Visual abstract reasoning tasks present challenges for deep neural networks, exposing limitations in their capabilities. In this work, we present a neural network model that addresses the challenges posed by Raven’s Progressive Matrices (RPM). Inspired by the two-stream hypothesis of visual processing, we introduce the Dual-stream Reasoning Network (DRNet), which utilizes two parallel branches to capture image features. On top of the two streams, a reasoning module first learns to merge the high-level features of the same image. Then, it employs a rule extractor to handle combinations involving the eight context images and each candidate image, extracting discrete abstract rules and utilizing an multilayer perceptron (MLP) to make predictions. Empirical results demonstrate that the proposed DRNet achieves state-of-the-art average performance across multiple RPM benchmarks. Furthermore, DRNet demonstrates robust generalization capabilities, even extending to various out-of-distribution scenarios. The dual streams within DRNet serve distinct functions by addressing local or spatial information. They are then integrated into the reasoning module, leveraging abstract rules to facilitate the execution of visual reasoning tasks. These findings indicate that the dual-stream architecture could play a crucial role in visual abstract reasoning.

Absolute pose regression has shown great potential in LiDAR localization, which learns to regress 6-DoF LiDAR poses through deep networks. However, recent regression methods suffer from scene ambiguities in challenging scenarios, leading to inaccurate and unstable localization. Inspired by neurobiological localization mechanisms, i.e., the firing mechanism of place cells, head-direction cells, and grid cells in mammalian brains, we propose a novel LiDAR localization framework called NIDALoc to achieve more robust and accurate results. First, we propose a Hebbian memory module, motivated by place cells, to preserve historical information, which helps refine local view features to reduce scene ambiguities. Specifically, the memory module stores scene information and then recalls it when revisiting an old place. Second, we propose a novel pose constrained framework, consisting of an orientation classification task and a grid center regression task, to regularize orientation and position estimation, respectively. The framework based on head-direction cells and grid cells constrains the absolute pose regression to reduce wrong predictions. Extensive experiments on two outdoor datasets demonstrate the effectiveness of NIDALoc, which outperforms state-of-the-art localization methods, especially in large-scale challenging scenes. The source code is available on the project website at https://github.com/PSYZ1234/NIDALoc.

How to transform a mixed flow of sensory and motor information into memory state of self-location and to build map representations of the environment are central questions in the navigation research. Studies in neuroscience have shown that place cells in the hippocampus of the rodent brains form dynamic cognitive representations of locations in the environment. We propose a neural-network model called sensory-motor integration network model (SeMINet) to learn cognitive map representations by integrating sensory and motor information while an agent is exploring a virtual environment. This biologically inspired model consists of a deep neural network representing visual features of the environment, a recurrent network of place units encoding spatial information by sensorimotor integration, and a secondary network to decode the locations of the agent from spatial representations. The recurrent connections between the place units sustain an activity bump in the network without the need of sensory inputs, and the asymmetry in the connections propagates the activity bump in the network, forming a dynamic memory state which matches the motion of the agent. A competitive learning process establishes the association between the sensory representations and the memory state of the place units, and is able to correct the cumulative path-integration errors. The simulation results demonstrate that the network forms neural codes that convey location information of the agent independent of its head direction. The decoding network reliably predicts the location even when the movement is subject to noise. The proposed SeMINet thus provides a brain-inspired neural-network model for cognitive map updated by both self-motion cues and visual cues.

In this paper, we develop a new classification method for manifold-valued data in the framework of probabilistic learning vector quantization. In many classification scenarios, the data can be naturally represented by symmetric positive definite matrices, which are inherently points that live on a curved Riemannian manifold. Due to the non-Euclidean geometry of Riemannian manifolds, traditional Euclidean machine learning algorithms yield poor results on such data. In this paper, we generalize the probabilistic learning vector quantization algorithm for data points living on the manifold of symmetric positive definite matrices equipped with Riemannian natural metric (affine-invariant metric). By exploiting the induced Riemannian distance, we derive the probabilistic learning Riemannian space quantization algorithm, obtaining the learning rule through Riemannian gradient descent. Empirical investigations on synthetic data, image data , and motor imagery electroencephalogram (EEG) data demonstrate the superior performance of the proposed method.

