Neural Engine Research Articles

Computers and chess masters: The role of AI intransforming elite human performance.

Advances in Artificial Intelligence (AI) have made significant strides in recent years, often supplementing rather than replacing human performance. The extent of their assistance at the highest levels of human performance remains unclear. We analyse over 11.6 million decisions of elite chess players, a domain commonly used as a testbed for AI and psychology due to its complexity and objective assessment. We investigated the impact of two AI chess revolutions: the first in the late 1990s with the rise of powerful PCs and internet access and the second in the late 2010s with deep learning-powered chess engines. The rate of human improvement mirrored AI advancements, but contrary to expectations, the quality of decisions mostly improved steadily over four decades, irrespective of age, with no distinct periods of rapid improvement. Only the youngest top players saw marked gains in the late 1990s, likely due to better access to knowledge and computers. Surprisingly, the recent wave of neural network-powered engines has not significantly impacted the best players - at least, not yet. Our research highlights AI's potential to enhance human capability in complex tasks, given the right conditions, even among the most elite performers.

Turing tests in chess: An experiment revealing the role of human subjectivity

With the growing capabilities of AI, technology is increasingly able to match or even surpass human performance. In the current study, focused on the game of chess, we investigated whether chess players could distinguish whether they were playing against a human or a computer, and how they achieved this. A total of 24 chess players each played eight 5 + 0 Blitz games from different starting positions. They played against (1) a human, (2) Maia, a neural network-based chess engine trained to play in a human-like manner, (3) Stockfish 16, the best chess engine available, downgraded to play at a lower level, and (4) Stockfish 16 at its maximal level. The opponent’s move time was fixed at 10 s. During the game, participants verbalized their thoughts, and after each game, they indicated by means of a questionnaire whether they thought they had played against a human or a machine and if there were particular moves that revealed the nature of the opponent. The results showed that Stockfish at the highest level was usually correctly identified as an engine, while Maia was often incorrectly identified as a human. The moves of the downgraded Stockfish were relatively often labeled as ‘strange’ by the participants. In conclusion, the Turing test, as applied here in a domain where computers can perform superhumanly, is essentially a test of whether the chess computer can devise suboptimal moves that correspond to human moves, and not necessarily a test of computer intelligence.

Siracusa: A 16 nm Heterogenous RISC-V SoC for Extended Reality With At-MRAM Neural Engine

Siracusa: A 16 nm Heterogenous RISC-V SoC for Extended Reality With At-MRAM Neural Engine

NEural Engine for Discovering Luminous Events (NEEDLE): identifying rare transient candidates in real time from host galaxy images

ABSTRACT Known for their efficiency in analysing large data sets, machine learning-based classifiers have been widely used in wide-field sky survey pipelines. The upcoming Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will generate millions of real-time alerts every night, enabling the discovery of large samples of rare events. Identifying such objects soon after explosion will be essential to study their evolution. Using ∼5400 transients from the Zwicky Transient Facility (ZTF) Bright Transient Survey as training and test data, we develop NEEDLE (NEural Engine for Discovering Luminous Events), a novel hybrid (convolutional neural network + dense neural network) classifier to select for two rare classes with strong environmental preferences: superluminous supernovae (SLSNe) preferring dwarf galaxies, and tidal disruption events (TDEs) occurring in the centres of nucleated galaxies. The input data includes (i) cutouts of the detection and reference images, (ii) photometric information contained directly in the alert packets, and (iii) host galaxy magnitudes from Pan-STARRS (Panoramic Survey Telescope and Rapid Response System). Despite having only a few tens of examples of the rare classes, our average (best) completeness on an unseen test set reaches 73 per cent (86 per cent) for SLSNe and 80 per cent (87 per cent) for TDEs. While very encouraging for completeness, this may still result in relatively low purity for the rare transients, given the large class imbalance in real surveys. However, the goal of NEEDLE is to find good candidates for spectroscopic classification, rather than to select pure photometric samples. Our system will be deployed as an annotator on the UK alert broker, Lasair, to provide predictions of real-time alerts from ZTF and LSST to the community.

