Memory Efficiency Research Articles

Сегнетоэлектрические материалы в современных вычислениях

Background: Ferroelectric materials, known for their spontaneous electric polarization, are reshaping the landscape of modern computing. These materials have unique qualities, such as non-volatility, high-speed operation, and energy efficiency, which are critical for advancing modern computer technology. Objective: This study intends to investigate integrating ferroelectric materials into current computer systems, emphasizing to improve memory devices and processors and solve the accompanying technological obstacles and possibilities. Methods: A thorough analysis of existing literature and experimental data was carried out to determine the capabilities and limitations of ferroelectric materials in computing applications. Key performance indicators examined include polarization retention, energy consumption, scalability, and integration with existing semiconductor technology. Results: The results show that ferroelectric materials considerably increase memory device performance and efficiency by allowing quicker write/read operations and lowering power use. However, concerns such as material deterioration, data retention, and integration complexity with silicon-based technology remain. Conclusion: Ferroelectric materials provide exciting prospects for the next generation of computer technology. While they provide significant gains in memory and processing power, addressing the technological obstacles is critical to their effective adoption into mainstream computer applications. Further research and development are needed to solve these issues and fully realize the promise of ferroelectric materials in improving computer performance.

Revolutionizing tomato disease detection in complex environments.

In the current agricultural landscape, a significant portion of tomato plants suffer from leaf diseases, posing a major challenge to manual detection due to the task's extensive scope. Existing detection algorithms struggle to balance speed with accuracy, especially when identifying small-scale leaf diseases across diverse settings. Addressing this need, this study presents FCHF-DETR (Faster-Cascaded-attention-High-feature-fusion-Focaler Detection-Transformer), an innovative, high-precision, and lightweight detection algorithm based on RT-DETR-R18 (Real-Time-Detection-Transformer-ResNet18). The algorithm was developed using a carefully curated dataset of 3147 RGB images, showcasing tomato leaf diseases across a range of scenes and resolutions. FasterNet replaces ResNet18 in the algorithm's backbone network, aimed at reducing the model's size and improving memory efficiency. Additionally, replacing the conventional AIFI (Attention-based Intra-scale Feature Interaction) module with Cascaded Group Attention and the original CCFM (CNN-based Cross-scale Feature-fusion Module) module with HSFPN (High-Level Screening-feature Fusion Pyramid Networks) in the Efficient Hybrid Encoder significantly enhanced detection accuracy without greatly affecting efficiency. To tackle the challenge of identifying challenging samples, the Focaler-CIoU loss function was incorporated, refining the model's performance throughout the dataset. Empirical results show that FCHF-DETR achieved 96.4% Precision, 96.7% Recall, 89.1% mAP (Mean Average Precision) 50-95 and 97.2% mAP50 on the test set, with a reduction of 9.2G in FLOPs (floating point of operations) and 3.6M in parameters. These findings clearly demonstrate that the proposed method improves detection accuracy and reduces computational complexity, addressing the dual challenges of precision and efficiency in tomato leaf disease detection.

A lightweight convolutional neural network-based feature extractor for visible images

Feature extraction networks (FENs), as the first stage in many computer vision tasks, play critical roles. Previous studies regarding FENs employed deeper and wider networks to attain higher accuracy, but their approaches were memory-inefficient and computationally intensive. Here, we present an accurate and lightweight feature extractor (RoShuNet) for visible images based on ShuffleNetV2. The provided improvements are threefold. To make ShuffleNetV2 compact without degrading its feature extraction ability, we propose an aggregated dual group convolutional module; to better aid the channel interflow process, we propose a γ-weighted shuffling module; to further reduce the complexity and size of the model, we introduce slimming strategies. Classification experiments demonstrate the state-of-the-art (SOTA) performance of RoShuNet, which yields an increase in accuracy and reduces the complexity and size of the model compared to those of ShuffleNetV2. Generalization experiments verify that the proposed method is also applicable to feature extraction tasks in semantic segmentation and multiple-object tracking scenarios, achieving comparable accuracy to that of other approaches with more memory and greater computational efficiency. Our method provides a novel perspective for designing lightweight models.

