Search

Federated Learning (FL) allows machine learning models to train locally on individual mobile devices, synchronizing model updates via a shared server. This approach safeguards user privacy; however, it also generates a heterogeneous training environment due to the varying performance capabilities across devices. As a result, straggler devices with lower performance often dictate the overall training time in FL. In this work, we aim to alleviate this performance bottleneck due to stragglers by dynamically balancing the training load across the system. We introduce Invariant Dropout, a method that extracts a sub-model based on the weight update threshold, thereby minimizing potential impacts on accuracy. Building on this dropout technique, we develop an adaptive training framework, Federated Learning using Invariant Dropout (FLuID). FLuID offers a lightweight sub-model extraction to regulate computational intensity, thereby reducing the load on straggler devices without affecting model quality. Our method leverages neuron updates from non-straggler devices to construct a tailored sub-model for each straggler based on client performance profiling. Furthermore, FLuID can dynamically adapt to changes in stragglers as runtime conditions shift. We evaluate FLuID using five real-world mobile clients. The evaluations show that Invariant Dropout maintains baseline model efficiency while alleviating the performance bottleneck of stragglers through a dynamic, runtime approach.

Background and Purpose: This work aims at optimizing and studying the feasibility of imaging the brachial plexus at 1.5T using 3D nerve-SHeath signal increased with INKed rest-tissue RARE imaging (3D SHINKEI) neurography sequence by comparing with routine sequences. Materials and Methods: The study was performed on a 1.5T Achieva scanner. It was designed in two parts: (a) Optimization of SHINKEI sequence at 1.5T; and (b) Feasibility study of the optimized SHINKEI sequence for generating clinical quality magnetic resonance neurography images at 1.5T. Simulations and volunteer experiments were conducted to optimize the T2 preparation duration for optimum nerve-muscle contrast at 1.5T. Images from the sequence under study and other routine sequences from 24 patients clinically referred for brachial plexus imaging were scored by a panel of radiologists for diagnostic quality. Injury detection efficacy of these sequences were evaluated against the surgical information available from seven patients. Results: T2 preparation duration of 50 ms gives the best contrast to noise between nerve and muscle. The images of 3D SHINKEI and short-term inversion recovery turbo spin-echo sequences are of similar diagnostic quality but significantly better than diffusion weighted imaging with background signal suppression. In comparison with the surgical findings, 3D SHINKEI has the lowest specificity; however, it had the highest sensitivity and predictive efficacy compared to other routine sequences. Conclusion: 3D SHINKEI sequence provides a good nerve–muscle contrast and has high predictive efficacy of nerve injury, indicating that it is a potential screening sequence candidate for brachial plexus scans at 1.5T also.

Recommendation systems are unique as they show a conflation of compute and memory intensity due to their deep learning and massive embedding tables. Training these models typically involve a hybrid CPU-GPU mode, where GPUs accelerate the deep learning portion and the CPUs store and process the memory-intensive embedding tables. The hybrid mode incurs a substantial CPU-to-GPU transfer time and relies on main memory bandwidth to feed embeddings to GPU for deep learning acceleration. Alternatively, we can store the entire embeddings across GPUs to avoid the transfer time and utilize the GPU's High Bandwidth Memory (HBM). This approach requires GPU-to-GPU backend communication and scales the number of GPUs with the size of the embedding tables. To overcome these concerns, this paper offers a heterogeneous acceleration pipeline, called Hotline. Hotline leverages the insight that only a small number of embedding entries are accessed frequently, and can easily fit in a single GPU's HBM. Hotline implements a data-aware and model-aware scheduling pipeline that utilizes the (1) CPU main memory for not-frequently-accessed embeddings and (2) GPUs' local memory for frequently-accessed embeddings. Hotline improves the training throughput by dynamically stitching the execution of popular and not-popular inputs through a novel hardware accelerator and feeding to the GPUs. Results on real-world datasets and recommender models show that Hotline reduces the average training time by 3x and 1.8x in comparison to Intel-optimized CPU-GPU DLRM and HugeCTR-optimized GPU-only baseline, respectively. Hotline increases the overall training throughput to 35.7 epochs/hour in comparison to 5.3 epochs/hour for the Intel-optimized DLRM baseline

