Context-Aware Compilation of DNN Training Pipelines across Edge and Cloud

Empowered by machine learning, edge devices including smartphones, wearable, and IoT devices have become growingly intelligent, raising conflicts with the limited resource. On-device model personalization is particularly hard as training models on edge devices is highly resource-intensive. In this work, we propose a novel training pipeline across the edge and the cloud, by taking advantage of the powerful cloud while keeping data local at the edge. Highlights of the design incorporate the parallel execution enabled by our feature replay, reduced communication cost by our error-feedback feature compression, as well as the context-aware deployment decision engine. Working as an integrated system, the proposed pipeline training framework not only significantly speeds up training, but also incurs little accuracy loss or additional memory/energy overhead. We test our system in a variety of settings including WiFi, 5G, household IoT, and on different training tasks such as image/text classification, image generation, to demonstrate its advantage over the state-of-the-art. Experimental results show that our system not only adapts well to, but also draws on the varying contexts, delivering a practical and efficient solution to edge-cloud model training.

A Data-Centric Augmentation Approach for Disturbed Sensor Image Segmentation

Deep learning methods have become increasingly popular for optical sensor image analysis. They are adaptable to specific tasks and simultaneously demonstrate a high degree of generalization capability. However, applying deep neural networks to problems with low availability of labeled training data can lead to a model being incapable of generalizing to possible scenarios that may occur in test data, especially with the occurrence of dominant imaging artifacts. We propose a data-centric augmentation approach based on generative adversarial networks that overlays the existing labeled data with synthetic artifacts that are generated from data not present in the training set. This augmentation leads to a more robust generalization capability in semantic segmentation. Our method does not need any additional labeling and does not lead to additional memory or time consumption during inference. Further, we find it to be more effective than comparable approaches that are based on procedurally generated disturbances and the direct use of real disturbances. Building upon the improved segmentation results, we observe that our approach leads to improvements of 22% in the F1-score for an evaluated detection problem, which promises significant robustness towards future disturbances. In the context of sensor-based data analysis, the compensation of image artifacts is a challenge. When the structures of interest are not clearly visible in an image, algorithms that can cope with artifacts are crucial for obtaining the desired information. Thereby, the high variation of artifacts, the combination of different types of artifacts, and their similarity to signals of interest are specific issues that have to be considered in the analysis. Despite the high generalization capability of deep learning-based approaches, their recent success was driven by the availability of large amounts of labeled data. Therefore, the provision of comprehensive labeled image data with different characteristics of image artifacts is of importance. At the same time, applying deep neural networks to problems with low availability of labeled data remains a challenge. This work presents a data-centric augmentation approach based on generative adversarial networks that augments the existing labeled data with synthetic artifacts generated from data not present in the training set. In our experiments, this augmentation leads to a more robust generalization in segmentation. Our method does not need additional labeling and does not lead to additional memory or time consumption during inference. Further, we find it to be more effective than comparable augmentations based on procedurally generated artifacts and the direct use of real artifacts. Building upon the improved segmentation results, we observe that our approach leads to improvements of 22% in the F1-score for an evaluated detection problem. Having achieved these results with an example sensor, we expect increased robustness against artifacts in future applications.

Voltage-driven gigahertz frequency tuning of spin Hall nano-oscillators

Spin Hall nano-oscillators (SHNOs) exploiting current-driven magnetization auto-oscillation have recently received much attention because of their potential for oscillator-based neuromorphic computing. Widespread neuromorphic application with SHNOs requires an energy-efficient way to tune oscillation frequency in broad ranges and store trained frequencies in SHNOs without the need for additional memory circuitry. Voltage control of oscillation frequency of SHNOs was experimentally demonstrated, but the voltage-driven frequency tuning was volatile and limited to megahertz ranges. Here, we show that the frequency of SHNO is controlled up to 2.1 GHz by a moderate electric field of 1.25 MV/cm. The large frequency tuning is attributed to the voltage-controlled magnetic anisotropy (VCMA) in a perpendicularly magnetized Ta/Pt/[Co/Ni]n/Co/AlOx structure. Moreover, non-volatile VCMA effect enables control of the cumulative frequency using repetitive voltage pulses, which mimic the potentiation and depression functions of biological synapses. Our results suggest that the voltage-driven frequency tuning of SHNOs facilitates the development of energy-efficient spin-based neuromorphic devices.