Task-Based Attentional Control: The Role of Anxiety and Age.

Attentional Control Theory (ACT) posits that anxiety impacts cognitive functioning through interference in working memory and processing efficiency, resulting in performance deficits in set-shifting and inhibition. Few studies have examined the effects of anxiety on set-shifting and inhibition in clinical samples or how these relationships might be affected by age. The current study tested whether increased age, elevated anxiety, and their interaction were associated with reduced performance on measures of set-shifting and inhibition. Symptom and neuropsychological testing data were obtained from outpatient participants presenting at an academic medical center (N = 521; mean age = 50.39years, SD = 22.35, range = 18-90; 47.4% female; 78.3% White). The Trail Making Test Difference score was used to assess set-shifting and the Stroop Color-Word Test Interference score was used to assess inhibition. After controlling for demographic variables, ADHD diagnosis, depression symptoms, and Mild Cognitive Impairment (MCI), both age and anxiety were significant predictors of set-shifting (β = 0.45 and β = 0.18, respectively, ps < 0.001) and inhibition (β = -0.37, p < 0.001 and β = -0.19, p = 0.001, respectively). No interaction was found between age and anxiety in the prediction of set-shifting or inhibition. Congruent with ACT, anxiety was associated with worse performance on measures of set-shifting and inhibition. Older age was an independent predictor of worse set-shifting and inhibition but did not moderate the relationship between anxiety and attentional control, suggesting that anxiety adversely affected working memory and processing efficiency equivalently across the adult lifespan. The results highlight the importance of anxiety assessment in neuropsychological evaluation in patients of all ages.

Foundation model-driven distributed learning for enhanced retinal age prediction.

The retinal age gap (RAG) is emerging as a potential biomarker for various diseases of the human body, yet its utility depends on machine learning models capable of accurately predicting biological retinal age from fundus images. However, training generalizable models is hindered by potential shortages of diverse training data. To overcome these obstacles, this work develops a novel and computationally efficient distributed learning framework for retinal age prediction. The proposed framework employs a memory-efficient 8-bit quantized version of RETFound, a cutting-edge foundation model for retinal image analysis, to extract features from fundus images. These features are then used to train an efficient linear regression head model for predicting retinal age. The framework explores federated learning (FL) as well as traveling model (TM) approaches for distributed training of the linear regression head. To evaluate this framework, we simulate a client network using fundus image data from the UK Biobank. Additionally, data from patients with type 1 diabetes from the UK Biobank and the Brazilian Multilabel Ophthalmological Dataset (BRSET) were utilized to explore the clinical utility of the developed methods. Our findings reveal that the developed distributed learning framework achieves retinal age prediction performance on par with centralized methods, with FL and TM providing similar performance (mean absolute error of 3.57 ± 0.18 years for centralized learning, 3.60 ± 0.16 years for TM, and 3.63 ± 0.19 years for FL). Notably, the TM was found to converge with fewer local updates than FL. Moreover, patients with type 1 diabetes exhibited significantly higher RAG values than healthy controls in all models, for both the UK Biobank and BRSET datasets (P < .001). The high computational and memory efficiency of the developed distributed learning framework makes it well suited for resource-constrained environments. The capacity of this framework to integrate data from underrepresented populations for training of retinal age prediction models could significantly enhance the accessibility of the RAG as an important disease biomarker.

