A 128-Channel Extreme Learning Machine-Based Neural Decoder for Brain Machine Interfaces

A new approach for neural decoding by inspiring of hyperdimensional computing for implantable intra-cortical BMIs

In the field of Brain Machine Interface (BMI), the process of translating motor intention into a machine command is denoted as decoding. However, despite recent advancements, decoding remains a formidable challenge within BMI. The utilization of current decoding algorithms in the field of BMI often involves computational complexity and requires the use of computers. This is primarily due to the reliance on mathematical models to address the decoding issue and perform subsequent output calculations. Unfortunately, computers are not feasible for implantable BMI systems due to their size and power consumption. To address this predicament, this study proposes a pioneering approach inspired by hyperdimensional computing. This approach first involves identifying the pattern of each stimulus by considering the normal firing rate distribution of each neuron. Subsequently, the newly observed firing pattern for each input is compared with the patterns detected at each moment for each neuron. The algorithm, which shares similarities with hyperdimensional computing, identifies the most similar pattern as the final output. This approach reduces the dependence on mathematical models. The efficacy of this method is assessed through the utilization of an authentic dataset acquired from the Frontal Eye Field (FEF) of two male rhesus monkeys. The output space encompasses eight possible angles. The results demonstrate an accuracy rate of 51.5% while exhibiting significantly low computational complexity, involving a mere 2050 adder operators. Furthermore, the proposed algorithm is implemented on a field-programmable gate array (FPGA) and as an ASIC designe in a standard CMOS 180 nm technology, underscoring its suitability for real-time implantable BMI applications. The implementation required only 2.3 Kbytes of RAM, occupied an area of 2.2 mm2, and consumed 9.32 µW at a 1.8 V power supply. Consequently, the proposed solution represents an accurate, low computational complexity, hardware-friendly, and real-time approach.

An approximate randomization-based neural network with dedicated digital architecture for energy-constrained devices

Variable energy constraints affect the implementations of neural networks on battery-operated embedded systems. This paper describes a learning algorithm for randomization-based neural networks with hard-limit activation functions. The approach adopts a novel cost function that balances accuracy and network complexity during training. From an energy-specific perspective, the new learning strategy allows to adjust, dynamically and in real time, the number of operations during the network’s forward phase. The proposed learning scheme leads to efficient predictors supported by digital architectures. The resulting digital architecture can switch to approximate computing at run time, in compliance with the available energy budget. Experiments on 10 real-world prediction testbeds confirmed the effectiveness of the learning scheme. Additional tests on limited-resource devices supported the implementation efficiency of the overall design approach.

In-sensor neural network for high energy efficiency analog-to-information conversion

This work presents an on-chip analog-to-information conversion technique that utilizes analog hyper-dimensional computing based on reservoir-computing paradigm to process electrocardiograph (ECG) signals locally in-sensor and reduce radio frequency transmission by more than three orders-of-magnitude. Instead of transmitting the naturally sparse ECG signal or extracted features, the on-chip analog-to-information converter analyzes the ECG signal through a nonlinear reservoir kernel followed by an artificial neural network, and transmits the prediction results. The proposed technique is demonstrated for detection of sepsis onset and achieves state-of-the-art accuracy and energy efficiency while reducing sensor power by [Formula: see text] with test-chips prototyped in 65 nm CMOS.

A Power-Efficient Brain-Machine Interface System With a Sub-mw Feature Extraction and Decoding ASIC Demonstrated in Nonhuman Primates