Quantum computers can speed up computationally hard problems. However, to realize their full potential, we must mitigate qubit errors (from noise) by developing noise-aware algorithms, compilers, and architectures. Thus, simulating quantum programs on classical computers with different noise models is a de-facto tool that is used by researchers and practitioners. Unfortunately, noisy quantum simulators iteratively execute the same circuit across multiple trials (shots), thereby incurring high-performance overheads. To address this, we propose a noisy simulation technique called Tree-Based Quantum Circuit Simulation (TQSim). TQSim exploits the reusability of the intermediate results during the noisy simulation and reduces computation. TQSim dynamically partitions a circuit into several subcircuits. It then reuses the intermediate results from these subcircuits during computation. As compared to a noisy Qulacs-based baseline simulator, TQSim achieves an average speedup of 2.51x across 48 different benchmark circuits. Additionally, across benchmarks, TQSim produces results with a normalized fidelity that is within the 0.016 range of the baseline normalized fidelity.

Scalable Solid-State Drives (SSDs) have revolutionized the way we store and access our data across datacenters and handheld devices. Unfortunately, scaling technology can have a significant environmental impact. Across the globe, most semiconductor manufacturing use electricity that is generated from coal and natural gas. For instance, manufacturing a Gigabyte of Flash emits 0.16 Kg CO$_2$ and is a significant fraction of the total carbon emission in the system. We estimate that manufacturing storage devices has resulted in 20 million metric tonnes of CO$_2$ emissions in 2021 alone. To better understand this concern, this paper compares the sustainability trade-offs between Hard Disk Drives (HDDs) and SSDs and recommends methodologies to estimate the embodied carbon costs of the storage system. In this paper, we outline four possible strategies to make storage systems sustainable. First, this paper recommends directions that help select the right medium of storage (SSD vs HDD). Second, this paper proposes lifetime extension techniques for SSDs. Third, this paper advocates for effective and efficient recycling and reuse of high-density multi-level cell-based SSDs. Fourth, specifically for hand-held devices, this paper recommends leveraging elasticity in cloud storage., Comment: In the proceedings of the 1st Workshop on Sustainable Computer Systems Design and Implementation (HotCarbon 2022)

2053 Background: Comprehensive molecular profiling reveals significant differences in treatment response among GBM patients. A mechanistic multi-omics biology model allows biosimulation of molecular effects of cell signaling, drugs and radiation on patient-specific in silico diseased cells. The Cellworks Singula Therapy Response Index (TRI) identifies the magnitude of disease control and survival for specific anti-tumor strategies. TRI ranks the anticipated outcome of each therapy with a continuous TRI Score, 0 to 100, for each patient’s unique genomic network. Methods: TRI’s ability to predict OS and PFS was prospectively evaluated in a retrospective cohort of 270 IDH wildtype GBM patients from The Cancer Genome Atlas (TCGA) with known clinical outcomes treated with physician prescribed therapies (PPT). The median age was 57.5 years for 162 males and 108 females. There were 73 MGMT methylated with median OS deceased of 17.1 months and living of 9.5 months and median PFS of 6.5 months. There were 197 MGMT unmethylated with median OS deceased of 14.0 months and living of 13.6 months and median PFS of 6.0 months. Stratified random sampling was used to split the data into independent training (N = 153) and validation (N = 117) subjects. Multivariate Cox Proportional Hazard and Proportional Odds models were used to model OS and PFS as a function of the pre-defined Singula TRI and clinical thresholds. Cox Proportional Hazards (PH) regression and likelihood ratio (LR) tests were used on the independent validation subjects to assess the hypothesis that Singula is predictive of OS and PFS above and beyond standard clinical factors. Results: In the validation set, Singula TRI was significantly predictive of OS and PFS in univariate analyses and remained significantly predictive in multivariate analyses which included age, sex, MGMT methylation status and drug class. Singula TRI facilitates selection of optimal personalized therapies by providing patient-specific estimates of OS and PFS for 18 NCCN guideline GBM therapies. Conclusions: Cellworks Singula was strongly predictive of OS and PFS and provided predictive value beyond physician prescribed therapy, patient age, sex and MGMT methylation status. This information may be used to estimate increases in OS and PFS when comparing Singula TRI recommended therapies versus standard care. These positive results suggest the utility of biosimulation-informed therapy selection to improve survival of GBM and merits validation in prospective clinical studies. [Table: see text]