ADVANCING MEMORY DENSITY: A NOVEL DESIGN FOR MULTIPLE-BIT-PER-CELL PHASE CHANGE MEMORY

Multiple-bit-per-cell phase-change memory (MPCM) has emerged as a promising solution to address the escalating demands for high-density, low-power, and fast-access memory in modern computing and data storage systems. This paper presents a novel device design aimed at enabling multiple bits per cell in phase-change memory, thereby significantly enhancing memory density while maintaining performance and reliability. Leveraging innovative material compositions and advanced fabrication techniques, the proposed design demonstrates the potential to push the boundaries of memory capacity, efficiency, and scalability. Through comprehensive simulation analysis and performance evaluations, we showcase the feasibility and advantages of the new device design, highlighting its potential to revolutionize memory architectures and meet the evolving needs of next-generation computing systems.

An Area-Efficient and Configurable Number Theoretic Transform Accelerator for Homomorphic Encryption

Homomorphic Encryption (HE) allows for arbitrary computation of encrypted data, offering a method for preserving privacy in cloud computations. However, efficiency remains a significant obstacle, particularly with the polynomial multiplication of large parameter sets, which occupies substantial computing and memory overhead. Prior studies proposed the use of Number Theoretic Transform (NTT) to accelerate polynomial multiplication, which proved efficient, owing to its low computational complexity. However, these efforts primarily focused on NTT designs for small parameter sets, and configurability and memory efficiency were not considered carefully. This paper focuses on designing a unified NTT/Inverse NTT (INTT) architecture with high area efficiency and configurability, which is more suitable for HE schemes. We adopt the Constant-Geometry (CG) NTT algorithm and propose a conflict-free access pattern, demonstrating a 16.7% reduction in coefficients of storage capacity compared to the state-of-the-art CG NTT design. Additionally, we propose a twiddle factor generation strategy to minimize storage for Twiddle Factors (TFs). The proposed architecture offers configurability of both compile time and runtime, allowing for the deployment of varying parallelism and parameter sets during compilation while accommodating a wide range of polynomial degrees and moduli after compilation. Experimental results on the Xilinx FPGA show that our design can achieve higher area efficiency and configurability compared with previous works. Furthermore, we explore the performance difference between precomputed TFs and online-generated TFs for the NTT architecture, aiming to show the importance of online generation-based NTT architecture in HE applications.

Memory-Efficient Trust Management for IoT Nodes Enabling Secure Communication in Trusted Environments

The Internet of Things (IoT) is changing quickly, so making sure that gadgets that are linked can talk to each other safely is very important. This paper describes a new way to handle trust that is specifically made for IoT nodes that work in safe settings. Conventional models frequently have a parcel of additional memory needs and are difficult to compute, which is made more regrettable by IoT gadgets that do not have a part of assets. Our proposed approach bargains with these issues by employing a belief administration framework that doesn't utilize a parcel of memory and finds an adjustment between security needs and the restricted assets that IoT hubs have. The foremost inventive thing around our strategy is that it can alter belief levels on the fly based on modern data and past experiences. Our framework takes up as small memory as conceivable whereas still having solid security measures since it employs lightweight cryptography conventions and quick information structures. This lets Internet of Things (IoT) gadgets utilize secure communication strategies without abating them down or making them less dependable. Context-aware belief assessment could be a key portion of our framework. This implies that belief levels are continually changed based on real-time encompassing variables and the way hubs are connected with each other. This adaptable alter makes beyond any doubt that belief choices remain in line with changing working conditions and threat scenarios. We moreover made beyond any doubt that our strategy works well with current IoT plans, permitting distinctive sorts of systems to work together and develop. We demonstrate that our system works in a wide range of IoT application cases by running a part of models and real-world tests on it. This study shows an easy-to-use and expandable way to improve trust management in IoT settings, with a focus on memory efficiency without losing safety. By using our approach, IoT stakeholders can reduce the risks that come with old trust models. This will make IoT environments more reliable and adaptable in the future.

Ultratough Supramolecular Polyurethane Featuring an Interwoven Network with Recyclability, Ideal Self-Healing and Editable Shape Memory Properties.

