difficult to obtain. Weak labels indicating whether a radiograph is likely to contain pulmonary nodules are typically easier to obtain at scale by parsing historical free-text radiological reports associated to the radiographs. Using a repositotory of over 700,000 chest radiographs, in this study we demonstrate that promising nodule detection performance can be achieved using weak labels through convolutional neural networks for radiograph classification. We propose two network architectures for the classification of images likely to contain pulmonary nodules using both weak labels and manually-delineated bounding boxes, when these are available. Annotated nodules are used at training time to deliver a visual attention mechanism informing the model about its localisation performance. The first architecture extracts saliency maps from high-level convolutional layers and compares the estimated position of a nodule against the ground truth, when this is available. A corresponding localisation error is then back-propagated along with the softmax classification error. The second approach consists of a recurrent attention model that learns to observe a short sequence of smaller image portions through reinforcement learning. When a nodule annotation is available at training time, the reward function is modified accordingly so that exploring portions of the radiographs away from a nodule incurs a larger penalty. Our empirical results demonstrate the potential advantages of these architectures in comparison to competing methodologies.

Can artificial intelligence (AI) learn complicated non-linear physics? Here we propose a novel deep learning approach that learns non-linear photon scattering physics and obtains accurate 3D distribution of optical anomalies. In contrast to the traditional black-box deep learning approaches to inverse problems, our deep network learns to invert the Lippmann-Schwinger integral equation which describes the essential physics of photon migration of diffuse near-infrared (NIR) photons in turbid media. As an example for clinical relevance, we applied the method to our prototype diffuse optical tomography (DOT). We show that our deep neural network, trained with only simulation data, can accurately recover the location of anomalies within biomimetic phantoms and live animals without the use of an exogenous contrast agent.

This paper proposes a new algorithm for controlling classification results by generating a small additive perturbation without changing the classifier network. Our work is inspired by existing works generating adversarial perturbation that worsens classification performance. In contrast to the existing methods, our work aims to generate perturbations that can enhance overall classification performance. To solve this performance enhancement problem, we newly propose a perturbation generation network (PGN) influenced by the adversarial learning strategy. In our problem, the information in a large external dataset is summarized by a small additive perturbation, which helps to improve the performance of the classifier trained with the target dataset. In addition to this performance enhancement problem, we show that the proposed PGN can be adopted to solve the classical adversarial problem without utilizing the information on the target classifier. The mentioned characteristics of our method are verified through extensive experiments on publicly available visual datasets.

We propose a practical method for $L_0$ norm regularization for neural networks: pruning the network during training by encouraging weights to become exactly zero. Such regularization is interesting since (1) it can greatly speed up training and inference, and (2) it can improve generalization. AIC and BIC, well-known model selection criteria, are special cases of $L_0$ regularization. However, since the $L_0$ norm of weights is non-differentiable, we cannot incorporate it directly as a regularization term in the objective function. We propose a solution through the inclusion of a collection of non-negative stochastic gates, which collectively determine which weights to set to zero. We show that, somewhat surprisingly, for certain distributions over the gates, the expected $L_0$ norm of the resulting gated weights is differentiable with respect to the distribution parameters. We further propose the \emph{hard concrete} distribution for the gates, which is obtained by "stretching" a binary concrete distribution and then transforming its samples with a hard-sigmoid. The parameters of the distribution over the gates can then be jointly optimized with the original network parameters. As a result our method allows for straightforward and efficient learning of model structures with stochastic gradient descent and allows for conditional computation in a principled way. We perform various experiments to demonstrate the effectiveness of the resulting approach and regularizer.

An important problem in training deep networks with high capacity is to ensure that the trained network works well when presented with new inputs outside the training dataset. Dropout is an effective regularization technique to boost the network generalization in which a random subset of the elements of the given data and the extracted features are set to zero during the training process. In this paper, a new randomized regularization technique in which we withhold a random part of the data without necessarily turning off the neurons/data-elements is proposed. In the proposed method, of which the conventional dropout is shown to be a special case, random data dropout is performed in an arbitrary basis, hence the designation Generalized Dropout. We also present a framework whereby the proposed technique can be applied efficiently to convolutional neural networks. The presented numerical experiments demonstrate that the proposed technique yields notable performance gain. Generalized Dropout provides new insight into the idea of dropout, shows that we can achieve different performance gains by using different bases matrices, and opens up a new research question as of how to choose optimal bases matrices that achieve maximal performance gain.

In this paper, we introduce and evaluate PROPEDEUTICA, a novel methodology and framework for efficient and effective real-time malware detection, leveraging the best of conventional machine learning (ML) and deep learning (DL) algorithms. In PROPEDEUTICA, all software processes in the system start execution subjected to a conventional ML detector for fast classification. If a piece of software receives a borderline classification, it is subjected to further analysis via more performance expensive and more accurate DL methods, via our newly proposed DL algorithm DEEPMALWARE. Further, we introduce delays to the execution of software subjected to deep learning analysis as a way to "buy time" for DL analysis and to rate-limit the impact of possible malware in the system. We evaluated PROPEDEUTICA with a set of 9,115 malware samples and 877 commonly used benign software samples from various categories for the Windows OS. Our results show that the false positive rate for conventional ML methods can reach 20%, and for modern DL methods it is usually below 6%. However, the classification time for DL can be 100X longer than conventional ML methods. PROPEDEUTICA improved the detection F1-score from 77.54% (conventional ML method) to 90.25%, and reduced the detection time by 54.86%. Further, the percentage of software subjected to DL analysis was approximately 40% on average. Further, the application of delays in software subjected to ML reduced the detection time by approximately 10%. Finally, we found and discussed a discrepancy between the detection accuracy offline (analysis after all traces are collected) and on-the-fly (analysis in tandem with trace collection). Our insights show that conventional ML and modern DL-based malware detectors in isolation cannot meet the needs of efficient and effective malware detection: high accuracy, low false positive rate, and short classification time.

Imitation learning has been commonly applied to solve different tasks in isolation. This usually requires either careful feature engineering, or a significant number of samples. This is far from what we desire: ideally, robots should be able to learn from very few demonstrations of any given task, and instantly generalize to new situations of the same task, without requiring task-specific engineering. In this paper, we propose a meta-learning framework for achieving such capability, which we call one-shot imitation learning. Specifically, we consider the setting where there is a very large set of tasks, and each task has many instantiations. For example, a task could be to stack all blocks on a table into a single tower, another task could be to place all blocks on a table into two-block towers, etc. In each case, different instances of the task would consist of different sets of blocks with different initial states. At training time, our algorithm is presented with pairs of demonstrations for a subset of all tasks. A neural net is trained that takes as input one demonstration and the current state (which initially is the initial state of the other demonstration of the pair), and outputs an action with the goal that the resulting sequence of states and actions matches as closely as possible with the second demonstration. At test time, a demonstration of a single instance of a new task is presented, and the neural net is expected to perform well on new instances of this new task. The use of soft attention allows the model to generalize to conditions and tasks unseen in the training data. We anticipate that by training this model on a much greater variety of tasks and settings, we will obtain a general system that can turn any demonstrations into robust policies that can accomplish an overwhelming variety of tasks. Videos available at https://bit.ly/nips2017-oneshot .

This is a simpler problem than, say, full document summarization. The approach follows the general approach used for neural machine translation with an encoder and a decoder.