SNOMED CT is a prominent clinical terminology that is globally used by healthcare professionals to record patient data. As such, it is important that SNOMED CT be free from errors and inconsistencies so that they do not trickle down to applications that rely on it. Studies have shown that, owing to its large and complex nature as well as its regular revisions, errors and inconsistencies are inherent in SNOMED CT. However, auditing resources are limited, and manual auditing is cumbersome. This calls for computational quality assurance techniques to identify areas that have a high likelihood of modeling discrepancies. This study presents a visualization tool that is based on a context-based auditing methodology and can be used as a crowdsourcing application to allow SNOMED CT users from across the globe to comment on the modeling of a concept that is found by the underlying algorithm to be inconsistent. A manual auditing of a sample of 50 concept pairs using the visualization tool found 66% of the pairs to exhibit inconsistencies. Tools such as this and the feedback obtained from SNOMED CT users through collective collaboration can be used by the curators of SNOMED CT to further enhance the terminology.

The growing prevalence and computational demands of Artificial Intelligence (AI) workloads has led to widespread use of hardware accelerators in their execution. Scaling the performance of AI accelerators across generations is pivotal to their success in commercial deployments. The intrinsic error-resilient nature of AI workloads present a unique opportunity for performance/energy improvement through precision scaling. Motivated by the recent algorithmic advances in precision scaling for inference and training, we designed RaPiD1, a 4-core AI accelerator chip supporting a spectrum of precisions, namely, 16 and 8-bit floating-point and 4 and 2-bit fixed-point. The 36mm2RaPiD chip fabricated in 7nm EUV technology delivers a peak 3.5 TFLOPS/W in HFP8 mode and 16.5 TOPS/W in INT4 mode at nominal voltage. Using a performance model calibrated to within 1% of the measurement results, we evaluated DNN inference using 4-bit fixed-point representation for a 4-core 1 RaPiD chip system and DNN training using 8-bit floating point representation for a 768 TFLOPs AI system comprising 4 32-core RaPiD chips. Our results show INT4 inference for batch size of 1 achieves 3 - 13.5 (average 7) TOPS/W and FP8 training for a mini-batch of 512 achieves a sustained 102 - 588 (average 203) TFLOPS across a wide range of applications.

Assuring the safety of autonomous vehicles is one of the most significant challenges in the automotive industry. Tech companies and automotive manufacturers use the idea of Operational Design Domain (ODD) to indicate where their Automated Driving Systems (ADS) can operate safely. By definition from SAE J3016, an ODD defines where the ADS is designed to operate. However, it is loosely defined in no particular format, and it is unclear how exactly to formulate the ODD, which leaves it up to the ADS developer to determine. This paper proposes a methodology to identify an ODD for an ADS with statistical data and risk tolerance, where the identified ODD is constituted of a geographical map where the risk of ADS operation is lower than the pre-determined risk threshold for a given set of environmental conditions. Two different ADSs are run through this method as an example to showcase the methodology and link the identified ODD directly to the calculated performance of the ADSs. This systematically generated ODD can mitigate potential safety issues by informing the limitations of the ADS to safety drivers, through geographic and environmental boundaries.

Automated Driving System (ADS) safety assessment is a crucial step before deployment on public roads. Despite the importance of ADS safety assurance to test ADS reliability, most of the existing work is strongly attached to a single testing data source (i.e. on-road collected testing data, simulation or test track). Each source has different fidelity levels and capabilities, therefore there is a lack of a solution that allows for all data sources to complement each other to enable agnostic end to end evaluation and contributes towards different testing goals. Evaluation of ADSs is considered as a mandatory step in the autonomous vehicle development life cycle, demanding a reliable and comprehensive method is important. Here, we propose a source-agnostic framework, which can perform ADS evaluation compatible with different testing sources. Our findings show that this comprehensive solution can save time, effort and money consumed in ADS evaluation.

In this paper, we propose a number of novel techniques and numerical representation formats that enable, for the very first time, the precision of training systems to be aggressively scaled from 8-bits to 4-bits. To enable this advance, we explore a novel adaptive Gradient Scaling technique (Gradscale) that addresses the challenges of insufficient range and resolution in quantized gradients as well as explores the impact of quantization errors observed during model training. We theoretically analyze the role of bias in gradient quantization and propose solutions that mitigate the impact of this bias on model convergence. Finally, we examine our techniques on a spectrum of deep learning models in computer vision, speech, and NLP. In combination with previously proposed solutions for 4-bit quantization of weight and activation tensors, 4-bit training shows a non-significant loss in accuracy across application domains while enabling significant hardware acceleration (> 7X over state-of-the-art FP16 systems).