Developing multifunctional polymers with excellent mechanical properties, outstanding shape memory characteristics, and good self-healing properties is a formidable challenge. Inspired by the woven cross-linking strategy, a series of supramolecular polyurethane (PU) with an interwoven network structure composed of covalent and supramolecular cross-linking nodes have been successfully synthesized by introducing the ureido-pyrimidinone (UPy) motifs into the PU skeleton. The best-performing sample exhibited ultrahigh strength (∼77.2 MPa) and toughness (∼312.7 MJ m-3), along with an ideal self-healing efficiency (up to 90.8% for 6 h) and satisfactory temperature-responsive shape memory effect (shape recovery rates up to 96.9%). Furthermore, it ensured recyclability. These favorable properties are mainly ascribed to the effective dissipation of strain energy due to the disassembly and reconfiguration of supramolecular nodes (i.e., quadruple hydrogen bonds (H-bonds) between UPy units), as well as the covalent cross-linking nodes that maintain the integrity of the polymer network structure. Thus, our work provides a universal strategy that breaks through the traditional contradictions and paves the way for the commercialization of high-performance multifunctional PU elastomers.

Coarse mesh finite difference acceleration of the random ray method

This paper presents the development and evaluation of the Coarse Mesh Finite Difference (CMFD) acceleration method to The Random Ray Method (TRRM). We demonstrate its effectiveness in accelerating the convergence of the fission sources and reducing the real statistical error in neutron transport calculations. TRRM treats the angular variable of the neutron flux stochastically and performs batch sampling of characteristic rays to transform the Method of Characteristics (MOC) into a stochastic process, that is capable of making significant strides in memory efficiency and computational performance for some problems. Despite its advantages, TRRM exhibits a potential challenge with a large number of inactive cycles and inherent inter-cycle correlation much like the Monte Carlo method. To address this, the CMFD acceleration method is explored and demonstrated to dramatically reduce the number of required inactive cycles and diminish inter-cycle correlation. Results from the numerical analysis of a 2D C5G7 core problem indicate that the application of CMFD leads to enhanced convergence, with the integration of a CMFD acceleration step every cycle offering the most substantial reduction in statistical noise and error. The study reveals that applying CMFD with every cycle effectively resolves the issue of needing inactive cycles and significantly lowers the inter-cycle correlation, thereby providing a more accurate estimation of standard deviation for pin power distributions. We conclude that using CMFD not only minimizes the number of inactive cycles of TRRM – much like normal Monte Carlo transport – but also lowers real statistical error effectively. For a targeted maximum standard deviation of 0.1% in the pin power, the addition of CMFD can decrease the number of necessary active cycles by 41% compared to standard TRRM, as demonstrated by the 2D C5G7 benchmark analysis.

Combining signal decomposition and deep learning model to predict noisy runoff coefficient

Predicting runoff coefficient (Rc), as an indicator of the catchment’s response to the rainfall-runoff process, remains a persistent challenge using different modelling techniques, especially in catchments with strong human manipulation. This study investigates the efficiency of the Long Short-Term Memory (LSTM) method in predicting Rc for the Rur catchment, in Germany. The period from 1961 to 2021 is considered, which is subject to human intervention and significant urbanization especially in the northern part of the catchment. An LSTM structure is defined by employing inputs at a monthly resolution including temperature, precipitation, soil water storage, and total evaporation with a look-back window of 1 to 6 months to model noisy Rc data of the study area. Two approaches using either undecomposed or decomposed Rc were employed in conjunction with the LSTM method, to mitigate the impact of noise associated with Rc. The results show that in the case of undecomposed Rc, the best performance of the LSTM structure was obtained with a 4-month look-back window, yielding Nash-Sutcliffe efficiency (NSE) of 0.55, 0.46, and 0.15 for training, validation, and test sets, respectively. These results highlight inadequate accuracy in accounting for the presence of noise in Rc. Therefore, in the second novel approach, we used maximal overlap discrete wavelet transform (MODWT) to decompose the Rc up to level 3 to reduce the complexity and distribute the noise effects across each level. The new approach showed high accuracy in modelling noisy data of Rc with NSE values of 0.97, 0.95, and 0.90 for training, validation, and test sets, respectively. The obtained results underscore the pivotal role of decomposition techniques in conjunction with LSTM to account for the presence of noise, especially in catchments with strong human manipulation.