An end-to-end deep convolutional neural network-based dose engine for parotid gland cancer seed implant brachytherapy.

Seed implant brachytherapy (SIBT) is a promising treatment modality for parotid gland cancers (PGCs). However, the current clinical standard dose calculation method based on the American Association of Physicists in Medicine (AAPM) Task Group 43 (TG-43) Report oversimplifies patient anatomy as a homogeneous water phantom medium, leading to significant dose calculation errors due to heterogeneity surrounding the parotid gland. Monte Carlo Simulation (MCS) can yield accurate dose distributions but the long computation time hinders its wide application in clinical practice. This paper aims to develop an end-to-end deep convolutional neural network-based dose engine (DCNN-DE) to achieve fast and accurate dose calculation for PGC SIBT. A DCNN model was trained using the patient's CT images and TG-43-based dose maps as inputs, with the corresponding MCS-based dose maps as the ground truth. The DCNN model was enhanced based on our previously proposed model by incorporating attention gates (AGs) and large kernel convolutions. Training and evaluation of the model were performed using a dataset comprising 188 PGC I-125 SIBT patient cases, and its transferability was tested on an additional 16 non-PGC head and neck cancers (HNCs) I-125 SIBT patient cases. Comparison studies were conducted to validate the superiority of the enhanced model over the original one and compare their overall performance. On the PGC testing dataset, the DCNN-DE demonstrated the ability to generate accurate dose maps, with percentage absolute errors (PAEs) of 0.67%±0.47% for clinical target volume (CTV) D90 and 1.04%±1.33% for skin D0.1cc. The comparison studies revealed that incorporating AGs and large kernel convolutions resulted in 8.2% (p<0.001) and 3.1% (p<0.001) accuracy improvement, respectively, as measured by dose mean absolute error. On the non-PGC HNC dataset, the DCNN-DE exhibited good transferability, achieving a CTV D90 PAE of 1.88%±1.73%. The DCNN-DE can generate a dose map in less than 10 ms. We have developed and validated an end-to-end DCNN-DE for PGC SIBT. The proposed DCNN-DE enables fast and accurate dose calculation, making it suitable for application in the plan optimization and evaluation process of PGC SIBT.

Diagnostic Tool for Early Detection of Rheumatic Disorders Using Machine Learning Algorithm and Predictive Models

Background: Rheumatic diseases are chronic diseases that affect joints, tendons, ligaments, bones, muscles, and other vital organs. Detection of rheumatic diseases is a complex process that requires careful analysis of heterogeneous content from clinical examinations, patient history, and laboratory investigations. Machine learning techniques have made it possible to integrate such techniques into the complex diagnostic process to identify inherent features that lead to disease formation, development, and progression for remedial measures. Methods: An automated diagnostic tool using a multilayer neural network computational engine is presented to detect rheumatic disorders and the type of underlying disorder for therapeutic strategies. Rheumatic disorders considered are rheumatoid arthritis, osteoarthritis, and systemic lupus erythematosus. The detection system was trained and tested using 70% and 30% respectively of labelled synthetic dataset of 100,000 records containing both single and multiple disorders. Results: The detection system was able to detect and predict underlying disorders with accuracy of 97.48%, sensitivity of 96.80%, and specificity of 97.50%. Conclusion: The good performance suggests that this solution is robust enough and can be implemented for screening patients for intervention measures. This is a much-needed solution in environments with limited specialists, as the solution promotes task-shifting from the specialist level to the primary healthcare physicians.

Designing a Deep Neural Network engine for LLC block reuse prediction to mitigate Soft Error in Multicore

