M. Smoluchowski Institute of Physics and Mark Kac Complex Systems Research Center, Jagiellonian University, PL-30-348 Kraków, Poland (Dated: September 25, 2018) We demonstrate that in residual neural networks (ResNets) dynamical isometry is achievable irrespectively of the activation function used. We do that by deriving, with the help of Free Probability and Random Matrix Theories, a universal formula for the spectral density of the input-output Jacobian at initialization, in the large network width and depth limit. The resulting singular value spectrum depends on a single parameter, which we calculate for a variety of popular activation functions, by analyzing the signal propagation in the artificial neural network. We corroborate our results with numerical simulations of both random matrices and ResNets applied to the CIFAR-10 classification problem. Moreover, we study the consequence of this universal behavior for the initial and late phases of the learning processes. We conclude by drawing attention to the simple fact, that initialization acts as a confounding factor between the choice of activation function and the rate of learning. We propose that in ResNets this can be resolved based on our results, by ensuring the same level of dynamical isometry at initialization. Deep Learning has achieved unparalleled success in fields such as object detection and recognition, language translation and speech recognition [2].

Deep neural networks (DNNs) generate much richer function spaces than shallow networks. Since the function spaces induced by shallow networks have several approximation theoretic drawbacks, this explains, however, not necessarily the success of deep networks. In this article we take another route by comparing the expressive power of DNNs with ReLU activation function to piecewise linear spline methods. We show that MARS (multivariate adaptive regression splines) is improper learnable by DNNs in the sense that for any given function that can be expressed as a function in MARS with $M$ parameters there exists a multilayer neural network with $O(M \log (M/\varepsilon))$ parameters that approximates this function up to sup-norm error $\varepsilon.$ We show a similar result for expansions with respect to the Faber-Schauder system. Based on this, we derive risk comparison inequalities that bound the statistical risk of fitting a neural network by the statistical risk of spline-based methods. This shows that deep networks perform better or only slightly worse than the considered spline methods. We provide a constructive proof for the function approximations.

The generative learning phase of Autoencoder (AE) and its successor Denosing Autoencoder (DAE) enhances the flexibility of data stream method in exploiting unlabelled samples. Nonetheless, the feasibility of DAE for data stream analytic deserves in-depth study because it characterizes a fixed network capacity which cannot adapt to rapidly changing environments. An automated construction of a denoising autoeconder, namely deep evolving denoising autoencoder (DEVDAN), is proposed in this paper. DEVDAN features an open structure both in the generative phase and in the discriminative phase where input features can be automatically added and discarded on the fly. A network significance (NS) method is formulated in this paper and is derived from the bias-variance concept. This method is capable of estimating the statistical contribution of the network structure and its hidden units which precursors an ideal state to add or prune input features. Furthermore, DEVDAN is free of the problem- specific threshold and works fully in the single-pass learning fashion. The efficacy of DEVDAN is numerically validated using nine non-stationary data stream problems simulated under the prequential test-then-train protocol where DEVDAN is capable of delivering an improvement of classification accuracy to recently published online learning works while having flexibility in the automatic extraction of robust input features and in adapting to rapidly changing environments.

Deep learning architectures have proved versatile in a number of drug discovery applications, including the modelling of in vitro compound activity. While controlling for prediction confidence is essential to increase the trust, interpretability and usefulness of virtual screening models in drug discovery, techniques to estimate the reliability of the predictions generated with deep learning networks remain largely underexplored. Here, we present Deep Confidence, a framework to compute valid and efficient confidence intervals for individual predictions using the deep learning technique Snapshot Ensembling and conformal prediction. Specifically, Deep Confidence generates an ensemble of deep neural networks by recording the network parameters throughout the local minima visited during the optimization phase of a single neural network. This approach serves to derive a set of base learners (i.e., snapshots) with comparable predictive power on average, that will however generate slightly different predictions for a given instance. The variability across base learners and the validation residuals are in turn harnessed to compute confidence intervals using the conformal prediction framework. Using a set of 24 diverse IC50 data sets from ChEMBL 23, we show that Snapshot Ensembles perform on par with Random Forest (RF) and ensembles of independently trained deep neural networks. In addition, we find that the confidence regions predicted using the Deep Confidence framework span a narrower set of values. Overall, Deep Confidence represents a highly versatile error prediction framework that can be applied to any deep learning-based application at no extra computational cost.