The role of attention and verbal rehearsal in remembering more valuable item-colour binding

ABSTRACT Selectively remembering more valuable information can improve memory efficiency. Such value effects have been observed on long-term memory for item-colour binding, but the possible contributory factors are unclear. The current study explored contributions from attention (Experiment 1) and verbal rehearsal (Experiment 2). Across two experiments, memory was superior for item-colour bindings that were associated with high (relative to low) point values at encoding, both in an immediate test and a delayed re-test. When availability of attentional resources was reduced during encoding, value only influenced immediate and not delayed memory (Experiment 1). This indicates that a transient value effect can be obtained with little attentional resources, but attentional resources are involved in creating a longer lasting effect. When articulatory suppression was implemented during encoding (Experiment 2), value effects were somewhat reduced in the immediate test and abolished in the delayed re-test, suggesting a role for verbal rehearsal in value effects on item-colour binding memory. These patterns of value effects did not interact with encoding presentation format (i.e., sequential vs. simultaneous presentation of objects). Together, these results suggest that attentional resources and verbal rehearsal both contribute to value effects on item-colour binding memory, with varying impacts on the durability of these effects.

Algorithmic optimization of quantum optical storage in solids

Quantum memory devices with high storage efficiency and bandwidth are essential elements for future quantum networks. Solid-state quantum memories can provide broadband storage, but they primarily suffer from low storage efficiency. We use passive optimization and algorithmic optimization techniques to demonstrate nearly a sixfold enhancement in quantum memory efficiency. In this regime, we demonstrate coherent and single-photon-level storage with a high signal-to-noise ratio. The optimization technique presented here can be applied to most solid-state quantum memories to significantly improve the storage efficiency without compromising the memory bandwidth. Published by the American Physical Society 2024

Biobased photothermal responsive shape memory polythioether/MXene nanocomposites with self-extinguishing performance

Thermosetting shape memory polymers (TS-SMPs) have a higher shape fixity ratio (Rf) and shape recovery (Rr) due to their structure integrity and thermal stability, and the glass transition temperature (Tg) usually influences the transition temperature (Ttrans) of TS-SMPs. Therefore, it is significant to tailor the Tg range and mechanical properties for TS-SMPs. In this work, the inherently flame-retarding bio-based MXene-reinforced polythioether nanocomposites with excellent shape memory performance, driven by photothermal response, were designed. Firstly, a photopolymerizable MXene nanomonomer (MPS-MXene) was obtained by introducing 3-(methacryloxy) propyltrimethoxysilane onto the surface of the MXene monolayer. Then, the bis(4-allyl-2-methoxyphenyl) phenyl phosphonate (BEP) was synthesized by eugenol (a lignin derivative) with phenyl phosphonic dichloride. The polythioether/MXene composites (BEP-T-xM) were synthesized through in-situ thiol-ene photopolymerization of MPS-MXene, BEP, and tris[2-(3-mercaptopropionyloxy)ethyl]isocyanurate. Surprisingly, compared with neat polythioether, the Tg of BEP-T-0.2 M composite increased from 19.6 to 34.6 °C, when only 0.2 wt% MPS-MXene was added. The photothermal conversion efficiency (η = 73.8 %) and shape memory performance with Rr (>98 %) were greatly improved. Consequently, the interesting phenomena such as “sunflowers growing towards the sunlight” and “thermal shrinkage” can be observed. Meanwhile, the tensile strength and Young’s modulus increased significantly by 141 % and 91 %, respectively, due to the good dispersion of rigid MPS-MXene nanosheets. Furthermore, all the polythioether/MXene composites showed self-extinguishing performance. This study will provide a novel strategy for designing biobased photothermal-responsive TS-SMPs.