Visual place recognition in changing environments is a challenging and critical task for autonomous robot navigation. Deep convolutional neural networks (ConvNets) have recently been used as efficient feature extractors and obtained excellent performance in place recognition. However the success of ConvNets' learning highly relies on the availability of large datasets with millions of labeled images, the collection of which is a tedious and costly burden. Thus we develop an unsupervised learning method (the siamese VisNet) to autonomously learn invariant features in changing environments from unlabeled images. The siamese VisNet has two identical branches of sub-networks. With a Hebbian-type of learning rule incorporating a trace of previous activity patterns, the siamese VisNet learns features with increasing invariance in changing environments from layer to layer. Experiments conducted on multiple datasets demonstrate the robustness of the siamese VisNet against viewpoint changes, appearance changes, and joint viewpoint-appearance changes. In addition, the siamese VisNet, with lower complexity in architecture, outperforms the state-of-the-art place recognition ConvNets such as the CaffeNet and the PlaceNet. The proposed siamese VisNet constitutes a biologically plausible yet efficient method for unsupervised place recognition.

The problem of designing waveform and quantisation thresholds is studied in a multistatic radar setting, where distributed receivers are connected to a fusion centre via capacity constraints. Different from the previous cloud radio‐multistatic radar system which utilises an additive quantisation Gaussian noise model, a real quantisation system is designed. The authors first optimise the waveform without quantisation. Then they compress the received signal at receivers into a scalar without any performance degradation. Furthermore, the scalar quantiser is adopted and quantisation thresholds are designed. Numerical simulations are performed and the effectiveness of combined waveform and thresholds optimisation strategy is demonstrated.

We propose a neural network model for reinforcement learning to control a robotic manipulator with unknown parameters and dead zones. The model is composed of three networks. The state of the robotic manipulator is predicted by the state network of the model, the action policy is learned by the action network, and the performance index of the action policy is estimated by a critic network. The three networks work together to optimize the performance index based on the reinforcement learning control scheme. The convergence of the learning methods is analyzed. Application of the proposed model on a simulated two-link robotic manipulator demonstrates the effectiveness and the stability of the model.

Tracking is a common topic in various areas of robotics research. Motivated by the hunting behavior of predators in nature, we propose a prey-predator model for efficient robot tracking. The head direction and speed of the pursuer is automatically adjusted according to the position and velocity of the prey. Under the situation with perception uncertainty, where the actual location of the prey is not observable, the pursuer predicts the location of the prey according to simple inference, an online adaptive autoregressive model, or an online adaptive echo state network. Simulation results demonstrate that the proposed prey-predator model is able to control the pursuer and to track the prey efficiently, even under perception uncertainty. Simple inference gives better results when the motion of the target is piecewise linear, while echo state network is more suitable when the dynamics of the target are more complex. The proposed prey-predator model thus provides an efficient method tracking targets with various statistical nature of trajectories for applications such as underwater robot tracking, human tracking and team formation.

A local-autoencoding (LAE) method is proposed for the parameter estimation in a Hidden Potts-Markov random field model. Due to sampling cost, Markov chain Monte Carlo methods are rarely used in real-time applications. Like other heuristic methods, LAE is based on a conditional independence assumption. It adapts, however, the parameters in a block-by-block style with a simple Hebbian learning rule. Experiments with given label fields show that the LAE is able to converge in far less time than required for a scan. It is also possible to derive an estimate for LAE based on a Cramer–Rao bound that is similar to the classical maximum pseudolikelihood method. As a general algorithm, LAE can be used to estimate the parameters in anisotropic label fields. Furthermore, LAE is not limited to the classical Potts model and can be applied to other types of Potts models by simple label field transformations and straightforward learning rule extensions. Experimental results on image segmentations demonstrate the efficiency and generality of the LAE algorithm.

The entorhino-hippocampal neural circuits of the mammalian brains are able to internally generate efficient spatial representations of large-scale environments. Hippocampal principal cells in mammalian brains form sparse codes of the positions in large environments. The underlying computational mechanisms of the formation of spatial memory in large-scale environments are still not well understood. We present a competitive learning network that integrates both spatial inputs from grid cells as well as contextual inputs from non-spatial cells. Through self-organisation, place units in the network form place fields at “randomly” interleaved locations, in the sense that both the recruitment process and the inter-field interval are memoryless, a property superior for spatial memory formation in large environments.

The spatial responses of many of the cells recorded in layer II of rodent medial entorhinal cortex (MEC) show a triangular grid pattern, which appears to provide an accurate population code for animal spatial position. In layer III, V and VI of the rat MEC, grid cells are also selective to head-direction and are modulated by the speed of the animal. Several putative mechanisms of grid-like maps were proposed, including attractor network dynamics, interactions with theta oscillations or single-unit mechanisms such as firing rate adaptation. In this paper, we present a new attractor network model that accounts for the conjunctive position-by-velocity selectivity of grid cells. Our network model is able to perform robust path integration even when the recurrent connections are subject to random perturbations.