Responses generated by neural conversational models tend to lack informativeness and diversity. We present Adversarial Information Maximization (AIM), an adversarial learning strategy that addresses these two related but distinct problems. To foster response diversity, we leverage adversarial training that allows distributional matching of synthetic and real responses. To improve informativeness, our framework explicitly optimizes a variational lower bound on pairwise mutual information between query and response. Empirical results from automatic and human evaluations demonstrate that our methods significantly boost informativeness and diversity.

There are time series that are amenable to recurrent neural network (RNN) solutions when treated as sequences, but some series, e.g. In order to address such situations, we introduce a unified RNN that handles five different feature types, each in a different manner. Our RNN framework separates sequential features into two groups dependent on their frequency, which we call sparse and dense features, and which affect cell updates differently. Further, we also incorporate time features at the sequential level that relate to the time between specified events in the sequence and are used to modify the cell's memory state. We also include two types of static (whole sequence level) features, one related to time and one not, which are combined with the encoder output. The experiments show that the modeling framework proposed does increase performance compared to standard cells. The study of time series has a long history and the literature for it covers many different methods (Hamilton (1994)). The study of asynchronous time series is an important subset of this. Asynchronous time series are series for which features are sampled at irregular time intervals, and at any given time step new values of any subset of features may be present. When a feature does not change values often it can be treated as being present only at times of change.

Computer-aided diagnosis systems for classification of different type of skin lesions have been an active field of research in recent decades. It has been shown that introducing lesions and their attributes masks into lesion classification pipeline can greatly improve the performance. In this paper, we propose a framework by incorporating transfer learning for segmenting lesions and their attributes based on the convolutional neural networks. The proposed framework is inspired by the well-known UNet architecture. It utilizes a variety of pre-trained networks in the encoding path and generates the prediction map by combining multi-scale information in decoding path using a pyramid pooling manner. To circumvent the lack of training data and increase the proposed model generalization, an extensive set of novel augmentation routines have been applied during the training of the network. Moreover, for each task of lesion and attribute segmentation, a specific loss function has been designed to obviate the training phase difficulties. Finally, the prediction for each task is generated by ensembling the outputs from different models. The proposed approach achieves promising results on the cross-validation experiments on the ISIC2018- Task1 and Task2 data sets.

Abstract--We propose a tool-use model that can detect the features of tools, target objects, and actions from the provided effects of object manipulation. We construct a model that enables robots to manipulate objects with tools, using infant learning as a concept. To realize this, we train sensory-motor data recorded during a tool-use task performed by a robot with deep learning. Experiments include four factors: (1) tools, (2) objects, (3) actions, and (4) effects, which the model considers simultaneously. For evaluation, the robot generates predicted images and motions given information of the effects of using unknown tools and objects. We confirm that the robot is capable of detecting features of tools, objects, and actions by learning the effects and executing the task. I. Introduction In recent years, robots have become part of human living space and have been expected to perform various tasks in complex environments. If robots could use tools as humans do, they could improve in versatility, overcome some physical limitations and adapt to the environment.

Interpretability is a key factor in the design of automatic classifiers for medical diagnosis. Deep learning models have been proven to be a very effective classification algorithm when trained in a supervised way with enough data. The main concern is the difficulty of inferring rationale interpretations from them. Different attempts have been done in last years in order to convert deep learning classifiers from high confidence statistical black box machines into self-explanatory models. In this paper we go forward into the generation of explanations by identifying the independent causes that use a deep learning model for classifying an image into a certain class. We use a combination of Independent Component Analysis with a Score Visualization technique. In this paper we study the medical problem of classifying an eye fundus image into 5 levels of Diabetic Retinopathy. We conclude that only 3 independent components are enough for the differentiation and correct classification between the 5 disease standard classes. We propose a method for visualizing them and detecting lesions from the generated visual maps.

Previous transfer learning methods based on deep network assume the knowledge should be transferred between the same hidden layers of the source domain and the target domains. This assumption doesn't always hold true, especially when the data from the two domains are heterogeneous with different resolutions. In such case, the most suitable numbers of layers for the source domain data and the target domain data would differ. As a result, the high level knowledge from the source domain would be transferred to the wrong layer of target domain. Based on this observation, "where to transfer" proposed in this paper should be a novel research frontier. We propose a new mathematic model named DT -LET to solve this heterogeneous transfer learning problem. In order to select the best matching of layers to transfer knowledge, we define specific loss function to estimate the corresponding relationship between high-level features of data in the source domain and the target domain. To verify this proposed cross-layer model, experiments for two cross-domain recognition/classification tasks are conducted, and the achieved superior results demonstrate the necessity of layer correspondence searching.