The resilience of Deep Learning (DL) training and inference workloads to low-precision computations, coupled with the demand for power-and area-efficient hardware accelerators for these workloads, has led to the emergence of 16-bit floating point formats as the precision of choice for DL hardware accelerators. This paper describes our optimized 16-bit format that has 6 exponent bits and 9 fraction bits, derived from a study of the range of values encountered in DL applications. We demonstrate that our format preserves the accuracy of DL networks, and we compare its ease-of-use for DL against IEEE-754 half-precision (5 exponent bits and 10 fraction bits) and bfloat16 (8 exponent bits and 7 fraction bits). Further, our format eliminated sub-normals and simplifies rounding modes and handling of corner cases. This streamlines floating-point unit logic and enables realization of a compact power-efficient computation engine.

Interoperability of health information has been declared as a national objective to be achieved using Electronic Health Records (EHRs). The HITECH Act of 2009 in the US helped facilitate the adoption of EHRs by authorizing incentive payments through Medicare and Medicaid to hospitals and clinicians for using EHRs to achieve improvements in health-care delivery. SNOMED CT, being a global standardized terminology and a requirement for the meaningful use of EHRs, will play an integral role in achieving widespread exchange of clinical information and decision support tools. However, quality issues with SNOMED CT may hinder the progress that we have seen with the adoption of EHRs at hospitals and clinics. Therefore, it becomes essential that efficient methods be formulated to assure the quality of SNOMED CT. This study presents a machine learning approach using a context-based, lexical method. The results show that such computational methods can effectively be used to identify problematic areas in SNOMED CT where the limited auditing resources could then be allocated.

Highly distributed training of Deep Neural Networks (DNNs) on future compute platforms (offering 100 of TeraOps/s of computational capacity) is expected to be severely communication constrained. To overcome this limitation, new gradient compression techniques are needed that are computationally friendly, applicable to a wide variety of layers seen in Deep Neural Networks and adaptable to variations in network architectures as well as their hyper-parameters. In this paper we introduce a novel technique - the Adaptive Residual Gradient Compression (AdaComp) scheme. AdaComp is based on localized selection of gradient residues and automatically tunes the compression rate depending on local activity. We show excellent results on a wide spectrum of state of the art Deep Learning models in multiple domains (vision, speech, language), datasets (MNIST, CIFAR10, ImageNet, BN50, Shakespeare), optimizers (SGD with momentum, Adam) and network parameters (number of learners, minibatch-size etc.). Exploiting both sparsity and quantization, we demonstrate end-to-end compression rates of ∼200× for fully-connected and recurrent layers, and ∼40× for convolutional layers, without any noticeable degradation in model accuracies.

The combination of growth in compute capabilities and availability of large datasets has led to a re-birth of deep learning. Deep Neural Networks (DNNs) have become state-of-the-art in a variety of machine learning tasks spanning domains across vision, speech, and machine translation. Deep Learning (DL) achieves high accuracy in these tasks at the expense of 100s of ExaOps of computation; posing significant challenges to efficient large-scale deployment in both resource-constrained environments and data centers.

SNOMED CT has continued to expand its adoption as a comprehensive clinical terminology. In 2013, the US government mandated SNOMED CT as a requirement for Stage 2 meaningful use criteria for Electronic Health Records. Studies have, however, identified inconsistencies in the content of SNOMED CT that may lessen its effectiveness when used for encoding patient data. Auditing thus becomes an integral part of SNOMED CT's evolution. However, manually auditing every concept in SNOMED CT can be a humongous and a tedious task. Algorithmic techniques are needed to identify concepts with inconsistencies. This study uses a lexical method to identify concepts with high likelihood of inconsistencies in their structural modeling. A sample of concepts derived by using this technique is manually audited. Over a third (36.7%) of the sample concepts are found to exhibit one or more inconsistencies in their structural modeling. Such technique can extend SNOMED International's own effort to maintain the quality of the content of SNOMED CT.

