Neural Computation Research Articles

On-Chip Photonic Synapses with All-Optical Memory and Neural Network Computation.

Inspired by the human brain, neural network computing was expected to break the bottleneck of traditional computing, but the integrated design still faces great challenges. Here, a readily integrated membrane-system photonic synapse was demonstrated. By pre-pulse training at 1064 nm (cutoff wavelength), the photonic synapse can be regulated both excitatory and inhibitory at tunable wavelengths (1200-2000 nm). Furthermore, more weights and memory functions were shown through the photonic synapse integrated network. Additionally, the digital recognition function of the single-layer perceptron neural network constructed by photonic synapses has been successfully demonstrated. Most of the biological synaptic functions were realized by the photonic synaptic network, and it had the advantages of compact structure, scalable, adjustable wavelength, and so on, which opens up a new idea for the study of the neural synaptic network.

Origin of the efficiency of spike timing-based neural computation for processing temporal information

Although the advantage of spike timing-based over rate-based network computation has been recognized, the underlying mechanism remains unclear. Using Tempotron and Perceptron as elementary neural models, we examined the intrinsic difference between spike timing-based and rate-based computations. For more direct comparison, we modified Tempotron computation into rate-based computation with the retention of some temporal information. Previous studies have shown that spike timing-based computation are computationally more powerful than rate-based computation in terms of the number of computational units required and the capability in classifying random patterns. Our study showed that spike timing-based and rate-based Tempotron computations provided similar capability in classifying random spike patterns, as well as in text sentiment classification and spam text detection. However, spike timing-based computation is superior in performing a task involving discriminating forward vs. reverse sequence of events, i.e., information mainly temporal in nature. Further studies revealed that this superiority required the asymmetry in the profile of the postsynaptic potential (PSP), and that temporal sequence information was converted to biased spatial distribution of synaptic weight modifications during learning. Thus, the intrinsic PSP asymmetry is a mechanistic basis for the high efficiency of spike timing-based computation for processing temporal information.

A Comparative Study of Ensemble Techniques Based on Genetic Programming: A Case Study in Semantic Similarity Assessment

The challenge of assessing semantic similarity between pieces of text through computers has attracted considerable attention from industry and academia. New advances in neural computation have developed very sophisticated concepts, establishing a new state of the art in this respect. In this paper, we go one step further by proposing new techniques built on the existing methods. To do so, we bring to the table the stacking concept that has given such good results and propose a new architecture for ensemble learning based on genetic programming. As there are several possible variants, we compare them all and try to establish which one is the most appropriate to achieve successful results in this context. Analysis of the experiments indicates that Cartesian Genetic Programming seems to give better average results.

Classification of motor imagery electroencephalogram signals by using adaptive cross-subject transfer learning.

Electroencephalogram (EEG)-based motor imagery (MI) classification is an important aspect in brain-computer interfaces (BCIs), which bridges between neural system and computer devices decoding brain signals into recognizable machine commands. However, due to the small number of training samples of MI electroencephalogram (MI-EEG) for a single subject and the great individual differences of MI-EEG among different subjects, the generalization and accuracy of the model on the specific MI task may be poor. To solve these problems, an adaptive cross-subject transfer learning algorithm is proposed, which is based on kernel mean matching (KMM) and transfer learning adaptive boosting (TrAdaBoost) method. First, the common spatial pattern (CSP) is used to extract the spatial features. Then, in order to make the feature distribution more similar among different subjects, the KMM algorithm is used to compute a sample weight matrix for aligning the mean between source and target domains and reducing distribution differences among different subjects. Finally, the sample weight matrix from KMM is used as the initialization weight of TrAdaBoost, and then TrAdaBoost is used to adaptively select source domain samples that are closer to the target task distribution to assist in building a classification model. In order to verify the effectiveness and feasibility of the proposed method, the algorithm is applied to BCI Competition IV datasets and in-house datasets. The results show that the average classification accuracy of the proposed method on the public datasets is 89.1%, and the average classification accuracy on the in-house datasets is 80.4%. Compared with the existing methods, the proposed method effectively improves the classification accuracy of MI-EEG signals. At the same time, this paper also applies the proposed algorithm to the in-house dataset, the results verify the effectiveness of the algorithm again, and the results of this study have certain clinical guiding significance for brain rehabilitation.

