Since 2014 transfer learning has become the key driver for the improvement of spatial saliency prediction; however, with stagnant progress in the last 3-5 years. We conduct a large-scale transfer learning study which tests different ImageNet backbones, always using the same read out architecture and learning protocol adopted from DeepGaze II. By replacing the VGG19 backbone of DeepGaze II with ResNet50 features we improve the performance on saliency prediction from 78% to 85%. However, as we continue to test better ImageNet models as backbones (such as EfficientNetB5) we observe no additional improvement on saliency prediction. By analyzing the backbones further, we find that generalization to other datasets differs substantially, with models being consistently overconfident in their fixation predictions. We show that by combining multiple backbones in a principled manner a good confidence calibration on unseen datasets can be achieved. This yields a significant leap in benchmark performance in and out-of-domain with a 15 percent point improvement over DeepGaze II to 93% on MIT1003, marking a new state of the art on the MIT/Tuebingen Saliency Benchmark in all available metrics (AUC: 88.3%, sAUC: 79.4%, CC: 82.4%).

Reinforcement learning (RL)-based neural architecture search (NAS) generally guarantees better convergence yet suffers from the requirement of huge computational resources compared with gradient-based approaches, due to the rollout bottleneck - exhaustive training of each sampled architecture on the proxy tasks. In this paper, we propose a general pipeline to accelerate the convergence of the rollout process as well as the RL process in NAS. It is motivated by the interesting observation that both the architecture and the parameter knowledge can be transferred between different search processes and even different tasks. We first introduce an uncertainty-aware critic (value function) in Proximal Policy Optimization (PPO) [27] to take advantage of the architecture knowledge in previous search processes, which stabilizes the training process and reduce the searching time by 4 times. In addition, an architecture knowledge pool together with a block similarity function is proposed to utilize parameter knowledge and reduces the searching time by 2 times. To the best of our knowledge, this is the first method that introduces a block-level weight sharing scheme in RL-based NAS. The block similarity function guarantees a 100% hit ratio with strict fairness [5]. Besides, we show that an off-policy correction factor used in "replay buffer" of RL optimization can further reduce half of the searching time. Experiments on the Mobile Neural Architecture Search (MNAS) [30] search space show that the proposed Fast Neural Architecture Search (FNAS) accelerates the standard RL-based NAS process by 10x (e.g., 20,000 GPU hours to 2,000 GPU hours for MNAS), and guarantees better performance on various vision tasks.

We present a Bayesian treatment for deep regression using an Errors-in-Variables model which accounts for the uncertainty associated with the input to the employed neural network. It is shown how the treatment can be combined with already existing approaches for uncertainty quantification that are based on variational inference. Our approach yields a decomposition of the predictive uncertainty into an aleatoric and epistemic part that is more complete and, in many cases, more consistent from a statistical perspective. We illustrate and discuss the approach along various toy and real world examples.

Recently many first and second order variants of SGD have been proposed to facilitate training of Deep Neural Networks (DNNs). A common limitation of these works stem from the fact that they use the same learning rate across all instances present in the dataset. This setting is widely adopted under the assumption that loss functions for each instance are similar in nature, and hence, a common learning rate can be used. In this work, we relax this assumption and propose an optimization framework which accounts for difference in loss function characteristics across instances. More specifically, our optimizer learns a dynamic learning rate for each instance present in the dataset. Learning a dynamic learning rate for each instance allows our optimization framework to focus on different modes of training data during optimization. When applied to an image classification task, across different CNN architectures, learning dynamic learning rates leads to consistent gains over standard optimizers. When applied to a dataset containing corrupt instances, our framework reduces the learning rates on noisy instances, and improves over the state-of-the-art. Finally, we show that our optimization framework can be used for personalization of a machine learning model towards a known targeted data distribution.