Intracortical brain-machine interfaces have shown promise for restoring function to people with paralysis, but their translation to portable and implantable devices is hindered by their high power consumption. Recent devices have drastically reduced power consumption compared to standard experimental brain-machine interfaces, but still require wired or wireless connections to computing hardware for feature extraction and inference. Here, we introduce a Neural Recording And Decoding (NeuRAD) application specific integrated circuit (ASIC) in 180 nm CMOS that can extract neural spiking features and predict two-dimensional behaviors in real-time. To reduce amplifier and feature extraction power consumption, the NeuRAD has a hardware accelerator for extracting spiking band power (SBP) from intracortical spiking signals and includes an M0 processor with a fixed-point Matrix Acceleration Unit (MAU) for efficient and flexible decoding. We validated device functionality by recording SBP from a nonhuman primate implanted with a Utah microelectrode array and predicting the one- and two-dimensional finger movements the monkey was attempting to execute in closed-loop using a steady-state Kalman filter (SSKF). Using the NeuRAD's real-time predictions, the monkey achieved 100% success rate and 0.82 s mean target acquisition time to control one-dimensional finger movements using just 581 μW. To predict two-dimensional finger movements, the NeuRAD consumed 588 μW to enable the monkey to achieve a 96% success rate and 2.4 s mean acquisition time. By employing SBP, ASIC brain-machine interfaces can close the gap to enable fully implantable therapies for people with paralysis.

Closed-Loop Neural Prostheses With On-Chip Intelligence: A Review and a Low-Latency Machine Learning Model for Brain State Detection

The application of closed-loop approaches in systems neuroscience and therapeutic stimulation holds great promise for revolutionizing our understanding of the brain and for developing novel neuromodulation therapies to restore lost functions. Neural prostheses capable of multi-channel neural recording, on-site signal processing, rapid symptom detection, and closed-loop stimulation are critical to enabling such novel treatments. However, the existing closed-loop neuromodulation devices are too simplistic and lack sufficient on-chip processing and intelligence. In this paper, we first discuss both commercial and investigational closed-loop neuromodulation devices for brain disorders. Next, we review state-of-the-art neural prostheses with on-chip machine learning, focusing on application-specific integrated circuits (ASIC). System requirements, performance and hardware comparisons, design trade-offs, and hardware optimization techniques are discussed. To facilitate a fair comparison and guide design choices among various on-chip classifiers, we propose a new energy-area (E-A) efficiency figure of merit that evaluates hardware efficiency and multi-channel scalability. Finally, we present several techniques to improve the key design metrics of tree-based on-chip classifiers, both in the context of ensemble methods and oblique structures. A novel Depth-Variant Tree Ensemble (DVTE) is proposed to reduce processing latency (e.g., by 2.5× on seizure detection task). We further develop a cost-aware learning approach to jointly optimize the power and latency metrics. We show that algorithm-hardware co-design enables the energy- and memory-optimized design of tree-based models, while preserving a high accuracy and low latency. Furthermore, we show that our proposed tree-based models feature a highly interpretable decision process that is essential for safety-critical applications such as closed-loop stimulation.

Biomedical Implants with Charge-Transfer Monitoring and Regulating Abilities

Transmembrane charge (ion/electron) transfer is essential for maintaining cellular homeostasis and is involved in many biological processes, from protein synthesis to embryonic development in organisms. Designing implant devices that can detect or regulate cellular transmembrane charge transfer is expected to sense and modulate the behaviors of host cells and tissues. Thus, charge transfer can be regarded as a bridge connecting living systems and human-made implantable devices. This review describes the mode and mechanism of charge transfer between organisms and nonliving materials, and summarizes the strategies to endow implants with charge-transfer regulating or monitoring abilities. Furthermore, three major charge-transfer controlling systems, including wired, self-activated, and stimuli-responsive biomedical implants, as well as the design principles and pivotal materials are systematically elaborated. The clinical challenges and the prospects for future development of these implant devices are also discussed.

Shared Prosthetic Control Based on Multiple Movement Intent Decoders

SIGNIFICANCE

A number of movement intent decoders exist in the literature that typically differ in the algorithms used and the nature of the outputs generated. Each approach comes with its own advantages and disadvantages. Combining the estimates of multiple algorithms may have better performance than any of the individual methods.

OBJECTIVE

This paper presents and evaluates a shared controller framework for prosthetic limbs based on multiple decoders of volitional movement intent.

METHODS

An algorithm to combine multiple estimates to control the prosthesis is developed in this paper. The capabilities of the approach are validated using a system that combines a Kalman filter-based decoder with a multilayer perceptron classifier-based decoder. The shared controller's performance is validated in online experiments where a virtual limb is controlled in real-time by amputee and intact-arm subjects. During the testing phase subjects controlled a virtual hand in real time to move digits to instructed positions using either a Kalman filter decoder, a multilayer perceptron decoder, or a linear combination of the two.