Recommender models are commonly used to suggest relevant items to a user for e-commerce and online advertisement-based applications. These models use massive embedding tables to store numerical representation of items' and users' categorical variables (memory intensive) and employ neural networks (compute intensive) to generate final recommendations. Training these large-scale recommendation models is evolving to require increasing data and compute resources. The highly parallel neural networks portion of these models can benefit from GPU acceleration however, large embedding tables often cannot fit in the limited-capacity GPU device memory. Hence, this paper deep dives into the semantics of training data and obtains insights about the feature access, transfer, and usage patterns of these models. We observe that, due to the popularity of certain inputs, the accesses to the embeddings are highly skewed with a few embedding entries being accessed up to 10000X more. This paper leverages this asymmetrical access pattern to offer a framework, called FAE, and proposes a hot-embedding aware data layout for training recommender models. This layout utilizes the scarce GPU memory for storing the highly accessed embeddings, thus reduces the data transfers from CPU to GPU. At the same time, FAE engages the GPU to accelerate the executions of these hot embedding entries. Experiments on production-scale recommendation models with real datasets show that FAE reduces the overall training time by 2.3X and 1.52X in comparison to XDL CPU-only and XDL CPU-GPU execution while maintaining baseline accuracy.

Presented and published at the Proceedings of International Conference on Manufacturing ICM2002, August 09-11,2002, Dhaka,Bangladesh - Department of Industrial and Production Engineering, BUET. The presented work is a development of a simple method tointerface servo controlled multi axis robot model and is termed as ARDIC (Advanced Robot Direct Interfacing Controller)It is a program in a common language for communicatingwith any robot driven by 24 V DC servo motors by building adatabase and perform the required operations . The program can teach robot movements and store them up in batch files which can be called at a later stage .The interaction between the user and the robot is maintained by a high level of PC interfacing. There is no time lag in governing the robot arm axis. Sending and receiving signals is done through PC ports . Motion commands can be given with the arrow keys or by cursor . The software runs on Dos, Unix and Windows operating system. A compatible hardware is developed to the tune of the working conditions of the robot under test (Scorbot, Eshed Robotec). The working of the robot is satisfactory.

Paper discusses a unique robot interface model that facilitates direct communication with any intelligent or learning system of artificial intelligence. Learning aspects improves efficiencies where the interface allows learning of movements for optimized routes and replication of action performed previously by selecting the shortest paths.

Paper discusses a unique robot interface model that facilitates direct communication with any intelligent or learning system of artificial intelligence. Learning aspects improves efficiencies where the interface allows learning of movements for optimized routes and replication of action performed previously by selecting the shortest paths.

Paper discusses a unique robot interface model that facilitates direct communication with any intelligent or learning system of artificial intelligence. Learning aspects improves efficiencies where the interface allows learning of movements for optimized routes and replication of action performed previously by selecting the shortest paths.

Presented and published at the Proceedings of International Conference on Manufacturing ICM2002, August 09-11,2002, Dhaka,Bangladesh - Department of Industrial and Production Engineering, BUET. The presented work is a development of a simple method tointerface servo controlled multi axis robot model and is termed as ARDIC (Advanced Robot Direct Interfacing Controller)It is a program in a common language for communicatingwith any robot driven by 24 V DC servo motors by building adatabase and perform the required operations . The program can teach robot movements and store them up in batch files which can be called at a later stage .The interaction between the user and the robot is maintained by a high level of PC interfacing. There is no time lag in governing the robot arm axis. Sending and receiving signals is done through PC ports . Motion commands can be given with the arrow keys or by cursor . The software runs on Dos, Unix and Windows operating system. A compatible hardware is developed to the tune of the working conditions of the robot under test (Scorbot, Eshed Robotec). The working of the robot is satisfactory.