Use of Neural Networks and Computer Vision for Spill and Waste Detection in Port Waters: An Application in the Port of Palma (MaJorca, Spain)

Water quality and pollution is the main environmental concern for ports and adjacent coastal waters. Therefore, the development of Port Environmental Management systems often relies on water pollution monitoring. Computer vision is a powerful and versatile tool for an exhaustive and systematic monitoring task. An investigation has been conducted at the Port of Palma de Mallorca (Spain) to assess the feasibility and evaluate the main opportunities and difficulties of the implementation of water pollution monitoring based on computer vision. Experiments on surface slicks and marine litter identification based on random image sets have been conducted. The reliability and development requirements of the method have been evaluated, concluding that computer vision is suitable for these monitoring tasks. Several computer vision techniques based on convolutional neural networks were assessed, finding that Image Classification is the most adequate for marine pollution monitoring tasks due to its high accuracy rates and low training requirements. Image set size for initial training and the possibility to improve accuracy through retraining with increased image sets were considered due to the difficulty in obtaining port spill images. Thus, we have found that progressive implementation can not only offer functional monitoring systems in a shorter time frame but also reduce the total development cost for a system with the same accuracy level.

Artificial neural networks and computer vision in medicine and surgery

Introduction: Artificial neural networks are becoming an essential technology in data analysis, and their influence is starting to permeate the field of medicine. Experimental surgery has been a long-term subject of study of our lab; this is naturally reflected in our interest in other areas of modern technologies including artificial neural networks and their advancements. In the current issue, we would like to explore this aspect of technical progress. The main goal is to critically evaluate the strengths and weaknesses of artificial neural network technology concerning its use in clinical and experimental surgery. Methods: The article is focused on in-silico modeling, particularly on the potential of neural networks in terms of image data processing in medicine. The text briefly summarizes the historical development of deep learning neural networks and their basic principles. Furthermore, basic taxonomy tasks are presented. Finally, potential learning problems and possible solutions are also mentioned. Results: The article points out various possible uses of artificial neural networks in biological applications. Several biomedical applications of artificial neural networks are used to describe the division and principles of the most common tasks of machine learning and deep learning such as classification, detection, and segmentation. Conclusion: The application of artificial neural network methods in medicine and surgery offers a considerable potential; by learning directly from the data, they make it possible to avoid lengthy and subjective setting of parameters by an expert engineer. Nevertheless, the use of an unbalanced dataset can lead to unexpected, although traceable errors. The solution is to collect a dataset large enough to enable both learning and verification of proper functionality.

Retracted: Convolutional Neural Network and Computer Vision-Based Action Correction Method for Long-Distance Running Technology

Retracted: Convolutional Neural Network and Computer Vision-Based Action Correction Method for Long-Distance Running Technology

Visual motion perception as online hierarchical inference

Identifying the structure of motion relations in the environment is critical for navigation, tracking, prediction, and pursuit. Yet, little is known about the mental and neural computations that allow the visual system to infer this structure online from a volatile stream of visual information. We propose online hierarchical Bayesian inference as a principled solution for how the brain might solve this complex perceptual task. We derive an online Expectation-Maximization algorithm that explains human percepts qualitatively and quantitatively for a diverse set of stimuli, covering classical psychophysics experiments, ambiguous motion scenes, and illusory motion displays. We thereby identify normative explanations for the origin of human motion structure perception and make testable predictions for future psychophysics experiments. The proposed online hierarchical inference model furthermore affords a neural network implementation which shares properties with motion-sensitive cortical areas and motivates targeted experiments to reveal the neural representations of latent structure.

Electret-Based Vertical Organic Synaptic Transistor With MXene for Neural Network-Based Computation

Organic synaptic transistors with excellent solution processability and biocompatibility have emerged as artificial electronic synapses. Regular organic synaptic transistors suffer from slight conductance variation and asymmetric conductance tuning, limiting the development of the model perception accuracy of the organic neuromorphic systems. Here, we first develop an electret-based vertical organic synaptic transistor (EVOST) with Mxene as the source electrode. The EVOST achieves linear conductance tuning by leveraging the nanoscale carrier transport channel length and high conductivity of MXene. Moreover, we develop an artificial neural recognition system composed of EVOSTs for recognizing the random images from the database, which achieves outstanding perception accuracy of 94.9%. The EVOST provides an alternative way for neuromorphic computing networks.

Quiet Trajectories as Neural Building Blocks.

Our concept of the neural mechanisms of working memory has recently undergone an upheaval, because of two transformative concepts: multivariate neural state trajectories and the activity-silent hypothesis. I will argue that putting these concepts together raises the difficult problem of "quiet trajectories," where future neural activity is not fully determined by current activity. However, this also promises new building blocks for neural computation.

Dynamics Are the Only Constant in Working Memory.

In this short perspective, we reflect upon our tendency to use oversimplified and idiosyncratic tasks in a quest to discover general mechanisms of working memory. We discuss how the work of Mark Stokes and collaborators has looked beyond localized, temporally persistent neural activity and shifted focus toward the importance of distributed, dynamic neural codes for working memory. A critical lesson from this work is that using simplified tasks does not automatically simplify the neural computations supporting behavior (even if we wish it were so). Moreover, Stokes' insights about multidimensional dynamics highlight the flexibility of the neural codes underlying cognition and have pushed the field to look beyond static measures of working memory.

