Date: Wednesday 11 March 2020
Time: 14:30 - 16:00
Location / Room: Berlioz
Chair:
Gilles Sassatelli, Université de Montpellier, FR
Co-Chair:
Rishad Shafik, University of Newcastle, GB
Recent advances in machine learning have pushed the boundaries of what is possible in self-adaptive and learning systems. This session pushes the state of art in runtime power and performance trade-offs for deep neural networks and self-optimizing embedded systems.
| Time | Label | Presentation Title Authors |
|---|---|---|
| 14:30 | 7.7.1 | ANYTIMENET: CONTROLLING TIME-QUALITY TRADEOFFS IN DEEP NEURAL NETWORK ARCHITECTURES Speaker: Jung-Eun Kim, Yale University, US Authors: Jung-Eun Kim1, Richard Bradford2 and Zhong Shao1 1Yale University, US; 2Collins Aerospace, US Abstract Deeper neural networks, especially those with extremely large numbers of internal parameters, impose a heavy computational burden in obtaining sufficiently high-quality results. These burdens are impeding the application of machine learning and related techniques to time-critical computing systems. To address this challenge, we are proposing an architectural approach for neural networks that adaptively trades off computation time and solution quality to achieve high-quality solutions with timeliness. We propose a novel and general framework,AnytimeNet, that gradually inserts additional layers, so users can expect monotonically increasing quality of solutions as more computation time is expended. The framework allows users to select on the fly when to retrieve a result during runtime. Extensive evaluation results on classification tasks demonstrate that our proposed architecture provides adaptive control of classification solution quality according to the available computation time. Download Paper (PDF; Only available from the DATE venue WiFi) |
| 15:00 | 7.7.2 | ANTIDOTE: ATTENTION-BASED DYNAMIC OPTIMIZATION FOR NEURAL NETWORK RUNTIME EFFICIENCY Speaker: Xiang Chen, George Mason University, US Authors: Fuxun Yu1, Chenchen Liu2, Di Wang3, Yanzhi Wang1 and Xiang Chen1 1George Mason University, US; 2University of Maryland, Baltimore County, US; 3Microsoft, US Abstract Convolutional Neural Networks (CNNs) achieved great cognitive performance at the expense of considerable computation load. To relieve the computation load, many optimization works are developed to reduce the model redundancy by identifying and removing insignificant model components, such as weight sparsity and filter pruning. However, these works only evaluate model components' static significance with internal parameter information, ignoring their dynamic interaction with external inputs. With per-input feature activation, the model component significance can dynamically change, and thus the static methods can only achieve sub-optimal results. Therefore, we propose a dynamic CNN optimization framework in this work. Based on the neural network attention mechanism, we propose a comprehensive dynamic optimization framework including (1) testing-phase channel and column feature map pruning, as well as (2) training-phase optimization by targeted dropout. Such a dynamic optimization framework has several benefits: (1) First, it can accurately identify and aggressively remove per-input feature redundancy with considering the model-input interaction; (2) Meanwhile, it can maximally remove the feature map redundancy in various dimensions thanks to the multi-dimension flexibility; (3) The training-testing co-optimization favors the dynamic pruning and helps maintain the model accuracy even with very high feature pruning ratio. Extensive experiments show that our method could bring 37.4%∼54.5% FLOPs reduction with negligible accuracy drop on various of test networks. Download Paper (PDF; Only available from the DATE venue WiFi) |
| 15:30 | 7.7.3 | USING LEARNING CLASSIFIER SYSTEMS FOR THE DSE OF ADAPTIVE EMBEDDED SYSTEMS Speaker: Fedor Smirnov, Friedrich-Alexander-Universität Erlangen-Nürnberg, DE Authors: Fedor Smirnov, Behnaz Pourmohseni and Jürgen Teich, Friedrich-Alexander-Universität Erlangen-Nürnberg, DE Abstract Modern embedded systems are not only becoming more and more complex but are also often exposed to dynamically changing run-time conditions such as resource availability or processing power requirements. This trend has led to the emergence of adaptive systems which are designed using novel approaches that combine a static off-line Design Space Exploration (DSE) with the consideration of the dynamic run-time behavior of the system under design. In contrast to a static design approach, which provides a single design solution