FSDedup: Feature-Aware and Selective Deduplication for Improving Performance of Encrypted Non-Volatile Main Memory

Enhancing the endurance, performance, and energy efficiency of encrypted Non-Volatile Main Memory (NVMM) can be achieved by minimizing written data through inline deduplication. However, existing approaches applying inline deduplication to encrypted NVMM suffer from substantial performance degradation due to high computing, memory footprint, and index-lookup overhead to generate, store, and query the cryptographic hash (fingerprint). In the preliminary ESD [ 14 ], we proposed the Error Correcting Code (ECC) assisted selective deduplication scheme, utilizing the ECC information as a fingerprint to identify similar data effectively and then leveraging the selective deduplication technique to eliminate a large amount of redundant data with high reference counts. In this article, we proposed FSDedup. Compared with ESD, FSDedup could leverage the prefetch cache to reduce the read overhead during similarity comparison and utilize the cache refresh mechanism to identify further and eliminate more redundant data. Extensive experimental evaluations demonstrate that FSDedup can enhance the performance of the NVMM system further than the ESD. Experimental results show that FSDedup can improve both write and read speed by up to 1.8×, enhance Instructions Per Cycle by up to 1.5×, and reduce energy consumption by up to 2.0×, compared to ESD.

The Effects of Auditory, Visual, and Cognitive Distractions on Cybersickness in Virtual Reality.

Cybersickness (CS) is one of the challenges that has hindered the widespread adoption of Virtual Reality (VR). Consequently, researchers continue to explore novel means to mitigate the undesirable effects associated with this affliction, one that may require a combination of remedies as opposed to a solitary stratagem. Inspired by research probing into the use of distractions as a means to control pain, we investigated the efficacy of this countermeasure against CS, studying how the introduction of temporally time-gated distractions affects this malady during a virtual experience featuring active exploration. Downstream of this, we studied how other aspects of the VR experience are affected by this intervention. We discuss the results of a between-subjects study manipulating the presence, sensory modality, and nature of periodic and short-lived (5-12 seconds) distractor stimuli across four experimental conditions: 1) no-distractors (ND); 2) auditory distractors (AD); 3) visual distractors (VD); 4) cognitive distractors (CD). Two of these conditions (VD and AD) formed a yoked control design wherein every matched pair of 'seers' and 'hearers' was periodically exposed to distractors that were identical in terms of content, temporality, duration, and sequence. In the CD condition, each participant had to periodically perform a 2-back working memory task, the duration and temporality of which was matched to distractors presented in each matched pair of the yoked conditions. These three conditions were compared to a baseline control group featuring no distractions. Results indicated that the reported sickness levels were lower in all three distraction groups in comparison to the control group. The intervention also increased the amount of time users were able to endure the VR simulation and avoided causing detriments to spatial memory and virtual travel efficiency. Overall, it appears that it may be possible to make users less consciously aware and bothered by the symptoms of CS, thereby reducing its perceived severity.

Energy-Efficient Sleep Apnea Detection Using a Hyperdimensional Computing Framework Based on Wearable Bracelet Photoplethysmography.

