Since the popularization of deep neural networks in the early 2010s, tailoring neural network architectures to specific tasks has been one of the main sources of activity for both academics and practitioners. Accordingly, a palette of empirical methods has been developed for automating the choice of neural networks hyperparameters (a process sometimes called Neural Architecture Search), including - but not limited to - random search [2, 1], genetic algorithms [16, 13], bayesian methods [24, 12] or reinforcement learning [29]. However, when the computational requirements for training a single model are high, such approaches might be too expensive or result in iteration cycles that are too long to be practically useful - though some work in that direction has been carried out recently [5, 14]. In other cases, when the loss function is only used as a proxy for the task at hand [25, 26, 10] or is not interpretable [8], a further perceptual evaluation is typically necessary to evaluate the quality of a model's outputs and such systematic approaches at least partially break down. In both cases, an efficient and quantitative method to analyze and compare neural network architectures would be highly desirable - be it only to come up with a limited set of plausible candidates to pass on to the more expensive (or manual) methods. In this paper, we introduce the notion of capacity allocation analysis, which is a systematic, quantitative andcomputationally efficient way to analyze neural network architectures by quantifying which dependencies between inputs and outputs a parameter of a set of parameters actually model. We develop a quantitative framework for assessing and comparing different architectures for a given task, providing insights that are complementary to the value of the loss function itself.

We examine the practice of joint training for neural network ensembles, in which a multi-branch architecture is trained via single loss. This approach has recently gained traction, with claims of greater accuracy per parameter along with increased parallelism. We introduce a family of novel loss functions generalizing multiple previously proposed approaches, with which we study theoretical and empirical properties of joint training. These losses interpolate smoothly between independent and joint training of predictors, demonstrating that joint training has several disadvantages not observed in prior work. However, with appropriate regularization via our proposed loss, the method shows new promise in resource limited scenarios and fault-tolerant systems, e.g., IoT and edge devices. Finally, we discuss how these results may have implications for general multi-branch architectures such as ResNeXt and Inception.

Variational inference with a factorized Gaussian posterior estimate is a widely used approach for learning parameters and hidden variables. Empirically, a regularizing effect can be observed that is poorly understood. In this work, we show how mean field inference improves generalization by limiting mutual information between learned parameters and the data through noise. We quantify a maximum capacity when the posterior variance is either fixed or learned and connect it to generalization error, even when the KL-divergence in the objective is rescaled. Our experiments demonstrate that bounding information between parameters and data effectively regularizes neural networks on both supervised and unsupervised tasks.

Abstract--In this paper, we present a framework to control a self-driving car by fusing raw information from RGB images and depth maps. A deep neural network architecture is used for mapping the vision and depth information, respectively, to steering commands. This fusion of information from two sensor sources allows to provide redundancy and fault tolerance in the presence of sensor failures. Even if one of the input sensors fails to produce the correct output, the other functioning sensor would still be able to maneuver the car. Such redundancy is crucial in the critical application of self-driving cars. The experimental results have showed that our method is capable of learning to use the relevant sensor information even when one of the sensors fail without any explicit signal. I. INTRODUCTION The fusion of different sensor modalities in the context of autonomous driving is a crucial aspect in order to be robust against sensor failures. We consider RGB and Depth camera sensors for training a control module to maneuver a self-driving car.

To widen their accessibility and increase their utility, intelligent agents must be able to learn complex behaviors as specified by (non-expert) human users. Moreover, they will need to learn these behaviors within a reasonable amount of time while efficiently leveraging the sparse feedback a human trainer is capable of providing. Recent work has shown that human feedback can be characterized as a critique of an agent's current behavior rather than as an alternative reward signal to be maximized, culminating in the COnvergent Actor-Critic by Humans (COACH) algorithm for making direct policy updates based on human feedback. Our work builds on COACH, moving to a setting where the agent's policy is represented by a deep neural network. We employ a series of modifications on top of the original COACH algorithm that are critical for successfully learning behaviors from high-dimensional observations, while also satisfying the constraint of obtaining reduced sample complexity. We demonstrate the effectiveness of our Deep COACH algorithm in the rich 3D world of Minecraft with an agent that learns to complete tasks by mapping from raw pixels to actions using only real-time human feedback in 10-15 minutes of interaction.