RESULTS

The shared controller results in statistically significant improvements over the component decoders. Specifically, certain degrees of shared control result in increases in the time-in-target metric and decreases in unintended movements.

CONCLUSION

The shared controller of this paper combines the good qualities of component decoders tested in this paper. Herein, combining a Kalman filter decoder with a classifier-based decoder inherits the flexibility of the Kalman filter decoder and the limited unwanted movements from the classifier-based decoder, resulting in a system that may be able to perform the tasks of everyday life more naturally and reliably.

Hardware Implementation of Deep Network Accelerators Towards Healthcare and Biomedical Applications

The advent of dedicated Deep Learning (DL) accelerators and neuromorphic processors has brought on new opportunities for applying both Deep and Spiking Neural Network (SNN) algorithms to healthcare and biomedical applications at the edge. This can facilitate the advancement of medical Internet of Things (IoT) systems and Point of Care (PoC) devices. In this paper, we provide a tutorial describing how various technologies including emerging memristive devices, Field Programmable Gate Arrays (FPGAs), and Complementary Metal Oxide Semiconductor (CMOS) can be used to develop efficient DL accelerators to solve a wide variety of diagnostic, pattern recognition, and signal processing problems in healthcare. Furthermore, we explore how spiking neuromorphic processors can complement their DL counterparts for processing biomedical signals. The tutorial is augmented with case studies of the vast literature on neural network and neuromorphic hardware as applied to the healthcare domain. We benchmark various hardware platforms by performing a sensor fusion signal processing task combining electromyography (EMG) signals with computer vision. Comparisons are made between dedicated neuromorphic processors and embedded AI accelerators in terms of inference latency and energy. Finally, we provide our analysis of the field and share a perspective on the advantages, disadvantages, challenges, and opportunities that various accelerators and neuromorphic processors introduce to healthcare and biomedical domains.

Multichannel parallel processing of neural signals in memristor arrays

Fully implantable neural interfaces with massive recording channels bring the gospel to patients with motor or speech function loss. As the number of recording channels rapidly increases, conventional complementary metal-oxide semiconductor (CMOS) chips for neural signal processing face severe challenges on parallelism scalability, computational cost, and power consumption. In this work, we propose a previously unexplored approach for parallel processing of multichannel neural signals in memristor arrays, taking advantage of their rich dynamic characteristics. The critical information of neural signal waveform is extracted and encoded in the memristor conductance modulation. A signal segmentation scheme is developed to adapt to device variations. To verify the fidelity of the processed results, seizure prediction is further demonstrated, with high accuracy above 95% and also more than 1000× improvement in power efficiency compared with CMOS counterparts. This work suggests that memristor arrays could be a promising multichannel signal processing module for future implantable neural interfaces.

Neural signal analysis with memristor arrays towards high-efficiency brain-machine interfaces

Brain-machine interfaces are promising tools to restore lost motor functions and probe brain functional mechanisms. As the number of recording electrodes has been exponentially rising, the signal processing capability of brain–machine interfaces is falling behind. One of the key bottlenecks is that they adopt conventional von Neumann architecture with digital computation that is fundamentally different from the working principle of human brain. In this work, we present a memristor-based neural signal analysis system, where the bio-plausible characteristics of memristors are utilized to analyze signals in the analog domain with high efficiency. As a proof-of-concept demonstration, memristor arrays are used to implement the filtering and identification of epilepsy-related neural signals, achieving a high accuracy of 93.46%. Remarkably, our memristor-based system shows nearly 400× improvements in the power efficiency compared to state-of-the-art complementary metal-oxide-semiconductor systems. This work demonstrates the feasibility of using memristors for high-performance neural signal analysis in next-generation brain–machine interfaces.

Designing energy efficient and high performance brain-machine interfaces with millions of recording electrodes for in-situ analysis remains a challenge. Here, the authors develop a memristor-based neural signal analysis system capable of filtering and identifying epilepsy-related brain activities with an accuracy of 93.46%.