Adrenocortical carcinoma is a rare entity and usually diagnosed at later stages which poses a treatment dilemma and usually results in a bad prognosis. It is the second most aggressive endocrine malignancy after anaplastic thyroid carcinoma. Here we report a case in a young male diagnosed incidentally on an ultrasound scan done for abdominal pain. On examination his abdomen was soft with no obvious mass palpable. We are reporting a brief summary of the diagnosis, work up and management of this rare entity in the hope that it may shed some more light on the appropriate management of this elusive condition.

Compression is seen as a simple technique to increase the effective cache capacity. Unfortunately, compression techniques either incur tag area overheads or restrict data placement to only include neighboring compressed cache blocks to mitigate tag area overheads. Ideally, we should be able to place arbitrary compressed cache blocks without any placement restrictions and tag area overheads. This paper proposes Touch��, a framework that enables storing multiple arbitrary compressed cache blocks within a physical cacheline without any tag area overheads. The Touch�� framework consists of three components. The first component, called the ``Signature'' (SIGN) engine, creates shortened signatures from the tag addresses of compressed blocks. Due to this, the SIGN engine can store multiple signatures in each tag entry. On a cache access, the physical cacheline is accessed only if there is a signature match (which has a negligible probability of false positive). The second component, called the ``Tag Appended Data'' (TADA) mechanism, stores the full tag addresses with data. TADA enables Touch�� to detect false positive signature matches by ensuring that the actual tag address is available for comparison. The third component, called the ``Superblock Marker'' (SMARK) mechanism, uses a unique marker in the tag entry to indicate the occurrence of compressed cache blocks from neighboring physical addresses in the same cacheline. Touch�� is completely hardware-based and achieves an average speedup of 12\% (ideal 13\%) when compared to an uncompressed baseline., Keywords: Compression, Caches, Tag Array, Data Array, Hashing

High capacity and scalable memory systems play a vital role in enabling our desktops, smartphones, and pervasive technologies like Internet of Things (IoT). Unfortunately, memory systems are becoming increasingly prone to faults. This is because we rely on technology scaling to improve memory density, and at small feature sizes, memory cells tend to break easily. Today, memory reliability is seen as the key impediment towards using high-density devices, adopting new technologies, and even building the next Exascale supercomputer. To ensure even a bare-minimum level of reliability, present-day solutions tend to have high performance, power and area overheads. Ideally, we would like memory systems to remain robust, scalable, and implementable while keeping the overheads to a minimum. This dissertation describes how simple cross-layer architectural techniques can provide orders of magnitude higher reliability and enable seamless scalability for memory systems while incurring negligible overheads., PhD thesis, Georgia Institute of Technology (May 2017)

In late November 2011, the National Aeronautics and Space Administration (NASA) plans to launch its robotic explorer to scour Mars for signs of the planet’s ability to support life. The Mars Science Laboratory (MSL) spacecraft is scheduled to lift off from Cape Canaveral Air Force Station in Florida, shuttling Curiosity, an SUV-sized rover with a hefty scientific payload, to the red planet’s surface. John Grotzinger, a member of the National Academy of Sciences and professor of geology at the California Institute of Technology, helps oversee the mission. He became involved in the quest after studying how changes in the Earth’s environment helped influence animal diversity in some parts of our planet. Here, Grotzinger discusses the MSL with PNAS.

This thesis examines the role of diagnosis—traditional and molecular—in three areas of medicine: personalized cancer treatment, treatment of infectious diseases and treatment of controversial disorders lacking unambiguous physiological bases. The thesis uses a mix of statistics, expert interviews and patient anecdotes to address in the form of three feature stories three aspects pertinent to the role of diagnosis in modern medicine. The first story addresses the challenges to developing diagnostic markers for truly personalized cancer therapy. The second story features a recent advance in molecular diagnostics that has transformed the treatment of infectious diseases, especially hitherto-unknown viral infections. The third story illustrates the plight of patients suffering from disorders whose very existence is controversial and for which doctors are unable to provide clear-cut diagnoses.

