Categorization and decision-making in a neurobiologically plausible spiking network using a STDP-like learning rule

Understanding how the human brain is able to efficiently perceive and understand a visual scene is still a field of ongoing research. Although many studies have focused on the design and optimization of neural networks to solve visual recognition tasks, most of them either lack neurobiologically plausible learning rules or decision-making processes. Here we present a large-scale model of a hierarchical spiking neural network (SNN) that integrates a low-level memory encoding mechanism with a higher-level decision process to perform a visual classification task in real-time. The model consists of Izhikevich neurons and conductance-based synapses for realistic approximation of neuronal dynamics, a spike-timing-dependent plasticity (STDP) synaptic learning rule with additional synaptic dynamics for memory encoding, and an accumulator model for memory retrieval and categorization. The full network, which comprised 71,026 neurons and approximately 133 million synapses, ran in real-time on a single off-the-shelf graphics processing unit (GPU). The network was constructed on a publicly available SNN simulator that supports general-purpose neuromorphic computer chips. The network achieved 92% correct classifications on MNIST in 100 rounds of random sub-sampling, which is comparable to other SNN approaches and provides a conservative and reliable performance metric. Additionally, the model correctly predicted reaction times from psychophysical experiments. Because of the scalability of the approach and its neurobiological fidelity, the current model can be extended to an efficient neuromorphic implementation that supports more generalized object recognition and decision-making architectures found in the brain.     

On the Necessity of Recurrent Processing during Object Recognition: It Depends on the Need for Scene Segmentation

Although feedforward activity may suffice for recognizing objects in isolation, additional visual operations that aid object recognition might be needed for real-world scenes. One such additional operation is figure-ground segmentation, extracting the relevant features and locations of the target object while ignoring irrelevant features. In this study of 60 human participants (female and male), we show objects on backgrounds of increasing complexity to investigate whether recurrent computations are increasingly important for segmenting objects from more complex backgrounds. Three lines of evidence show that recurrent processing is critical for recognition of objects embedded in complex scenes. First, behavioral results indicated a greater reduction in performance after masking objects presented on more complex backgrounds, with the degree of impairment increasing with increasing background complexity. Second, electroencephalography (EEG) measurements showed clear differences in the evoked response potentials between conditions around time points beyond feedforward activity, and exploratory object decoding analyses based on the EEG signal indicated later decoding onsets for objects embedded in more complex backgrounds. Third, deep convolutional neural network performance confirmed this interpretation. Feedforward and less deep networks showed a higher degree of impairment in recognition for objects in complex backgrounds compared with recurrent and deeper networks. Together, these results support the notion that recurrent computations drive figure-ground segmentation of objects in complex scenes.SIGNIFICANCE STATEMENT The incredible speed of object recognition suggests that it relies purely on a fast feedforward buildup of perceptual activity. However, this view is contradicted by studies showing that disruption of recurrent processing leads to decreased object recognition performance. Here, we resolve this issue by showing that how object recognition is resolved and whether recurrent processing is crucial depends on the context in which it is presented. For objects presented in isolation or in simple environments, feedforward activity could be sufficient for successful object recognition. However, when the environment is more complex, additional processing seems necessary to select the elements that belong to the object and by that segregate them from the background.     

Self partitioning neural networks for target recognition

Automatic target recognition (ATR) is a domain in which the neural network technology has been applied with limited success. The domain is characterized by large training sets with dissimilar target images carrying conflicting information. This paper presents a novel method for quantifying the degree of non-cooperation that exists among the target members of the training set. Both the network architecture and the training algorithm are considered in the computation of the non-cooperation measures. Based on these measures, the self partitioning neural network (SPNN) approach partitions the target vectors into an appropriate number of groups and trains one subnetwork to recognize the targets in each group. A fusion network combines the outputs of the subnetworks to produce the final response. This method automatically determines the number of subnetworks needed without excessive computation. The subnetworks are simple with only one hidden layer and one unit in the output layer. They are topologically identical to one another. The simulation results indicate that the method is robust and capable of self organization to overcome the ill effects of the non-cooperating targets in the training set. The self partitioning approach improves the classification accuracy and reduces the training time of neural networks significantly. It is also shown that a trained self partitioning neural network is capable of learning new training vectors without retraining on the combined training set (i.e., the training set consisting of the previous and newly acquired training vectors).     