Recently, deep learning has been an area of intense researching. However, as a kind of computing intensive task, deep learning highly relies on the the scale of the GPU memory, which is usually expensive and scarce. Although there are some extensive works have been proposed for dynamic GPU memory management, they are hard to be applied to systems with multitasking dynamic workloads, such as in-database machine learning system. In this paper, we demonstrated TENSILE, a method of managing GPU memory in tensor granularity to reduce the GPU memory peak, with taking the multitasking dynamic workloads into consideration. As far as we know, TENSILE is the first method which is designed to manage multiple workloads' GPU memory using. We implement TENSILE on our own deep learning framework, and evaluated its performance. The experiment results shows that our method can achieve less time overhead than prior works with more GPU memory saved.

Knowledge distillation, i.e., one classifier being trained on the outputs of another classifier, is an empirically very successful technique for knowledge transfer between classifiers. It has even been observed that classifiers learn much faster and more reliably if trained with the outputs of another classifier as soft labels, instead of from ground truth data. So far, however, there is no satisfactory theoretical explanation of this phenomenon. In this work, we provide the first insights into the working mechanisms of distillation by studying the special case of linear and deep linear classifiers. Specifically, we prove a generalization bound that establishes fast convergence of the expected risk of a distillation-trained linear classifier. From the bound and its proof we extract three key factors that determine the success of distillation: * data geometry -- geometric properties of the data distribution, in particular class separation, has a direct influence on the convergence speed of the risk; * optimization bias -- gradient descent optimization finds a very favorable minimum of the distillation objective; and * strong monotonicity -- the expected risk of the student classifier always decreases when the size of the training set grows.

Mobility on demand (MoD) systems show great promise in realizing flexible and efficient urban transportation. However, significant technical challenges arise from operational decision making associated with MoD vehicle dispatch and fleet rebalancing. For this reason, operators tend to employ simplified algorithms that have been demonstrated to work well in a particular setting. To help bridge the gap between novel and existing methods, we propose a modular framework for fleet rebalancing based on model-free reinforcement learning (RL) that can leverage an existing dispatch method to minimize system cost. In particular, by treating dispatch as part of the environment dynamics, a centralized agent can learn to intermittently direct the dispatcher to reposition free vehicles and mitigate against fleet imbalance. We formulate RL state and action spaces as distributions over a grid partitioning of the operating area, making the framework scalable and avoiding the complexities associated with multiagent RL. Numerical experiments, using real-world trip and network data, demonstrate that this approach has several distinct advantages over baseline methods including: improved system cost; high degree of adaptability to the selected dispatch method; and the ability to perform scale-invariant transfer learning between problem instances with similar vehicle and request distributions.

Deep ensembles can be seen as the current state-of-the-art for uncertainty quantification in deep learning. While the approach was originally proposed as an non-Bayesian technique, arguments towards its Bayesian footing have been put forward as well. We show that deep ensembles can be viewed as an approximate Bayesian method by specifying the corresponding assumptions. Our finding leads to an improved approximation which results in an increased epistemic part of the uncertainty. Numerical examples suggest that the improved approximation can lead to more reliable uncertainties. Analytical derivations ensure easy calculation of results.

Abstract--We describe how some problems (interpretability, lack of object-orientedness) of modern deep networks potentially could be solved by adapting a biologically plausible saccadic mechanism of perception. A sketch of such a saccadic vision model is proposed. Proof of concept experimental results are provided to support the proposed approach. One of the most human-readable representations of a visual Deep convolutional networks are often used today in applied scene is the semantic scene graph: if it is present, the task problems as one of the basic components of learning systems. of generating the text describing the scene is trivial [7]. The On some tasks, for example, the task of modeling faces, it is nodes of such a graph are usually nouns that name objects possible to achieve representations with good interpretability on the stage. The node can be assigned its coordinates on the [2].

We introduce AndroidEnv, an open-source platform for Reinforcement Learning (RL) research built on top of the Android ecosystem. AndroidEnv allows RL agents to interact with a wide variety of apps and services commonly used by humans through a universal touchscreen interface. Since agents train on a realistic simulation of an Android device, they have the potential to be deployed on real devices. In this report, we give an overview of the environment, highlighting the significant features it provides for research, and we present an empirical evaluation of some popular reinforcement learning agents on a set of tasks built on this platform.