“…Confusion appears to occur just after the articulation of a major conceptual advance that served to greatly clarify a problem of exceptional importance.”- Ira Mellman, 1996. What could be more fitting than the domain of protein trafficking to elucidate the above statement made by one of the several pioneers in the field? Ever since the pioneering groundwork laid down by Blobel and colleagues, emphasising protein translocation across intracellular membranes, the field of protein trafficking has been a playground of debates, dogma-reversals and rediscoveries. The possession of a valid cellular address tag is the basic requirement for the delivery of a given protein at its intracellular destination. However, the complexity involved in the foray of proteins from their site of synthesis to their site of function is within the scope of no comprehensive treatise. In this thesis, the work done on two individual transport steps of two different proteins has been summarised. In the first part of this thesis, the trafficking of the cation-dependent (CD-) mannose 6- phosphate receptor (MPR) has been studied. The CD-MPR cycles between the TGN and the plasma membrane, through the early and late endosomal compartments. It performs the important function of transport of lysosomal enzymes to lysosomes, a process which ensures the correct biogenesis of lysosomes. However, it is important that the receptor itself be excluded from lysosomes and safely retrieved to the TGN from late endosomes in order to avoid degradation in lysosomes. This is essential to ensure that the CD-MPR is available to support several rounds of lysosomal enzyme transport. This retrieval step has been shown to depend on a pair of aromatic residues F18W19 in the cytoplasmic tail of the receptor. Mutation of the residues to alanines has been shown to result in massive mislocalisation of the CD-MPR in lysosomes, the W19 residue being more crucial to this function and the F18 residue playing a contributory role. The retrieval has also been shown to take place in a Rab9 dependent manner using the cytosolic adaptor protein TIP47 (Tail Interacting Protein of 47 kDa). TIP47 specifically interacts with the diaromatic motif to effect this transport step. In this study, we demonstrated a strict requirement for di-aromaticity at the positions 18 and 19 of the cytosolic tail of the CDMPR both for correct intracellular sorting in vivo and optimal TIP47 interaction in vitro, thus demonstrating the significance of the di-aromatic motif in endosomal sorting and establishing the highly specific nature of this interaction. This also established a paradigm for the CD-MPR as a representative member of a generic family of diaromatic motif containing proteins. The second part of this thesis deals with the trafficking of the human mannose 6- phosphate uncovering enzyme (UCE). The recognition of the mannose 6-phosphate tag on lysosomal enzymes by the MPRs is facilitated by UCE which exposes the recognition signal on the lysosomal enzymes in a two-step enzymatic reaction: the first starts in the cis-Golgi and is mediated by a phosphotransferase and the second, mediated by UCE, occurs in the TGN. At steady state, UCE is mostly localised to the TGN and it cycles between the TGN and the plasma membrane. It is rapidly internalised from the surface in a clathrin dependent endocytic pathway and the internalisation has been shown to be mediated by a critical tyrosine-488 residue in its cytoplasmic tail. The transmembrane domain and first 11 residues of the cytoplasmic tail of UCE have been shown to be involved in its TGN retention. In this study, we identified the residues involved in TGN exit of UCE using a combination of biochemical and confocal immunofluorescence methods. Using a high dimensional neural network capable of identifying differences between images not visible to the eye, we determined that the residues 492QEMN were involved in TGN exit of UCE. The same method was also used to analyse the individual contribution of each amino acid in the sequence and it was found that residue Q492 is the most important to the exit function while residues M494 and N495 also contribute. The identification of a trans-Golgi network exit signal in its cytoplasmic tail elucidates the trafficking pathway of uncovering enzyme, a crucial player in lysosomal biogenesis. With these two analyses, we contributed to a better understanding of signal sequences involved in intracellular protein trafficking of two related proteins both involved in lysosomal biogenesis.