Last level cache (LLC), a major contender of chip area, exhibits the highest sensitivity to soft error. Block reuse prediction is used to exhibit selective protection of LLC blocks. However, traditional reuse predictors fail to accurately understand the cache access pattern. This work explores machine learning for block reuse prediction that realizes the features that affect reuse likelihood and therefore, better predicts the reusable blocks. To off-load the reuse calculation from the processing cores, a dedicated reuse predictor engine is designed. This engine implements a deep neural network with unsupervised learning and calculates reuse probabilities offline without affecting the execution path. Priority blocks with high reuse likelihood are protected using replications in shared LLC. Low priority blocks are invalidated to reduce their vulnerable time. A replication aware replacement policy is also developed to protect the replicated/reserved blocks from eviction. The scheme is evaluated in Multi2Sim 5.0 simulation framework with SPEC-CPU benchmarks considering bulk-CMOS, FDSOI and FinFET technologies. The results indicate 41.29% and 43.14% reductions in miss rate, 92.29% and 89.56% reductions in vulnerability, 5.73% and 6.31% increase in write-back rate, 17.55% and 16.27% reductions in average memory access time, 2.17% and 1.54% increase in Instructions-per-cycle (IPC) and 1.31% and 0.77% reductions in normalized-execution-cycle over the baseline (with SECDED protection) for integer and floating point benchmarks respectively. Marginal increase in power consumption by 1.33% and dynamic energy by 0.43% are observed at 6.32% area overhead.

Convolutional Neural Network-Based Deep Learning Engine for Mastoidectomy Instrument Recognition and Movement Tracking.

To develop a convolutional neural network-based computer vision model to recognize and track 2 mastoidectomy surgical instruments-the drill and the suction-irrigator-from intraoperative video recordings of mastoidectomies. Technological development and model validation. Academic center. Ten 1-minute videos of mastoidectomies done for cochlear implantation by varying levels of resident surgeons were collected. For each video, containing 900 frames, an open-access computer vision annotation tool was used to annotate the drill and suction-irrigator class images with bounding boxes. A mastoidectomy instrument tracking module, which extracts the center coordinates of bounding boxes, was developed using a feature pyramid network and layered with DETECTRON, an open-access faster-region-based convolutional neural network. Eight videos were used to train the model, and 2 videos were used for testing. Outcome measures included Intersection over Union (IoU) ratio, accuracy, and average precision. For an IoU of 0.5, the mean average precision for the drill was 99% and 86% for the suction-irrigator. The model proved capable of generating maps of drill and suction-irrigator stroke direction and distance for the entirety of each video. This computer vision model can identify and track the drill and suction-irrigator from videos of intraoperative mastoidectomies performed by residents with excellent precision. It can now be employed to retrospectively study objective mastoidectomy measures of expert and resident surgeons, such as drill and suction-irrigator stroke concentration, economy of motion, speed, and coordination, setting the stage for characterization of objective expectations for safe and efficient mastoidectomies.

A neural engine for image processing

A neural engine for image processing

CINE: A 4K-UHD Energy-Efficient Computational Imaging Neural Engine With Overlapped Stripe Inference and Structure-Sparse Kernel

CINE: A 4K-UHD Energy-Efficient Computational Imaging Neural Engine With Overlapped Stripe Inference and Structure-Sparse Kernel

Psycholinguistics, Corpus Linguistics, and Text Analytics: Potential for Interaction

This article explores the possibilities of interaction between psycholinguistics, corpus linguistics, and text mining. When it comes to big data processing, the recent advances in psycholinguistics must be taken into account. In our digital reality, we should no longer view text as a mere sum of individual language units that are assembled based on their similarities in meaning. The insights of psycholinguistics into text have been revised, as in-depth text analysis unravels the new roles of the sender and the recipient, bringing to the fore qualitative methods for processing large data arrays. Without studying the mechanisms of speech generation and perception, it is impossible to improve language neural networks and machine translation engines. Here, the potential of psycholinguistics for training specialists in text analytics is discussed. As an example, the advantages of integrating the disciplines “Psycholinguistics” and “Psycholinguistic Methods for Studying Language and Text” into the psychological and linguistic module of the Master’s program “Text Analytics in Education and Science”, which has been launched at Kazan Federal University (Russia), are shown. The psycholinguistic component of the discipline “Corpus Linguistics” for Master’s students is described.