ResOT: Resource-Efficient Oblique Trees for Neural Signal Classification

Classifiers that can be implemented on chip with minimal computational and memory resources are essential for edge computing in emerging applications such as medical and IoT devices. This paper introduces a machine learning model based on oblique decision trees to enable resource-efficient classification on a neural implant. By integrating model compression with probabilistic routing and implementing cost-aware learning, our proposed model could significantly reduce the memory and hardware cost compared to state-of-the-art models, while maintaining the classification accuracy. We trained the resource-efficient oblique tree with power-efficient regularization (ResOT-PE) on three neural classification tasks to evaluate the performance, memory, and hardware requirements. On seizure detection task, we were able to reduce the model size by 3.4× and the feature extraction cost by 14.6× compared to the ensemble of boosted trees, using the intracranial EEG from 10 epilepsy patients. In a second experiment, we tested the ResOT-PE model on tremor detection for Parkinson's disease, using the local field potentials from 12 patients implanted with a deep-brain stimulation (DBS) device. We achieved a comparable classification performance as the state-of-the-art boosted tree ensemble, while reducing the model size and feature extraction cost by 10.6× and 6.8×, respectively. We also tested on a 6-class finger movement detection task using ECoG recordings from 9 subjects, reducing the model size by 17.6× and feature computation cost by 5.1×. The proposed model can enable a low-power and memory-efficient implementation of classifiers for real-time neurological disease detection and motor decoding.

Quantifying Leaf Chlorophyll Concentration of Sorghum from Hyperspectral Data Using Derivative Calculus and Machine Learning

Leaf chlorophyll concentration (LCC) is an important indicator of plant health, vigor, physiological status, productivity, and nutrient deficiencies. Hyperspectral spectroscopy at leaf level has been widely used to estimate LCC accurately and non-destructively. This study utilized leaf-level hyperspectral data with derivative calculus and machine learning to estimate LCC of sorghum. We calculated fractional derivative (FD) orders starting from 0.2 to 2.0 with 0.2 order increments. Additionally, 43 common vegetation indices (VIs) were calculated from leaf spectral reflectance factor to make comparisons with reflectance-based data. Within the modeling pipeline, three feature selection methods were assessed: Pearson’s correlation coefficient (PCC), partial least squares based variable importance in the projection (VIP), and random forest-based mean decrease impurity (MDI). Finally, we used partial least squares regression (PLSR), random forest regression (RFR), support vector regression (SVR), and extreme learning regression (ELR) to estimate the LCC of sorghum. Results showed that: (1) increasing derivative order can show improved model performance until certain order for reflectance-based analysis; however, it is inconclusive to state that a particular order is optimal for estimating LCC of sorghum; (2) VI-based modeling outperformed derivative augmented reflectance factor-based modeling; (3) mean decrease impurity was found effective in selecting sensitive features from large feature space (reflectance-based analysis), whereas simple Pearson’s correlation coefficient worked better with smaller feature space (VI-based analysis); and (4) SVR outperformed all other models within reflectance-based analysis; alternatively, ELR with VIs from original reflectance yielded slightly better results compared to all other models.

An Efficient Combination among sMRI, CSF, Cognitive Score, and APOE ε4 Biomarkers for Classification of AD and MCI Using Extreme Learning Machine

