Difference between revisions of "Resource:Seminar"

From MobiNetS
Jump to: navigation, search
(wenliang updates seminars)
m
Line 13: Line 13:
|speaker=Kaiwen}}
|speaker=Kaiwen}}
{{Latest_seminar
{{Latest_seminar
|abstract = This paper proposes Mandheling, the first system that enables highly resource-efficient on-device training by orchestrating the mixed-precision training with on-chip Digital Signal Processing (DSP) offloading. Mandheling fully explores the advantages of DSP in integer-based numerical calculation by four novel techniques: (1) a CPU-DSP co-scheduling scheme to mitigate the overhead from DSP-unfriendly operators; (2) a self-adaptive rescaling algorithm to reduce the overhead of dynamic rescaling in backward propagation; (3) a batch-splitting algorithm to improve the DSP cache efficiency; (4) a DSP-compute subgraph reusing mechanism to eliminate the preparation overhead on DSP. We have fully implemented Mandheling and demonstrated its effectiveness through extensive experiments. The results show that, compared to the state-of-the-art DNN engines from
|abstract = This paper proposes Mandheling, the first system that enables highly resource-efficient on-device training by orchestrating the mixed-precision training with on-chip Digital Signal Processing (DSP) offloading. Mandheling fully explores the advantages of DSP in integer-based numerical calculation by four novel techniques: (1) a CPU-DSP co-scheduling scheme to mitigate the overhead from DSP-unfriendly operators; (2) a self-adaptive rescaling algorithm to reduce the overhead of dynamic rescaling in backward propagation; (3) a batch-splitting algorithm to improve the DSP cache efficiency; (4) a DSP-compute subgraph reusing mechanism to eliminate the preparation overhead on DSP. We have fully implemented Mandheling and demonstrated its effectiveness through extensive experiments. The results show that, compared to the state-of-the-art DNN engines from TFLite and MNN, Mandheling reduces the per-batch training time by 5.5× and the energy consumption by 8.9× on average. In end-to-end training tasks, Mandheling reduces up to 10.7× convergence time and 13.1× energy consumption, with only 1.9%–2.7% accuracy loss compared to the FP32 precision setting.
TFLite and MNN, Mandheling reduces the per-batch training time by 5.5× and the energy consumption by 8.9× on average. In end-to-end training tasks, Mandheling reduces up to 10.7× convergence time and 13.1× energy consumption, with only 1.9%–2.7% accuracy loss compared to the FP32 precision setting.
|confname=Mobicom 2022
|confname=Mobicom 2022
|link=https://arxiv.org/pdf/2206.07509.pdf
|link=https://arxiv.org/pdf/2206.07509.pdf

Revision as of 22:13, 13 October 2022

Time: 2022-10-18 16:30
Address: 4th Research Building A527-B
Useful links: Readling list; Schedules; Previous seminars.

Latest

  1. [ICNP 2022] CONST: Exploiting Spatial-Temporal Correlation for Multi-Gateway based Reliable LoRa Reception, Kaiwen
    Abstract: As a representative technology of low power wide area network, LoRa has been widely adopted to many appli-cations. A fundamental question in LoRa is how to improve its reception quality in ultra-low SNR scenarios. Different from existing studies that exploit either spatial or temporal correlation for LoRa reception recovery, this paper jointly leverages the fine-grained spatial-temporal correlation among multiple gateways. We exploit the spatial and temporal correlation in LoRa packets to jointly process received signals so that the fine-grained offsets including Central Frequency Offset (CFO), Sampling Time Offset (STO) and Sampling Frequency Offset (SFO) are well compensated, and signals from multiple gateways are combined coherently. Moreover, a deep learning based soft decoding scheme is developed to integrate the energy distribution of each symbol into the decoder to further enhance the coding gain in a LoRa packet. We evaluate our work with commodity LoRa devices (i.e., Semtech SX1278) and gateways (i.e., USRP-B210) in both indoor and outdoor environments. Extensive experiment results show that our work achieves 4.6dB higher signal-to-noise ratio (SNR) and 1.5× lower bit error rate (BER) compared with existing approaches.
  2. [Mobicom 2022] Mandheling: Mixed-Precision On-Device DNN Training with DSP Offloading, Wenjie
    Abstract: This paper proposes Mandheling, the first system that enables highly resource-efficient on-device training by orchestrating the mixed-precision training with on-chip Digital Signal Processing (DSP) offloading. Mandheling fully explores the advantages of DSP in integer-based numerical calculation by four novel techniques: (1) a CPU-DSP co-scheduling scheme to mitigate the overhead from DSP-unfriendly operators; (2) a self-adaptive rescaling algorithm to reduce the overhead of dynamic rescaling in backward propagation; (3) a batch-splitting algorithm to improve the DSP cache efficiency; (4) a DSP-compute subgraph reusing mechanism to eliminate the preparation overhead on DSP. We have fully implemented Mandheling and demonstrated its effectiveness through extensive experiments. The results show that, compared to the state-of-the-art DNN engines from TFLite and MNN, Mandheling reduces the per-batch training time by 5.5× and the energy consumption by 8.9× on average. In end-to-end training tasks, Mandheling reduces up to 10.7× convergence time and 13.1× energy consumption, with only 1.9%–2.7% accuracy loss compared to the FP32 precision setting.
  3. [TMC 2022] Imitation Learning Enabled Task Scheduling for Online Vehicular Edge Computing, Zhenguo
    Abstract: Vehicular edge computing (VEC) is a promising paradigm based on the Internet of vehicles to provide computing resources for end users and relieve heavy traffic burden for cellular networks. In this paper, we consider a VEC network with dynamic topologies, unstable connections and unpredictable movements. Vehicles inside can offload computation tasks to available neighboring VEC clusters formed by onboard resources, with the purpose of both minimizing system energy consumption and satisfying task latency constraints. For online task scheduling, existing researches either design heuristic algorithms or leverage machine learning, e.g., deep reinforcement learning (DRL). However, these algorithms are not efficient enough because of their low searching efficiency and slow convergence speeds for large-scale networks. Instead, we propose an imitation learning enabled online task scheduling algorithm with near-optimal performance from the initial stage. Specially, an expert can obtain the optimal scheduling policy by solving the formulated optimization problem with a few samples offline. For online learning, we train agent policies by following the expert’s demonstration with an acceptable performance gap in theory. Performance results show that our solution has a significant advantage with more than 50 percent improvement compared with the benchmark.


History

History

2024

2023

2022

2021

2020

  • [Topic] [ The path planning algorithm for multiple mobile edge servers in EdgeGO], Rong Cong, 2020-11-18

2019

2018

2017

Instructions

请使用Latest_seminar和Hist_seminar模板更新本页信息.

    • 修改时间和地点信息
    • 将当前latest seminar部分的code复制到这个页面中
    • 将{{Latest_seminar... 修改为 {{Hist_seminar...，并增加对应的日期信息|date=
    • 填入latest seminar各字段信息
    • link请务必不要留空，如果没有link则填本页地址 https://mobinets.org/index.php?title=Resource:Seminar
  • 格式说明
    • Latest_seminar:

{{Latest_seminar
|confname=
|link=
|title=
|speaker=
}}

    • Hist_seminar

{{Hist_seminar
|confname=
|link=
|title=
|speaker=
|date=
}}

Retrieved from "http://mobinets.cn/site/index.php?title=Resource:Seminar&oldid=2540"