Fractional Spiking Neuron: Fractional Leaky Integrate-and-Fire Circuit Described with Dendritic Fractal Model.

As dendrites are essential parts of neurons, they are crucial factors for neuronal activities to follow multiple timescale dynamics, which ultimately affect information processing and cognition. However, in the common SNN (Spiking Neural Networks), the hardware-based LIF (Leaky Integrate-and-Fire) circuit only simulates the single timescale dynamic of soma without relating dendritic morphologies, which may limit the capability of simulating neurons to process information. This study proposes the dendritic fractal model mainly for quantifying dendritic morphological effects containing branch and length. To realize this model, We design multiple analog fractional-order circuits (AFCs) which match their extended structures and parameters with the dendritic features. Then introducing AFC into FLIF (Fractional Leaky Integrate-and-Fire) neuron circuits can demonstrate the same multiple timescale dynamics of spiking patterns as biological neurons, including spiking adaptation, inter-spike variability with power-law distribution, first-spike latency, and intrinsic memory. By contrast, it further enhances the degree of mimicry of neuron models and provides a more accurate model for understanding neural computation and cognition mechanisms.

Server load and network-aware adaptive deep learning inference offloading for edge platforms

This work presents DIAMOND, a deep neural network computation offloading scheme consisting of a lightweight client-to-server latency profiling component combined with a server inference time estimation module to accurately assess the expected latency of a deep learning model inference. Latency predictions for both the network and server are comprehensively used to make dynamic (partial) model offloading decisions at the client in run-time. Compared to previous work, DIAMOND targets to minimize network latency estimation overhead and considers the concurrent processing nature of state-of-the-art deep learning inference server designs. Our extensive evaluations with an NVIDIA Jetson Nano client connected to an NVIDIA Triton server shows that DIAMOND completes inference operations with noticeably reduced computational/energy overhead and latency compared to previously proposed model offloading approaches. Furthermore, our results show that DIAMOND well-adapts to practical server load and network dynamics.

Dissociation and hierarchy of human visual pathways for simultaneously coding facial identity and expression

Humans have an extraordinary ability to recognize facial expression and identity from a single face simultaneously and effortlessly, however, the underlying neural computation is not well understood. Here, we optimized a multi-task deep neural network to classify facial expression and identity simultaneously. Under various optimization training strategies, the best-performing model consistently showed ‘share-separate’ organization. The two separate branches of the best-performing model also exhibited distinct abilities to categorize facial expression and identity, and these abilities increased along the facial expression or identity branches toward high layers. By comparing the representational similarities between the best-performing model and functional magnetic resonance imaging (fMRI) responses in the human visual cortex to the same face stimuli, the face-selective posterior superior temporal sulcus (pSTS) in the dorsal visual cortex was significantly correlated with layers in the expression branch of the model, and the anterior inferotemporal cortex (aIT) and anterior fusiform face area (aFFA) in the ventral visual cortex were significantly correlated with layers in the identity branch of the model. Besides, the aFFA and aIT better matched the high layers of the model, while the posterior FFA (pFFA) and occipital facial area (OFA) better matched the middle and early layers of the model, respectively. Overall, our study provides a task-optimization computational model to better understand the neural mechanism underlying face recognition, which suggest that similar to the best-performing model, the human visual system exhibits both dissociated and hierarchical neuroanatomical organization when simultaneously coding facial identity and expression.

Objectivization of virus titration method in GMP-regulated environment using artificial intelligence-based classification system

The implementation of artificial intelligence-based systems (AI) into highly regulated industries faces significant regulatory and logistical challenges. The lack of established practices and guidelines, risk aversion attitude, perception of AI technology as insufficiently tested and unreliable are among the most important factors preventing wider adoption of AI solutions into pharmaceutical industry. Here we demonstrate by an example development and validation of neural network-based computer vision systems for binary classification of results from virus titration assays often used in research and production of vaccines and antivirals. The systems were shown to effectively classify images of cells from the titration assays to negative (non-infected) or positive (virus infected) with an accuracy of over 99%. The methods have been validated according to GMP principles and related guidelines. Regulatory and logistical challenges of implementation of AI-based solutions into GMP-compliant systems are discussed. Although the implementation of AI-systems into GMP-regulated environment brings specific challenges related to introduction of novel technologies, future benefits for early adopters could outweigh the initial efforts.