Alzheimer's disease (AD) is the most common cause of dementia and a progressive neurodegenerative condition, characterized by a decline in cognitive function. Symptoms usually appear gradually and worsen over time, becoming severe enough to interfere with individual daily tasks. Thus, the accurate diagnosis of both AD and the prodromal stage (i.e., mild cognitive impairment (MCI)) is crucial for timely treatment. As AD is inherently dynamic, the relationship between AD indicators is unclear and varies over time. To address this issue, we first aimed at investigating differences in atrophic patterns between individuals with AD and MCI and healthy controls (HCs). Then we utilized multiple biomarkers, along with filter- and wrapper-based feature selection and an extreme learning machine- (ELM-) based approach, with 10-fold cross-validation for classification. Increasing efforts are focusing on the use of multiple biomarkers, which can be useful for the diagnosis of AD and MCI. However, optimum combinations have yet to be identified and most multimodal analyses use only volumetric measures obtained from magnetic resonance imaging (MRI). Anatomical structural MRI (sMRI) measures have also so far mostly been used separately. The full possibilities of using anatomical MRI for AD detection have thus yet to be explored. In this study, three measures (cortical thickness, surface area, and gray matter volume), obtained from sMRI through preprocessing for brain atrophy measurements; cerebrospinal fluid (CSF), for quantification of specific proteins; cognitive score, as a measure of cognitive performance; and APOE ε4 allele status were utilized. Our results show that a combination of specific biomarkers performs well, with accuracies of 97.31% for classifying AD vs. HC, 91.72% for MCI vs. HC, 87.91% for MCI vs. AD, and 83.38% for MCIs vs. MCIc, respectively, when evaluated using the proposed algorithm. Meanwhile, the areas under the curve (AUC) from the receiver operating characteristic (ROC) curves combining multiple biomarkers provided better classification performance. The proposed features combination and selection algorithm effectively classified AD and MCI, and MCIs vs. MCIc, the most challenging classification task, and therefore could increase the accuracy of AD classification in clinical practice. Furthermore, we compared the performance of the proposed method with SVM classifiers, using a cross-validation method with Alzheimer's Disease Neuroimaging Initiative (ADNI) datasets.

Inhibition of Long-Term Variability in Decoding Forelimb Trajectory Using Evolutionary Neural Networks With Error-Correction Learning

Objective: In brain machine interfaces (BMIs), the functional mapping between neural activities and kinematic parameters varied over time owing to changes in neural recording conditions. The variability in neural recording conditions might result in unstable long-term decoding performance. Relevant studies trained decoders with several days of training data to make them inherently robust to changes in neural recording conditions. However, these decoders might not be robust to changes in neural recording conditions when only a few days of training data are available. In time-series prediction and feedback control system, an error feedback was commonly adopted to reduce the effects of model uncertainty. This motivated us to introduce an error feedback to a neural decoder for dealing with the variability in neural recording conditions. Approach: We proposed an evolutionary constructive and pruning neural network with error feedback (ECPNN-EF) as a neural decoder. The ECPNN-EF with partially connected topology decoded the instantaneous firing rates of each sorted unit into forelimb movement of a rat. Furthermore, an error feedback was adopted as an additional input to provide kinematic information and thus compensate for changes in functional mapping. The proposed neural decoder was trained on data collected from a water reward-related lever-pressing task for a rat. The first 2 days of data were used to train the decoder, and the subsequent 10 days of data were used to test the decoder. Main Results: The ECPNN-EF under different settings was evaluated to better understand the impact of the error feedback and partially connected topology. The experimental results demonstrated that the ECPNN-EF achieved significantly higher daily decoding performance with smaller daily variability when using the error feedback and partially connected topology. Significance: These results suggested that the ECPNN-EF with partially connected topology could cope with both within- and across-day changes in neural recording conditions. The error feedback in the ECPNN-EF compensated for decreases in decoding performance when neural recording conditions changed. This mechanism made the ECPNN-EF robust against changes in functional mappings and thus improved the long-term decoding stability when only a few days of training data were available.

Sparse Ensemble Machine Learning to Improve Robustness of Long-Term Decoding in iBMIs

This paper presents a novel sparse ensemble based machine learning approach to enhance robustness of intracortical Brain Machine Interfaces (iBMIs) in the face of non-stationary distribution of input neural data across time. Each classifier in the ensemble is trained on a randomly sampled (with replacement) set of input channels. These sparse connections ensure that with a high chance, few of the base classifiers should be less affected by the variations in some of the recording channels. We have tested the generality of this technique on different base classifiers - linear discriminant analysis (LDA), support vector machine (SVM), extreme learning machine (ELM) and multilayer perceptron (MLP). Results show decoding accuracy improvements of up to ≈21 %, 13%, 19%, 10% in non-human primate (NHP) A and 7%, 9%, 7%, 9% in NHP B across test days while using the sparse ensemble approach over a single classifier model for LDA, SVM, ELM and MLP algorithms respectively. Furthermore, improvements of up to ≈7(14)%, 8(15)%, 9(19)%, 7(15)% in NHP A and 8(15)%, 12(20)%, 15(23)%, 12(19)% in NHP B over Random Forest (Long-short Term Memory) have been obtained by sparse ensemble LDA, SVM, ELM, MLP respectively.