Dynamic and Real-Time Object Detection Based on Deep Learning for Home Service Robots

Home service robots operating indoors, such as inside houses and offices, require the real-time and accurate identification and location of target objects to perform service tasks efficiently. However, images captured by visual sensors while in motion states usually contain varying degrees of blurriness, presenting a significant challenge for object detection. In particular, daily life scenes contain small objects like fruits and tableware, which are often occluded, further complicating object recognition and positioning. A dynamic and real-time object detection algorithm is proposed for home service robots. This is composed of an image deblurring algorithm and an object detection algorithm. To improve the clarity of motion-blurred images, the DA-Multi-DCGAN algorithm is proposed. It comprises an embedded dynamic adjustment mechanism and a multimodal multiscale fusion structure based on robot motion and surrounding environmental information, enabling the deblurring processing of images that are captured under different motion states. Compared with DeblurGAN, DA-Multi-DCGAN had a 5.07 improvement in Peak Signal-to-Noise Ratio (PSNR) and a 0.022 improvement in Structural Similarity (SSIM). An AT-LI-YOLO method is proposed for small and occluded object detection. Based on depthwise separable convolution, this method highlights key areas and integrates salient features by embedding the attention module in the AT-Resblock to improve the sensitivity and detection precision of small objects and partially occluded objects. It also employs a lightweight network unit Lightblock to reduce the network’s parameters and computational complexity, which improves its computational efficiency. Compared with YOLOv3, the mean average precision (mAP) of AT-LI-YOLO increased by 3.19%, and the detection precision of small objects, such as apples and oranges and partially occluded objects, increased by 19.12% and 29.52%, respectively. Moreover, the model inference efficiency had a 7 ms reduction in processing time. Based on the typical home activities of older people and children, the dataset Grasp-17 was established for the training and testing of the proposed method. Using the TensorRT neural network inference engine of the developed service robot prototype, the proposed dynamic and real-time object detection algorithm required 29 ms, which meets the real-time requirement of smooth vision.

Automatic planning for head and neck seed implant brachytherapy based on deep convolutional neural network dose engine.

Seed implant brachytherapy (SIBT) is an effective treatment modality for head and neck (H&N) cancers; however, current clinical planning requires manual setting of needle paths and utilizes inaccurate dose calculation algorithms. This study aims to develop an accurate and efficient deep convolutional neural network dose engine (DCNN-DE) and an automatic SIBT planning method for H&N SIBT. A cohort of 25 H&N patients who received SIBT was utilized to develop and validate the methods. The DCNN-DE was developed based on 3D-unet model. It takes single seed dose distribution from a modified TG-43 method, the CT image and a novel inter-seed shadow map (ISSM) as inputs, and predicts the dose map of accuracy close to the one from Monte Carlo simulations (MCS). The ISSM was proposed to better handle inter-seed attenuation. The accuracy and efficacy of the DCNN-DE were validated by comparing with other methods taking MCS dose as reference. For SIBT planning, a novel strategy inspired by clinical practice was proposed to automatically generate parallel or non-parallel potential needle paths that avoid puncturing bone and critical organs. A heuristic-based optimization method was developed to optimize the seed positions to meet clinical prescription requirements. The proposed planning method was validated by re-planning the 25 cases and comparing with clinical plans. The absolute percentage error in the TG-43 calculation for CTV V100 and D90 was reduced from 5.4% and 13.2% to 0.4% and 1.1% with DCNN-DE, an accuracy improvement of 93% and 92%, respectively. The proposed planning method could automatically obtain a plan in 2.5±1.5 min. The generated plans were judged clinically acceptable with dose distribution comparable with those of the clinical plans. The proposed method can generate clinically acceptable plans quickly with high accuracy in dose evaluation, and thus has a high potential for clinical use in SIBT.

Neural network models selection scheme for health mobile app development

Mobile healthcare application (mHealth app) assists the frontline health worker in providing necessary health services to the patient. Unfortunately, existing mHealth apps continue to have accuracy issues and limited number of disease detection systems. Thus, an intelligent disease diagnostics system may help medical staff as well as people in poor communities in rural areas. This study proposes a scheme for simultaneously selecting the best neural network models for intelligent disease detection systems on mobile devices. To find the best models for a given dataset, the proposed scheme employs neural network models capable of evolving altered neural network architectures. Eight neural network models are developed simultaneously and then implemented on the Android Studio platform. Mobile health applications use pre-trained neural network models to provide users with disease prediction results. The performance of the mobile application is measured against the existing available datasets. The trained neural network engines perform admirably, detecting 7 out of 8 diseases with high accuracy ranging from 86.00% to 100% and a low detection accuracy of 63%. The detection times vary from 0.01 to 0.057 seconds. The developed mHealth app may be used by health workers and patients to improve resource-poor community health services and patients' healthcare quality.