Combining neural computation and genetic programming for observational causality detection and causal modelling

Combining neural computation and genetic programming for observational causality detection and causal modelling

Oscillatory discharges in the auditory midbrain of the big brown bat contribute to coding of echo delay.

Subsequent to his breakthrough discovery of delay-tuned neurons in the bat's auditory midbrain and cortex, Albert Feng proposed that neural computations for echo delay involve intrinsic oscillatory discharges generated in the inferior colliculus (IC). To explore further the presence of these neural oscillations, we recorded multiple unit activity with a novel annular low impedance electrode from the IC of anesthetized big brown bats and Seba's short-tailed fruit bats. In both species, responses to tones, noise bursts, and FM sweeps contain long latency components, extending up to 60ms post-stimulus onset, organized in periodic, oscillatory-like patterns at frequencies of 360-740Hz. Latencies of this oscillatory activity resemble the wide distributions of single neuron response latencies in the IC. In big brown bats, oscillations lasting up to 30ms after pulse onset emerge in response to single FM pulse-echo pairs, at particular pulse-echo delays. Oscillatory responses to pulses and evoked responses to echoes overlap extensively at short echo delays (5-7ms), creating interference-like patterns. At longer echo delays, responses are separately evident to both pulses and echoes, with less overlap. These results extend Feng's reports of IC oscillations, and point to different processing mechanisms underlying perception of short vs long echo delays.

Lactate supply overtakes glucose when neural computational and cognitive loads scale up

Neural computational power is determined by neuroenergetics, but how and which energy substrates are allocated to various forms of memory engram is unclear. To solve this question, we asked whether neuronal fueling by glucose or lactate scales differently upon increasing neural computation and cognitive loads. Here, using electrophysiology, two-photon imaging, cognitive tasks, and mathematical modeling, we show that both glucose and lactate are involved in engram formation, with lactate supporting long-term synaptic plasticity evoked by high-stimulation load activity patterns and high attentional load in cognitive tasks and glucose being sufficient for less demanding neural computation and learning tasks. Indeed, we show that lactate is mandatory for demanding neural computation, such as theta-burst stimulation, while glucose is sufficient for lighter forms of activity-dependent long-term potentiation (LTP), such as spike timing-dependent plasticity (STDP). We find that subtle variations of spike number or frequency in STDP are sufficient to shift the on-demand fueling from glucose to lactate. Finally, we demonstrate that lactate is necessary for a cognitive task requiring high attentional load, such as the object-in-place task, and for the corresponding invivo hippocampal LTP expression but is not needed for a less demanding task, such as a simple novel object recognition. Overall, these results demonstrate that glucose and lactate metabolism are differentially engaged in neuronal fueling depending on the complexity of the activity-dependent plasticity and behavior.

Convolutional neural network-based portable computer vision system for freshness assessment of crayfish (Prokaryophyllus clarkii).

Developing novel techniques for freshness assessment are of the utmost importance in yield and trade of aquatic products. The crayfish (Prokaryophyllus clarkii) is one of the most popular freshwater products in China, and its food safety should be a serious concern. In this study, a convolutional neural network (CNN)-based portable computer vision system for freshness assessment of crayfish method was proposed. A portable microscope was utilized to collect the microscopic images of crayfish with different freshness levels. The convolutional neural network was constructed and then optimized to extract features from the microscopic images. For the pictures from the portable microscope, the prediction accuracies of freshness could be 86.5% and 83.3% when the optimized networks were applied. The results indicate that the convolutional neural network-based portable computer vision system may provide an alternative way for the freshness assessment in the crayfish industrial chain. PRACTICAL APPLICATION: Portable computer vision system was constructed by a portable microscope connected to a mobile phone. The freshness of crayfish could be rapidly assessed by analyzing the pictures of crayfish using the system. The convolutional neural network-based portable computer vision system may provide an alternative way for the freshness assessment in the crayfish industrial chain.

Single-Transistor Neuron with Excitatory-Inhibitory Spatiotemporal Dynamics Applied for Neuronal Oscillations.

Brain-inspired neuromorphic computing systems with the potential to drive the next wave of artificial intelligence demand a spectrum of critical components beyond simple characteristics. An emerging research trend is to achieve advanced functions with ultracompact neuromorphic devices. In this work, a single-transistor neuron is demonstrated that implements excitatory-inhibitory (E-I) spatiotemporal integration and a series of essential neuron behaviors. Neuronal oscillations, the fundamental mode of neuronal communication, that construct high-dimensional population code to achieve efficient computing in the brain, can also be demonstrated by the neuron transistors. The highly scalable E-I neuron can be the basic building block for implementing core neuronal circuit motifs and large-scale architectural plans to replicate energy-efficient neural computations, forming the foundation of future integrated neuromorphic systems.