Learning-dependent Changes in the Responses to Visual Stimuli of Neurons in a Recognition Memory System

When young chicks are trained by exposing them to a conspicuous object they learn its characteristics. The learning process is known as imprinting. In the present study neuronal activity in a region crucial for imprinting was shown to be affected by training and by the object on which the chicks had been trained. The region is the intermediate and medial part of the left hyperstriatum ventrale (left IMHV). No such effects were found in a visual projection area, the left hyperstriatum accessorium. Domestic chicks were imprinted on either a rotating red box (n=7 chicks) or a rotating blue box (n=8). When the chicks were approximately 48 h old they were anaesthetized and multiple-unit activity was recorded in simultaneous, single penetrations through each of the two regions. Records were also made from eight dark-reared chicks. Whilst recording, the red or blue box, placed in front of the contralateral eye, was switched on to give a total of 20 rotations, the interval between each rotation being 10 s. The alternative stimulus was then presented 20 times. Unit activity in the 3 s before and after stimulus onset was compared and the data for each of the 20 presentations were combined. In the left IMHV 18 out of a total of 115 recording sites (16%) responded significantly to the stimuli; in the left hyperstriatum accessorium 39 out of 126 recording sites (26%) did so. Measures of unit activity at each recording site were combined for a given penetration to provide a 'mean penetration response'. The response to the red box differed from the response to the blue box in the left IMHV of dark-reared chicks. After training with the blue box the response to both boxes was similar to the response to the blue box in dark-reared birds. After training with the red box the response to both boxes was similar to the response to the red box in dark-reared birds. No significant effects were found in the left hyperstriatum accessorium. The two training boxes were virtually identical apart from the differences in colour and brightness. Training appeared to stabilize the response of the visually naive left IMHV to the training stimulus whilst changing its response to the alternative, but similar stimulus. That is, one consequence of training is that the two stimuli are placed in the same category, and this neural change may provide a basis for stimulus generalization. The underlying neural system is modelled and a mechanism that allows such stimuli to be discriminated is proposed.     

Recognition molecules myelin-associated glycoprotein and tenascin-C inhibit integrin-mediated adhesion of neural cells to collagen

Because of the importance of collagens in mediating cell-substrate interactions and the association of collagens with neural recognition molecules in the peripheral nervous system, the ability of neural recognition molecules to modify the substrate properties of collagens, in particular collagen type I, for cell adhesion was determined. Two cell lines, the N2A neuroblastoma and PC12 pheochromocytoma, were investigated for their capacity to adhere to different collagen types in the absence or presence of several neural recognition molecules. Adhesion of N2A or PC12 cells and membrane vesicles from PC12 cells to collagen type I was reduced when the collagen had been preincubated prior to its application as substrate with the extracellular domain of myelin-associated glycoprotein (s-MAG) or, as control, fibroblast tenascin-C (F-tenascin). In mixture with other collagen types, s-MAG was only able to reduce the adhesiveness of collagen types III and V, but not of collagen types II and IV. F-tenascin reduced the adhesiveness of all collagen types tested. In contrast to F-tenascin, s-MAG had to be present during fibrillogenesis to exert its effect, indicating that it must be coassembled into the collagen fibril to block the cell binding site. Cell adhesion to collagen type I was dependent on Mg²⁺ or Mn²⁺ and inhibited by a monoclonal antibody to the α₁ integrin subunit. The combined observations indicate that s-MAG and F-tenascin interfere with cell binding, most probably by modifying the integrin binding site, and that the two molecules act by different mechanisms, both leading to reduction of adhesion. © 1995 Wiley-Liss, Inc.     