Towards Intelligent Intracortical BMI (i 2BMI): Low-Power Neuromorphic Decoders That Outperform Kalman Filters

Fully-implantable wireless intracortical Brain Machine Interfaces (iBMI) is one of the most promising next frontiers in the nascent field of neurotechnology. However, scaling the number of channels in such systems by another 10× is difficult due to power and bandwidth requirements of the wireless transmitter. One promising solution for that is to include more processing, up to the decoder, in the implant so that transmission data-rate is reduced drastically. Earlier work on neuromorphic decoder chips only showed classification of discrete states. We present results for continuous state decoding using a low-power neuromorphic decoder chip termed Spike-input Extreme Learning Machine (SELMA) that implements a nonlinear decoder without memory and its memory-based version with time-delayed bins, SELMA-bins. We have compared SELMA, SELMA-bins against state-of-the-art Steady-State Kalman Filter (SSKF), a linear decoder with memory, across two different datasets involving a total of 4 non-human primates (NHPs). Results show at least a 10% (20%) or more increase in the fraction of variance accounted for (FVAF) by SELMA (SELMA-bins) over SSKF across the datasets. Estimated energy consumption comparison shows SELMA (SELMA-bins) consuming ≈ 9 nJ/update (23 nJ/update) against SSKF's ≈ 7.4 nJ/update for an iBMI with a 10 degree of freedom control. Thus, SELMA yields better performance against SSKF while consuming energy in the same range as SSKF whereas SELMA-bins performs the best with moderately increased energy consumption, albeit far less than energy required for raw data transmission. This paves the way for reducing transmission data rates in future scaled iBMI systems.

Dual Activation Function-Based Extreme Learning Machine (ELM) for Estimating Grapevine Berry Yield and Quality

Reliable assessment of grapevine productivity is a destructive and time-consuming process. In addition, the mixed effects of grapevine water status and scion-rootstock interactions on grapevine productivity are not always linear. Despite the potential opportunity of applying remote sensing and machine learning techniques to predict plant traits, there are still limitations to previously studied techniques for vine productivity due to the complexity of the system not being adequately modeled. During the 2014 and 2015 growing seasons, hyperspectral reflectance spectra were collected using a handheld spectroradiometer in a vineyard designed to investigate the effects of irrigation level (0%, 50%, and 100%) and rootstocks (1103 Paulsen, 3309 Couderc, SO4 and Chambourcin) on vine productivity. To assess vine productivity, it is necessary to measure factors related to fruit ripeness and not just yield, as an over cropped vine may produce high-yield but poor-quality fruit. Therefore, yield, Total Soluble Solids (TSS), Titratable Acidity (TA) and the ratio TSS/TA (maturation index, IMAD) were measured. A total of 20 vegetation indices were calculated from hyperspectral data and used as input for predictive model calibration. Prediction performance of linear/nonlinear multiple regression methods and Weighted Regularized Extreme Learning Machine (WRELM) were compared with our newly developed WRELM-TanhRe. The developed method is based on two activation functions: hyperbolic tangent (Tanh) and rectified linear unit (ReLU). The results revealed that WRELM and WRELM-TanhRe outperformed the widely used multiple regression methods when model performance was tested with an independent validation dataset. WRELM-TanhRe produced the highest prediction accuracy for all the berry yield and quality parameters (R2 of 0.522–0.682 and RMSE of 2–15%), except for TA, which was predicted best with WRELM (R2 of 0.545 and RMSE of 6%). The results demonstrate the value of combining hyperspectral remote sensing and machine learning methods for improving of berry yield and quality prediction.

