Accelerator Design for Deep Learning Training

Agrawal, Ankur; Chen, Chia-Yu; Choi, Jungwook; Gopalakrishnan, Kailash; Oh, Jinwook; Shukla, Sunil; Srinivasan, Vijayalakshmi; Venkataramani, Swagath; Zhang, Wei

doi:10.1145/3061639.3072944

Cited by 2 publications

(4 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Order By: Relevance

“…DNN components will always be part of an automotive system consisting of also conventional hardware and software components. Several researchers claim that that DNN components is a prerequisite for autonomous driving 28,2,79 . However, how to integrate such components in a system is an open question.…”

Section: Complementing Dnns With Conventional Componentsmentioning

confidence: 99%

Safely Entering the Deep: A Review of Verification and Validation for Machine Learning and a Challenge Elicitation in the Automotive Industry

Borg¹,

Englund²,

Wnuk³

et al. 2019

JASEN

View full text Add to dashboard Cite

Deep Neural Networks (DNN) will emerge as a cornerstone in automotive software engineering. However, developing systems with DNNs introduces novel challenges for safety assessments. This paper reviews the state-of-the-art in verification and validation of safety-critical systems that rely on machine learning. Furthermore, we report from a workshop series on DNNs for perception with automotive experts in Sweden, confirming that ISO 26262 largely contravenes the nature of DNNs. We recommend aerospace-to-automotive knowledge transfer and systems-based safety approaches, e.g., safety cage architectures and simulated system test cases.

show abstract

Section: Complementing Dnns With Conventional Componentsmentioning

confidence: 99%

Safely Entering the Deep: A Review of Verification and Validation for Machine Learning and a Challenge Elicitation in the Automotive Industry

Borg¹,

Englund²,

Wnuk³

et al. 2019

JASEN

View full text Add to dashboard Cite

show abstract

“…Some recent case studies for applying approximate computing to video processing [7], signal processing [8] and communication systems [9] have shown early feasibility. The research in the field of approximate computing has been led by seminal contributions from the industry players such as Intel [10], IBM [11,12], and Microsoft [13], as well as several research groups from academia [5,6,[14][15][16]. [6,17]).…”

Section: What Is Approximate Computing?mentioning

confidence: 99%

“…Companies are already resorting to approximate computing to obtain energy and cost optimized application-specific integrated circuits (ASICs), e.g., Google has made a custom ASIC named the Tensor Processing Unit to run machine learning-based tasks at scale in their data centers, while requiring fewer transistor per operation. Similarly, IBM exploited the error resilience of Deep Neural Networks to loss of numerical precision for better area and power efficiency systems [12]. We believe it is only a matter of time before these ideas would find way to network ASICs and that more research is needed on how these technologies will interplay with the network-layer stack.…”

Section: Algorithmsmentioning

confidence: 99%

“…Similarly, self organizing networks (SON) proposed for future cellular networks [70] use machine learning algorithms for various self organization phases. Cost, area and power efficient hardware at the cost of 100 percent accuracy are proposed for machine learning and deep learning applications [12]. The use cases for SONs using approximate hardware can reduce the cost of futuristic cellular systems.…”

Section: Case Study: Approximate 5g Network In Rural/low Income Areasmentioning

confidence: 99%

See 1 more Smart Citation

Approximate Networking for Universal Internet Access

et al. 2017

View full text Add to dashboard Cite

Despite the best efforts of networking researchers and practitioners, an ideal Internet experience is inaccessible to an overwhelming majority of people the world over, mainly due to the lack of cost-efficient ways of provisioning high-performance, global Internet. In this paper, we argue that instead of an exclusive focus on a utopian goal of universally accessible "ideal networking" (in which we have a high throughput and quality of service as well as low latency and congestion), we should consider providing "approximate networking" through the adoption of context-appropriate trade-offs. In this regard, we propose to leverage the advances in the emerging trend of "approximate computing" that rely on relaxing the bounds of precise/exact computing to provide new opportunities for improving the area, power, and performance efficiency of systems by orders of magnitude by embracing output errors in resilient applications. Furthermore, we propose to extend the dimensions of approximate computing towards various knobs available at network layers. Approximate networking can be used to provision "Global Access to the Internet for All" (GAIA) in a pragmatically tiered fashion, in which different users around the world are provided a different context-appropriate (but still contextually functional) Internet experience.

show abstract

Accelerator Design for Deep Learning Training

Cited by 2 publications

References 1 publication

Safely Entering the Deep: A Review of Verification and Validation for Machine Learning and a Challenge Elicitation in the Automotive Industry

Safely Entering the Deep: A Review of Verification and Validation for Machine Learning and a Challenge Elicitation in the Automotive Industry

Approximate Networking for Universal Internet Access

Contact Info

Product

Resources

About