Place learning and object recognition by rats subjected to transection of the fimbria-fornix and/or ablation of the prefrontal cortex

The acquisition of a water maze-based allocentric place learning task and an exploration based object recognition task were studied in four groups of rats: animals in which the fimbria-fornix had been transected, rats who had received bilateral ablations of the anteromedial prefrontal cortex, animals in which both of these structures had been lesioned, and a sham operated control group. None of the groups showed impairments of object recognition. Ablations of the prefrontal cortex caused a mild impairment in the acquisition of the place learning task. The two fimbria-fornix transected groups exhibited a severe impairment during the acquisition of this task. All groups reached criterion level task performance eventually. All groups were subjected to a number of behavioural and pharmacological challenges in order to elucidate the neural and cognitive mechanisms of this behavioural recovery. During a no-platform session both the fimbria-fornix transected group and the prefrontally ablated group demonstrated a normal preference for the former platform position. The combined lesion group, however, failed to show a similar preference for this position. The outcome of the pharmacological challenges demonstrated that while the task performance of all four groups relied equally on catecholaminergic mediation, only the task solution of the fimbria-fornix transected group was significantly impaired by disturbance of the catecholaminergic systems. The data indicated a high likelihood that prefrontal cortical mechanisms contribute to the recovery of allocentric place learning after fimbria-fornix transections.     

On-line Supervised Adaptive Training Using Radial Basis Function Networks

A new recursive supervised training algorithm is derived for the radial basis neural network architecture. The new algorithm combines the procedures of on-line candidate regressor selection with the conventional Givens QR based recursive parameter estimator to provide efficient adaptive supervised network training. A new concise on-line correlation based performance monitoring scheme is also introduced as an auxiliary device to detect structural changes in temporal data processing applications. Practical and simulated examples are included to demonstrate the effectiveness of the new procedures. Copyright © 1996 Elsevier Science Ltd.     

Application of deep learning to improve sleep scoring of wrist actigraphy

BackgroundEstimation of sleep parameters by wrist actigraphy is highly dependent on performance of the interpretative algorithm (IA) that converts movement data into sleep/wake scores.Research questions(1) Does the actigraphy mode of operation -Proportional Integrating Measure (PIM) or Zero Crossing Mode (ZCM), responsive respectively to intensity and frequency of movements- impact sleep scoring; and (2) Can a better performing sleep scoring IA be developed by a deep learning approach combining PIM/ZCM data.Study design and MethodsZCM and PIM plus electroencephalographic (EEG) data of 40 healthy adults (17 female, mean age: 26.7 years) were obtained from a single in-home nighttime sleep study. Effect of mode of operation was first evaluated by applying several classic deep learning models to PIM only, ZCM only, and combined ZCM/PIM data. After, a novel deep learning model was developed incorporating combined ZCM/PIM data, and its performance was compared with existing Cole-Kripke, rescored Cole-Kripke, Sadeh, and UCSD IAs.ResultsRelative to the EEG reference, ZCM/PIM combined mode produced higher agreement of scoring sleep/wake epochs than only ZCM or PIM modes. The proposed novel deep learning model showed 87.7% accuracy (0.2–1% higher than the other IAs), 94.1% sensitivity (0.7–4.3% lower than the other IAs), 64.0% specificity (9.9–21.5% higher than the other IAs), and 59.9% Kappa agreement (∼6.9–11.6% higher than other IAs) in detecting sleep epochs. The proposed deep learning model did not differ significantly from the reference EEG in estimating sleep onset latency (SOL), wake after sleep onset (WASO), total sleep time (TST), and sleep efficiency (SE). Amount of bias and minimum detectable change in estimating SOL, WASO, TST and SE by the deep learning model was smaller than other four IAs.InterpretationThe proposed novel deep learning algorithm simultaneously incorporating ZCM/PIM mode data performs significantly better in assessing sleep than existing conventional IAs.     

An improved time series prediction by applying the layer-by-layer learning method to FIR neural networks

The FIR neural network model was recently proposed for time series prediction and gave good results. However, the learning algorithm used for the FIR network is a kind of gradient descent method and hence inherits all the well-known problems of the method. Recently a new learning algorithm called the optimization layer by layer was proposed for the regular multilayer perceptron network, and showed a great improvement in the learning time as well as the performance of the network.

In this paper we develop a new learning algorithm for the FIR neural network model by applying the idea of the optimization layer by layer to the model. The results of the experiment, using two popular time series prediction problems, show that the new algorithm is far better in learning time and more accurate in prediction performance than the original learning algorithm.