In sparse-view Computed Tomography (CT), only a small number of projection images are taken around the object, and sinogram interpolation method has a significant impact on final image quality. When the amount of sparsity (the amount of missing views in sinogram data) is not high, conventional interpolation methods have yielded good results. When the amount of sparsity is high, more advanced sinogram interpolation methods are needed. Recently, several deep learning (DL) based sinogram interpolation methods have been proposed. However, those DL-based methods have mostly tested so far on computer simulated sinogram data rather experimentally acquired sinogram data. In this study, we developed a sinogram interpolation method for sparse-view micro-CT based on the combination of U-Net and residual learning. We applied the method to sinogram data obtained from sparse-view micro-CT experiments, where the sparsity reached 90%. The interpolated sinogram by the DL neural network was fed to FBP algorithm for reconstruction. The result shows that both RMSE and SSIM of CT image are greatly improved. The experimental results demonstrate that this sinogram interpolation method produce significantly better results over standard linear interpolation methods when the sinogram data are extremely sparse.

Learning automatically the best activation function for the task is an active topic in neural network research. At the moment, despite promising results, it is still difficult to determine a method for learning an activation function that is at the same time theoretically simple and easy to implement. Moreover, most of the methods proposed so far introduce new parameters or adopt different learning techniques. In this work we propose a simple method to obtain trained activation function which adds to the neural network local subnetworks with a small amount of neurons. Experiments show that this approach could lead to better result with respect to using a pre-defined activation function, without introducing a large amount of extra parameters that need to be learned.

Understanding learning and generalization of deep architectures has been a major research objective in the recent years with notable theoretical progress. A main focal point of generalization studies stems from the success of excessively large networks which defy the classical wisdom of uniform convergence and learnability. We study empirically the layer-wise functional structure of overparameterized deep models. We provide evidence for the heterogeneous characteristic of layers. To do so, we introduce the notion of (post training) re-initialization and re-randomization robustness. We show that layers can be categorized into either `robust' or `critical'. In contrast to critical layers, resetting the robust layers to their initial value has no negative consequence, and in many cases they barely change throughout training. Our study provides further evidence that mere parameter counting or norm accounting is too coarse in studying generalization of deep models, and flatness or robustness analysis of the model parameters needs to respect the network architectures.

Recently, Graph Neural Networks (GNNs) are trending in the machine learning community as a family of architectures that specializes in capturing the features of graph-related datasets, such as those pertaining to social networks and chemical structures. Unlike for other families of the networks, the representation power of GNNs has much room for improvement, and many graph networks to date suffer from the problem of underfitting. In this paper we will introduce a Graph Warp Module, a supernode-based auxiliary network module that can be attached to a wide variety of existing GNNs in order to improve the representation power of the original networks. Through extensive experiments on molecular graph datasets, we will show that our GWM indeed alleviates the underfitting problem for various existing networks, and that it can even help create a network with the state-of-the-art generalization performance.

Although recurrent neural networks (RNNs) for reinforcement learning (RL) have addressed unique advantages in various aspects, e. g., solving memory-dependent tasks and meta-learning, very few studies have demonstrated how RNNs can solve the problem of hierarchical RL by autonomously developing hierarchical control. In this paper, we propose a novel model-free RL framework called ReMASTER, which combines an off-policy actor-critic algorithm with a multiple timescale stochastic recurrent neural network for solving memory-dependent and hierarchical tasks. We performed experiments using a challenging continuous control task and showed that: (1) Internal representation necessary for achieving hierarchical control autonomously develops through exploratory learning. (2) Stochastic neurons in RNNs enable faster relearning when adapting to a new task which is a recomposition of sub-goals previously learned.