Single Recurrent Network Research Articles

Single Image Deraining Using Bilateral Recurrent Network

Single image deraining has received considerable progress based on deep convolutional neural network (CNN). In existing deep deraining methods, CNNs are deployed to extract rain streaks while failing in learning direct mapping from rainy image to clean background image, and their architectures become more and more complicated. In this work, we first propose a single recurrent network (SRN) by recursively unfolding a shallow residual network, where a recurrent layer is adopted to propagate deep features across multiple stages. This simple SRN is effective not only in learning residual mapping for extracting rain streaks, but also in learning direct mapping for predicting clean background image. Furthermore, two SRNs are coupled to simultaneously exploit rain streak layer and clean background image layer. Instead of naive combination, we propose bilateral LSTMs, which not only can respectively propagate deep features of rain streak layer and background image layer across stages, but also bring the interplay between these two SRNs, finally forming bilateral recurrent network (BRN). The experimental results demonstrate that our BRN notably outperforms state-of-the-art deep deraining networks on synthetic datasets quantitatively and qualitatively. The proposed methods also perform more favorably in terms of generalization performance on real-world rainy dataset. All the source code and pre-trained models are available at https://github.com/csdwren/RecDerain.

Deterministic networks for probabilistic computing

Neuronal network models of high-level brain functions such as memory recall and reasoning often rely on the presence of some form of noise. The majority of these models assumes that each neuron in the functional network is equipped with its own private source of randomness, often in the form of uncorrelated external noise. In vivo, synaptic background input has been suggested to serve as the main source of noise in biological neuronal networks. However, the finiteness of the number of such noise sources constitutes a challenge to this idea. Here, we show that shared-noise correlations resulting from a finite number of independent noise sources can substantially impair the performance of stochastic network models. We demonstrate that this problem is naturally overcome by replacing the ensemble of independent noise sources by a deterministic recurrent neuronal network. By virtue of inhibitory feedback, such networks can generate small residual spatial correlations in their activity which, counter to intuition, suppress the detrimental effect of shared input. We exploit this mechanism to show that a single recurrent network of a few hundred neurons can serve as a natural noise source for a large ensemble of functional networks performing probabilistic computations, each comprising thousands of units.

Neural Dynamics Indicate Parallel Integration of Environmental and Self-Motion Information by Place and Grid Cells.

Place cells and grid cells in the hippocampal formation are thought to integrate sensory and self-motion information into a representation of estimated spatial location, but the precise mechanism is unknown. We simulated a parallel attractor system in which place cells form an attractor network driven by environmental inputs and grid cells form an attractor network performing path integration driven by self-motion, with inter-connections between them allowing both types of input to influence firing in both ensembles. We show that such a system is needed to explain the spatial patterns and temporal dynamics of place cell firing when rats run on a linear track in which the familiar correspondence between environmental and self-motion inputs is changed. In contrast, the alternative architecture of a single recurrent network of place cells (performing path integration and receiving environmental inputs) cannot reproduce the place cell firing dynamics. These results support the hypothesis that grid and place cells provide two different but complementary attractor representations (based on self-motion and environmental sensory inputs, respectively). Our results also indicate the specific neural mechanism and main predictors of hippocampal map realignment and make predictions for future studies.

In-air handwritten Chinese text recognition with temporal convolutional recurrent network

As a new human-computer interaction way, in-air handwriting allows users to perform gesture-based writing in the midair. However, most existing in-air handwriting systems mainly focus on recognizing either isolated characters/words or only a small number of texts, making those systems far from practical applications. Instead, here we present a 3D in-air handwritten Chinese text recognition (IAHCTR) system for the first time, and construct the first public large-scale IAHCT dataset. Moreover, a novel architecture, named the temporal convolutional recurrent network (TCRN), is proposed for online HCTR. Specifically, the TCRN first applies the 1-dimensional convolution to extract local contextual features from low-level trajectories, and then it utilizes the recurrent network to capture long-term dependencies of high-level outputs. Compared with the state-of-the-art architecture, the TCRN not only avoids the domain-specific knowledge for feature image extraction, but also attains higher training efficiency with a more compact model. Empirically, this TCRN also outperforms the single recurrent network with faster prediction and higher accuracy. Experiments on CASIA-OLHWDB2 & ICDAR-2013 demonstrate that the TCRN yields the best result in comparison to the state-of-the-art methods for online HCTR.

Delay Selection by Spike-Timing-Dependent Plasticity in Recurrent Networks of Spiking Neurons Receiving Oscillatory Inputs

Learning rules, such as spike-timing-dependent plasticity (STDP), change the structure of networks of neurons based on the firing activity. A network level understanding of these mechanisms can help infer how the brain learns patterns and processes information. Previous studies have shown that STDP selectively potentiates feed-forward connections that have specific axonal delays, and that this underlies behavioral functions such as sound localization in the auditory brainstem of the barn owl. In this study, we investigate how STDP leads to the selective potentiation of recurrent connections with different axonal and dendritic delays during oscillatory activity. We develop analytical models of learning with additive STDP in recurrent networks driven by oscillatory inputs, and support the results using simulations with leaky integrate-and-fire neurons. Our results show selective potentiation of connections with specific axonal delays, which depended on the input frequency. In addition, we demonstrate how this can lead to a network becoming selective in the amplitude of its oscillatory response to this frequency. We extend this model of axonal delay selection within a single recurrent network in two ways. First, we show the selective potentiation of connections with a range of both axonal and dendritic delays. Second, we show axonal delay selection between multiple groups receiving out-of-phase, oscillatory inputs. We discuss the application of these models to the formation and activation of neuronal ensembles or cell assemblies in the cortex, and also to missing fundamental pitch perception in the auditory brainstem.

Adaptive behavior from fixed weight networks

An attribute often associated with intelligent systems is the capability of exhibiting adaptive behavior. For the purposes of this paper, we define adaptation as a system's ability to recognize change through its sensed inputs, and to approximately adjust its behavior in response to the perceived change. This paper explores the notion that a fixed weight, time-lagged recurrent network architecture can be made to exhibit adaptive behavior in the manner defined above after network training has been completed, i.e., to exhibit adaptation in the absence of an explicit learning mechanism. We describe in this paper network training procedures that enable the learning of adaptive behaviors, and we provide empirical evidence of the adaptive capability for a single recurrent network that has been trained to perform one-time-step prediction for any one of several distinct time series.