as a compromise between the possible run-time situations, the off-line DSE of these so-called hybrid design approaches yields a set of configuration alternatives, so that at run time, it becomes possible to dynamically choose the option most suited for the current situation. However, most of these approaches still use optimizers which were primarily developed for static design. Consequently, modeling complex dynamic environments or run-time requirements is either not possible or comes at the cost of a significant computation overhead or results of poor quality. As a remedy, this paper introduces Learning Optimizer Constrained by ALtering conditions (LOCAL), a novel optimization framework for the DSE of adaptive embedded systems. Following the structure of Learning Classifier System (LCS) optimizers, the proposed framework optimizes a strategy, i.e., a set of conditionally applicable solutions for the problem at hand, instead of a set of independent solutions. We show how the proposed framework—which can be used for the optimization of any adaptive system—is used for the optimization of dynamically reconfigurable many-core systems and provide experimental evidence that the hereby obtained strategy offers superior embeddability compared to the solutions provided by a s.o.t.a. hybrid approach which uses an evolutionary algorithm. Download Paper (PDF; Only available from the DATE venue WiFi) |
| 16:00 | IP3-13, 760 | EFFICIENT TRAINING ON EDGE DEVICES USING ONLINE QUANTIZATION Speaker: Michael Ostertag, University of California, San Diego, US Authors: Michael Ostertag1, Sarah Al-Doweesh2 and Tajana Rosing1 1University of California, San Diego, US; 2King Abdulaziz City of Science and Technology, SA Abstract Sensor-specific calibration functions offer superior performance over global models and single-step calibration procedures but require prohibitive levels of sampling in the input feature space. Sensor self-calibration by gathering training data through collaborative calibration or self-analyzing predictive results allows these sensors to gather sufficient information. Resource-constrained edge devices are then stuck between high communication costs for transmitting training data to a centralized server and high memory requirements for storing data locally. We propose online dataset quantization that maximizes the diversity of input features, maintaining a representative set of data from a larger stream of training data points. We test the effectiveness of online dataset quantization on two real-world datasets: air quality calibration and power prediction modeling. Online Dataset Quantization outperforms reservoir sampling and performs equally to offline methods. Download Paper (PDF; Only available from the DATE venue WiFi) |
| 16:01 | IP3-14, 190 | MULTI-AGENT ACTOR-CRITIC METHOD FOR JOINT DUTY-CYCLE AND TRANSMISSION POWER CONTROL Speaker: Sota Sawaguchi, CEA-Leti, FR Authors: Sota Sawaguchi1, Jean-Frédéric Christmann2, Anca Molnos2, Carolynn Bernier2 and Suzanne Lesecq2 1CEA, FR; 2CEA-Leti, FR Abstract Energy-harvesting Internet of Things (EH-IoT) wireless networks have gained attention due to their infinite operation and maintenance-free property. However, maintaining energy neutral operation (ENO) of EH-IoT devices, such that the harvested and consumed energy are matched during a certain time period, is crucial. Guaranteeing this ENO condition and optimal power-performance trade-off under various workloads and transient wireless channel quality is particularly challenging. This paper proposes a multi-agent actor-critic method for modulating both the transmission duty-cycle and the transmitter output power based on the state-of-buffer (SoB) and the state-of-charge (SoC) information as a state. Thanks to these buffers, system uncertainties, especially harvested energy and wireless link conditions, are addressed effectively. Differently from the state-of-the-art, our solution does not require any model of the wireless transceiver nor any measurement of wireless channel quality. Simulation results of a solar powered EH-IoT node using real-life outdoor solar irradiance data show that the proposed method achieves better performance without system fails throughout a year compared to the state-of-the-art that suffers some system downtime. Our approach also predicts almost no system fails during five years of operation. This proves that our approach can adapt to the change in energy-harvesting and wireless channel quality, all without direct observations. Download Paper (PDF; Only available from the DATE venue WiFi) |
| 16:00 | End of session | |