Deep Neural Networks (DNNs) have emerged as a powerful and versatile set of techniques showing successes on challenging artificial intelligence (AI) problems. Applications in domains such as image/video processing, autonomous cars, natural language processing, speech synthesis and recognition, genomics and many others have embraced deep learning as the foundation. DNNs achieve superior accuracy for these applications with high computational complexity using very large models which require 100s of MBs of data storage, exaops of computation and high bandwidth for data movement. In spite of these impressive advances, it still takes days to weeks to train state of the art Deep Networks on large datasets - which directly limits the pace of innovation and adoption. In this paper, we present a multi-pronged approach to address the challenges in meeting both the throughput and the energy efficiency goals for DNN training.

This pictorial illustrates how objects can illuminate people's experience of public space, an approach we call itinerant probes. Itinerant probes are not individualized, mediated artifacts, but instead dynamic events that en- liven people's personal and collective memories. Building on the probes literature and recent ecological perspectives, we describe three probes related to lighting that we investigated at three public sites. Our explorations with these probes highlight the intimate histories associated with public spaces.

SNOMED CT is important for clinical applications, such as Electronic Health Record (EHR) encoding. However, inconsistency in modeling its concepts may prevent SNOMED CT from providing proper support for clinical use. This study provides an effective methodology for locating inconsistently modeled SNOMED CT concepts. One can expect lexically similar concepts to be modeled similarly. Positional similarity sets, sets of lexically similar concepts having only one different word at the same position of their names, are introduced. Concepts in such sets have a higher likelihood of being unjustifiably inconsistently modeled. A technique to incorporate three structural indicators into the selected sets is provided to further improve the likelihood of finding inconsistently modeled concepts. An analysis of a sample of 50 such sets and for each of these three indicators is performed. The sample of positional similarity sets is found to have 18.6% inconsistent concepts. The use of structural indicators is shown to further improve the likelihood of finding inconsistently modeled concepts up to 41.6% with high statistical significance when compared to the previous sample of positional similarity sets. Positional similarity sets with different structural indicators are shown to help identify inconsistencies in concept modeling with high likelihood. Furthermore, such sets enable the comparison of concept modeling in the context of other lexically similar concepts, which enhances the effectiveness of corrections by auditors. Such quality assurance methods can be used to supplement IHTSDO's own efforts in order to improve the quality of SNOMED CT.

Training of large-scale deep neural networks is often constrained by the available computational resources. We study the effect of limited precision data representation and computation on neural network training. Within the context of low-precision fixed-point computations, we observe the rounding scheme to play a crucial role in determining the network’s behavior during training. Our results show that deep networks can be trained using only 16-bit wide fixed-point number representation when using stochastic rounding, and incur little to no degradation in the classification accuracy. We also demonstrate an energy-efficient hardware accelerator that implements low-precision fixed-point arithmetic with stochastic rounding

OBJECTIVE: Large and complex terminologies, such as Systematized Nomenclature of Medicine-Clinical Terms (SNOMED CT), are prone to errors and inconsistencies. Abstraction networks are compact summarizations of the content and structure of a terminology. Abstraction networks have been shown to support terminology quality assurance. In this paper, we introduce an abstraction network derivation methodology which can be applied to SNOMED CT target hierarchies whose classes are defined using only hierarchical relationships (ie, without attribute relationships) and similar description-logic-based terminologies. METHODS: We introduce the tribal abstraction network (TAN), based on the notion of a tribe-a subhierarchy rooted at a child of a hierarchy root, assuming only the existence of concepts with multiple parents. The TAN summarizes a hierarchy that does not have attribute relationships using sets of concepts, called tribal units that belong to exactly the same multiple tribes. Tribal units are further divided into refined tribal units which contain closely related concepts. A quality assurance methodology that utilizes TAN summarizations is introduced. RESULTS: A TAN is derived for the Observable entity hierarchy of SNOMED CT, summarizing its content. A TAN-based quality assurance review of the concepts of the hierarchy is performed, and erroneous concepts are shown to appear more frequently in large refined tribal units than in small refined tribal units. Furthermore, more erroneous concepts appear in large refined tribal units of more tribes than of fewer tribes. CONCLUSIONS: In this paper we introduce the TAN for summarizing SNOMED CT target hierarchies. A TAN was derived for the Observable entity hierarchy of SNOMED CT. A quality assurance methodology utilizing the TAN was introduced and demonstrated.