SCANet: Securing the Weights With Superparamagnetic-MTJ Crossbar Array Networks.

Deep neural networks (DNNs) form a critical infrastructure supporting various systems, spanning from the iPhone neural engine to imaging satellites and drones. The design of these neural cores is often proprietary or a military secret. Nevertheless, they remain vulnerable to model replication attacks that seek to reverse engineer the network's synaptic weights. In this article, we propose SCANet (Superparamagnetic-MTJ Crossbar Array Networks), a novel defense mechanism against such model stealing attacks by utilizing the innate stochasticity in superparamagnets. When used as the synapse in DNNs, superparamagnetic magnetic tunnel junctions (s-MTJs) are shown to be significantly more secure than prior memristor-based solutions. The thermally induced telegraphic switching in the s-MTJs is robust and uncontrollable, thus thwarting the attackers from obtaining sensitive data from the network. Using a mixture of both superparamagnetic and conventional MTJs in the neural network (NN), the designer can optimize the time period between the weight updation and the power consumed by the system. Furthermore, we propose a modified NN architecture that can prevent replication attacks while minimizing power consumption. We investigate the effect of the number of layers in the deep network and the number of neurons in each layer on the sharpness of accuracy degradation when the network is under attack. We also explore the efficacy of SCANet in real-time scenarios, using a case study on object detection.

Design of high performance and energy efficient convolution array for convolution neural network-based image inference engine

The energy efficiency of CNN-based inference engines predominately depends upon Giga-operations-per-second and power consumption. The sparse-based accelerator compresses the insignificant inputs (input feature maps & weights), skips the inefficient computations, and improves energy efficiency. A sparse accelerator for weights could impact the accuracy of the inference. Therefore, a sparse network for Input Feature Maps (IFMs) is considered. MATLAB-based sparsity analysis is done layer-wise on the pre-trained CNN models like AlexNet, VGG-16, VGG-19, ResNet-18 & ResNet-34. Layer-wise analysis reveals that ∼18%–90% of the IFMs are zeros. Besides, IFMs and Weights adopted the 16-bit Fix/Float data format to maintain an accuracy as close as 97% with Single Precision Floating Point (SPFP). A 3 × 1 Convolutional array with improved Zero-detect-Skip (CZS3×1) control units for multiplier and adder/subtractor arrays is proposed. The modified rectified linear unit (RELU) converts IFM values ≤ 0 to zero and sets Detection-Bit (DB) to 1. These DBs decide the mode of effective zero-skip operations in CZS3×1. A 3 × 3 Compressed Processing Element (CPE) is designed using three CZS3×1. The 20-CPEs convolution array architecture is implemented in 65 nm technology libraries. The performance of 90 Giga-operations per second (GOP/s) and energy efficiency of 3.42 Tera operations per second per watt (TOPS/W) were attained for the proposed CPE. The CPE with improved control strategy enhanced the performance by a factor of 2.45 while consuming 8.8 times less energy on average than the state-of-the-art CNN accelerators.

CPU frequency scheduling of real-time applications on embedded devices with temporal encoding-based deep reinforcement learning