A Sub- μW/Ch Analog Front-End for ∆-Neural Recording With Spike-Driven Data Compression

We present a fully implantable neural recording IC with a spike-driven data compression scheme to improve the power efficiency and preserve crucial data for monitoring brain activities. A difference between two consecutive neural signals, ∆-neural signal, is sampled in each channel to reduce the full dynamic range and the required resolution of an analog-to-digital converter (ADC), enabling the whole analog chain to be operated at a 0.5-V supply. A set of multiple ∆-signals are stored in analog memory to extract the magnitude and frequency features of the incoming neural signals, which are utilized to discriminate spikes in these signals instantaneously after the acquisition in the analog domain. The energy- and area-efficient successive approximation ADC is implemented and only converts detected spikes, decreasing the power dissipation and the amount of neural data. A prototype 16-channel neural interface IC was fabricated using a 0.18-μm CMOS process, and each component in the analog front-end was fully characterized. We successfully demonstrated precise spike detection through both in vitro and in vivo acquisition of the neural signal. The prototype chip consumed 0.88 μW/channel at a 0.5-V supply for the recording and compressed about 89% of neural data, saving the power consumption and bandwidth in the system.

Identification of Alzheimer's disease and mild cognitive impairment using multimodal sparse hierarchical extreme learning machine

Different modalities such as structural MRI, FDG-PET, and CSF have complementary information, which is likely to be very useful for diagnosis of AD and MCI. Therefore, it is possible to develop a more effective and accurate AD/MCI automatic diagnosis method by integrating complementary information of different modalities. In this paper, we propose multi-modal sparse hierarchical extreme leaning machine (MSH-ELM). We used volume and mean intensity extracted from 93 regions of interest (ROIs) as features of MRI and FDG-PET, respectively, and used p-tau, t-tau, and A β 42 as CSF features. In detail, high-level representation was individually extracted from each of MRI, FDG-PET, and CSF using a stacked sparse extreme learning machine auto-encoder (sELM-AE). Then, another stacked sELM-AE was devised to acquire a joint hierarchical feature representation by fusing the high-level representations obtained from each modality. Finally, we classified joint hierarchical feature representation using a kernel-based extreme learning machine (KELM). The results of MSH-ELM were compared with those of conventional ELM, single kernel support vector machine (SK-SVM), multiple kernel support vector machine (MK-SVM) and stacked auto-encoder (SAE). Performance was evaluated through 10-fold cross-validation. In the classification of AD vs. HC and MCI vs. HC problem, the proposed MSH-ELM method showed mean balanced accuracies of 96.10% and 86.46%, respectively, which is much better than those of competing methods. In summary, the proposed algorithm exhibits consistently better performance than SK-SVM, ELM, MK-SVM and SAE in the two binary classification problems (AD vs. HC and MCI vs. HC).

An Extreme Learning Machine-Based Neuromorphic Tactile Sensing System for Texture Recognition

Despite significant advances in computational algorithms and development of tactile sensors, artificial tactile sensing is strikingly less efficient and capable than the human tactile perception. Inspired by efficiency of biological systems, we aim to develop a neuromorphic system for tactile pattern recognition. We particularly target texture recognition as it is one of the most necessary and challenging tasks for artificial sensory systems. Our system consists of a piezoresistive fabric material as the sensor to emulate skin, an interface that produces spike patterns to mimic neural signals from mechanoreceptors, and an extreme learning machine (ELM) chip to analyze spiking activity. Benefiting from intrinsic advantages of biologically inspired event-driven systems and massively parallel and energy-efficient processing capabilities of the ELM chip, the proposed architecture offers a fast and energy-efficient alternative for processing tactile information. Moreover, it provides the opportunity for the development of low-cost tactile modules for large-area applications by integration of sensors and processing circuits. We demonstrate the recognition capability of our system in a texture discrimination task, where it achieves a classification accuracy of 92% for categorization of ten graded textures. Our results confirm that there exists a tradeoff between response time and classification accuracy (and information transfer rate). A faster decision can be achieved at early time steps or by using a shorter time window. This, however, results in deterioration of the classification accuracy and information transfer rate. We further observe that there exists a tradeoff between the classification accuracy and the input spike rate (and thus energy consumption). Our work substantiates the importance of development of efficient sparse codes for encoding sensory data to improve the energy efficiency. These results have a significance for a wide range of wearable, robotic, prosthetic, and industrial applications.