Sleep apnea syndrome (SAS) is a common sleep disorder, which has been shown to be an important contributor to major neurocognitive and cardiovascular sequelae. Considering current diagnostic strategies are limited with bulky medical devices and high examination expenses, a large number of cases go undiagnosed. To enable large-scale screening for SAS, wearable photoplethysmography (PPG) technologies have been used as an early detection tool. However, existing algorithms are energy-intensive and require large amounts of memory resources, which are believed to be the major drawbacks for further promotion of wearable devices for SAS detection. In this paper, an energy-efficient method of SAS detection based on hyperdimensional computing (HDC) is proposed. Inspired by the phenomenon of chunking in cognitive psychology as a memory mechanism for improving working memory efficiency, we proposed a one-dimensional block local binary pattern (1D-BlockLBP) encoding scheme combined with HDC to preserve dominant dynamical and temporal characteristics of pulse rate signals from wearable PPG devices. Our method achieved 70.17 % accuracy in sleep apnea segment detection, which is comparable with traditional machine learning methods. Additionally, our method achieves up to 67× lower memory footprint, 68× latency reduction, and 93× energy saving on the ARM Cortex-M4 processor. The simplicity of hypervector operations in HDC and the novel 1D-BlockLBP encoding effectively preserve pulse rate signal characteristics with high computational efficiency. This work provides a scalable solution for long-term home-based monitoring of sleep apnea, enhancing the feasibility of consistent patient care.

Binary segmentation of relief patterns on point clouds

Analysis of 3D textures, also known as relief patterns is a challenging task that requires separating repetitive surface patterns from the underlying global geometry. Existing works classify entire surfaces based on one or a few patterns by extracting ad-hoc statistical properties. Unfortunately, these methods are not suitable for objects with multiple geometric textures and perform poorly on more complex shapes. In this paper, we propose a neural network for binary segmentation to infer per-point labels based on the presence of surface relief patterns. We evaluated the proposed architecture on a high resolution point cloud dataset, surpassing the state-of-the-art, while maintaining memory and computation efficiency.

Getting value out of working memory through strategic prioritisation: Implications for storage and control.

Working memory is an active system responsible for the temporary maintenance and processing of information in the support of cognition and action. In keeping with this, a growing body of research has explored the close links between working memory and attention, and how these might be harnessed to impact performance and possibly improve working memory efficiency. This is theoretically and practically important, given that working memory is a central hub in complex cognition yet is extremely capacity- and resource-limited. We review work carried out over the last 10 years or so looking at how high "value" items in working memory can be strategically prioritised through selective attention, drawing principally from visual working memory paradigms with young adult participants, while also discussing how the core effects extend to different task domains and populations. A consistent set of core findings emerges, with improved memory for items that are allocated higher value but no change in overall task performance, and a recency advantage regardless of point allocation when items are encountered sequentially. Value-directed prioritisation is effortful, under top-down strategic control, and appears to vary with perceptual distraction and executive load. It is driven by processes operating during encoding, maintenance, and retrieval, though the extent to which these are influenced by different features of the task context remains to be mapped out. We discuss implications for working memory, attention, and strategic control, and note some possible future directions of travel for this promising line of research.

Benchmarking Test-Time DNN Adaptation at Edge with Compute-In-Memory

The prediction accuracy of deep neural networks (DNNs) deployed at the edge can deteriorate over time due to shifts in the data distribution. For heightened robustness, it’s crucial for DNNs to continually refine and improve their predictive capabilities. However, adaptation in resource-limited edge environments is fraught with challenges: (i) new labeled data might be unavailable; (ii) on-device adaptation is a necessity as cloud connections may be inaccessible; and (iii) the adaptation procedure should prioritize speed, memory efficiency, and energy conservation. Compute-In-Memory (CIM) has recently garnered attention for its computational efficacy and superior operational bandwidth. Additionally, emerging lightweight unsupervised DNN adaptation techniques during test-time have showcased promising results in enhancing model accuracy for data with noise. This article pioneers a holistic benchmarking exploration of these methods, assessing their performance and energy efficacy across diverse CIM architectures in edge and autonomous systems. Our findings reveal that the proposed adaptation strategies can adapt to both environment shifts and inherent hardware noise. Engaging in a thorough cross-layer algorithm-hardware-technology co-design space exploration, we highlight pivotal trade-offs among accuracy, performance, and energy for various DNN adaptation techniques and CIM configurations.