Small devices are frequently used in IoT and smart-city applications to perform periodic dedicated tasks with soft deadlines. This work focuses on developing methods to derive efficient power-management methods for periodic tasks on small devices. We first study the limitations of the existing Linux built-in methods used in small devices. We illustrate three typical workload/system patterns that are challenging to manage with Linux’s built-in solutions. We develop a reinforcement-learning-based technique with temporal encoding to derive an effective DVFS governor even with the presence of the three system patterns. The derived governor uses only one performance counter, the same as the built-in Linux mechanism, and does not require an explicit task model for the workload. We implemented a prototype system on the Nvidia Jetson Nano Board and experimented with it with six applications, including two self-designed and four benchmark applications. Under different deadline constraints, our approach can quickly derive a DVFS governor that can adapt to performance requirements and outperform the built-in Linux approach in energy saving. On Mibench workloads, with performance slack ranging from 0.04 s to 0.4 s, the proposed method can save 3%–11% more energy compared to Ondemand. AudioReg and FaceReg applications tested have 5%–14% energy-saving improvement. We have open-sourced the implementation of our in-kernel quantized neural network engine. The codebase can be found at: https://github.com/coladog/tinyagent.

Blockchain and Machine Learning Inspired Secure Smart Home Communication Network.

With the increasing growth rate of smart home devices and their interconnectivity via the Internet of Things (IoT), security threats to the communication network have become a concern. This paper proposes a learning engine for a smart home communication network that utilizes blockchain-based secure communication and a cloud-based data evaluation layer to segregate and rank data on the basis of three broad categories of Transactions (T), namely Smart T, Mod T, and Avoid T. The learning engine utilizes a neural network for the training and classification of the categories that helps the blockchain layer with improvisation in the decision-making process. The contributions of this paper include the application of a secure blockchain layer for user authentication and the generation of a ledger for the communication network; the utilization of the cloud-based data evaluation layer; the enhancement of an SI-based algorithm for training; and the utilization of a neural engine for the precise training and classification of categories. The proposed algorithm outperformed the Fused Real-Time Sequential Deep Extreme Learning Machine (RTS-DELM) system, the data fusion technique, and artificial intelligence Internet of Things technology in providing electronic information engineering and analyzing optimization schemes in terms of the computation complexity, false authentication rate, and qualitative parameters with a lower average computation complexity; in addition, it ensures a secure, efficient smart home communication network to enhance the lifestyle of human beings.

A case study on the evaluation of Korean-English legal translations by generic and custom neural machine translation engines

This paper addresses the quality of Korean-English legal translation outputs generated by three neural machine translation engines, namely NMT1, NMT2 and NMT3. NMT1 and NMT2 are widely available multilingual generic machine translation engines, whereas NMT3 is a custom machine translation engine for legal translation. In both automatic and human evaluations of the three engines’ English translations of Korean statutes, the custom engine outperformed the other two NMTs. BLEU and METEOR scores revealed that NMT3’s output was more similar to human reference translation. Human evaluation results confirmed NMT3’s strong performance, outperforming the generic engines in accuracy and fluency while underperforming in terminology. However, further analysis revealed that NMT3 was superior in handling terms comprising compound nouns and sentences over 25 words. Although the custom engine’s performance still requires substantial human involvement for quality translation, it showed a potential for the legal domain, demonstrating human parity in short sentences. The growing demand for speedy translation under budget constraints calls for further research on technology-based translation tools and human-computer interaction for legal translation.

Robust Classification and Detection of Big Medical Data Using Advanced Parallel K-Means Clustering, YOLOv4, and Logistic Regression.

Big-medical-data classification and image detection are crucial tasks in the field of healthcare, as they can assist with diagnosis, treatment planning, and disease monitoring. Logistic regression and YOLOv4 are popular algorithms that can be used for these tasks. However, these techniques have limitations and performance issue with big medical data. In this study, we presented a robust approach for big-medical-data classification and image detection using logistic regression and YOLOv4, respectively. To improve the performance of these algorithms, we proposed the use of advanced parallel k-means pre-processing, a clustering technique that identified patterns and structures in the data. Additionally, we leveraged the acceleration capabilities of a neural engine processor to further enhance the speed and efficiency of our approach. We evaluated our approach on several large medical datasets and showed that it could accurately classify large amounts of medical data and detect medical images. Our results demonstrated that the combination of advanced parallel k-means pre-processing, and the neural engine processor resulted in a significant improvement in the performance of logistic regression and YOLOv4, making them more reliable for use in medical applications. This new approach offers a promising solution for medical data classification and image detection and may have significant implications for the field of healthcare.