A Comparison of Low-Complexity Real-Time Feature Extraction for Neuromorphic Speech Recognition

This paper presents a real-time, low-complexity neuromorphic speech recognition system using a spiking silicon cochlea, a feature extraction module and a population encoding method based Neural Engineering Framework (NEF)/Extreme Learning Machine (ELM) classifier IC. Several feature extraction methods with varying memory and computational complexity are presented along with their corresponding classification accuracies. On the N-TIDIGITS18 dataset, we show that a fixed bin size based feature extraction method that votes across both time and spike count features can achieve an accuracy of 95% in software similar to previously report methods that use fixed number of bins per sample while using ~3× less energy and ~25× less memory for feature extraction (~1.5× less overall). Hardware measurements for the same topology show a slightly reduced accuracy of 94% that can be attributed to the extra correlations in hardware random weights. The hardware accuracy can be increased by further increasing the number of hidden nodes in ELM at the cost of memory and energy.

A Noise Filtering Algorithm for Event-Based Asynchronous Change Detection Image Sensors on TrueNorth and Its Implementation on TrueNorth

Asynchronous event-based sensors, or “silicon retinae,” are a new class of vision sensors inspired by biological vision systems. The output of these sensors often contains a significant number of noise events along with the signal. Filtering these noise events is a common preprocessing step before using the data for tasks such as tracking and classification. This paper presents a novel spiking neural network-based approach to filtering noise events from data captured by an Asynchronous Time-based Image Sensor on a neuromorphic processor, the IBM TrueNorth Neurosynaptic System. The significant contribution of this work is that it demonstrates our proposed filtering algorithm outperforms the traditional nearest neighbor noise filter in achieving higher signal to noise ratio (~10 dB higher) and retaining the events related to signal (~3X more). In addition, for our envisioned application of object tracking and classification under some parameter settings, it can also generate some of the missing events in the spatial neighborhood of the signal for all classes of moving objects in the data which are unattainable using the nearest neighbor filter.

A Streaming PCA VLSI Chip for Neural Data Compression

Neural recording system miniaturization and integration with low-power wireless technologies require compressing neural data before transmission. Feature extraction is a procedure to represent data in a low-dimensional space; its integration into a recording chip can be an efficient approach to compress neural data. In this paper, we propose a streaming principal component analysis algorithm and its microchip implementation to compress multichannel local field potential (LFP) and spike data. The circuits have been designed in a 65-nm CMOS technology and occupy a silicon area of 0.06 mm. Throughout the experiments, the chip compresses LFPs by 10 at the expense of as low as 1% reconstruction errors and 144-nW/channel power consumption; for spikes, the achieved compression ratio is 25 with 8% reconstruction errors and 3.05-W/channel power consumption. In addition, the algorithm and its hardware architecture can swiftly adapt to nonstationary spiking activities, which enables efficient hardware sharing among multiple channels to support a high-channel count recorder.

Neuromorphic Hardware Architecture Using the Neural Engineering Framework for Pattern Recognition

We present a hardware architecture that uses the neural engineering framework (NEF) to implement large-scale neural networks on field programmable gate arrays (FPGAs) for performing massively parallel real-time pattern recognition. NEF is a framework that is capable of synthesising large-scale cognitive systems from subnetworks and we have previously presented an FPGA implementation of the NEF that successfully performs nonlinear mathematical computations. That work was developed based on a compact digital neural core, which consists of 64 neurons that are instantiated by a single physical neuron using a time-multiplexing approach. We have now scaled this approach up to build a pattern recognition system by combining identical neural cores together. As a proof of concept, we have developed a handwritten digit recognition system using the MNIST database and achieved a recognition rate of 96.55%. The system is implemented on a state-of-the-art FPGA and can process 5.12 million digits per second. The architecture and hardware optimisations presented offer high-speed and resource-efficient means for performing high-speed, neuromorphic, and massively parallel pattern recognition and classification tasks.