Analysis and Comparison of Table-based Arithmetic to Boolean Masking

Masking is a popular technique to protect cryptographic implementations against side-channel attacks and comes in several variants including Boolean and arithmetic masking. Some masked implementations require conversion between these two variants, which is increasingly the case for masking of post-quantum encryption and signature schemes. One way to perform Arithmetic to Boolean (A2B) mask conversion is a table-based approach first introduced by Coron and Tchulkine, and later corrected and adapted by Debraize in CHES 2012. In this work, we show both analytically and experimentally that the table-based A2B conversion algorithm proposed by Debraize does not achieve the claimed resistance against differential power analysis due to a non-uniform masking of an intermediate variable. This non-uniformity is hard to find analytically but leads to clear leakage in experimental validation. To address the non-uniform masking issue, we propose two new A2B conversions: one that maintains efficiency at the cost of additional memory and one that trades efficiency for a reduced memory footprint. We give analytical and experimental evidence for their security, and will make their implementations, which are shown to be free from side-channel leakage in 100.000 power traces collected on the ARM Cortex-M4, available online. We conclude that when designing side-channel protection mechanisms, it is of paramount importance to perform both a theoretical analysis and an experimental validation of the method.

Q-Learning based Routing Protocol to Enhance Network Lifetime in WSNs

In resource constraint Wireless Sensor Networks (WSNs), enhancement of network lifetime has been one of the significantly challenging issues for the researchers. Researchers have been exploiting machine learning techniques, in particular reinforcement learning, to achieve efficient solutions in the domain of WSN. The objective of this paper is to apply Q-learning, a reinforcement learning technique, to enhance the lifetime of the network, by developing distributed routing protocols. Q-learning is an attractive choice for routing due to its low computational requirements and additional memory demands. To facilitate an agent running at each node to take an optimal action, the approach considers node’s residual energy, hop length to sink and transmission power. The parameters, residual energy and hop length, are used to calculate the Q-value, which in turn is used to decide the optimal next-hop for routing. The proposed protocols’ performance is evaluated through NS3 simulations, and compared with AODV protocol in terms of network lifetime, throughput and end-to-end delay.

DEEP REINFORCEMENT LEARNING WITH SPARSE DISTRIBUTED MEMORY FOR “WATER WORLD” PROBLEM SOLVING

Context. Machine learning is one of the actively developing areas of data processing. Reinforcement learning is a class of machine learning methods where the problem involves mapping the sequence of environmental states to agent’s actions. Significant progress in this area has been achieved using DQN-algorithms, which became one of the first classes of stable algorithms for learning using deep neural networks. The main disadvantage of this approach is the rapid growth of RAM in real-world tasks. The approach proposed in this paper can partially solve this problem. Objective. The aim is to develop a method of forming the structure and nature of access to the sparse distributed memory with increased information content to improve reinforcement learning without additional memory. Method. A method of forming the structure and modification of sparse distributed memory for storing previous transitions of the actor in the form of prototypes is proposed. The method allows increasing the informativeness of the stored data and, as a result, to improve the process of creating a model of the studied process by intensifying the learning of the deep neural network. Increasing the informativeness of the stored data is the result of this sequence of actions. First, we compare the new transition and the last saved transition. To perform this comparison, this method introduces a rate estimate for the distance between transitions. If the distance between the new transition and the last saved transition is smaller than the specified threshold, the new transition is written in place of the previous one without increasing the amount of memory. Otherwise, we create a new prototype in memory while deleting the prototype that has been stored in memory the longest. Results. The work of the proposed method was studied during the solution of the popular “Water World” test problem. The results showed a 1.5-times increase in the actor’s survival time in a hostile environment. This result was achieved by increasing the informativeness of the stored data without increasing the amount of RAM. Conclusions. The proposed method of forming and modifying the structure of sparse distributed memory allowed to increase the informativeness of the stored data. As a result of this approach, improved reinforcement learning parameters on the example of the “Water World” problem by increasing the accuracy of the model of the physical process represented by a deep neural network.

GenMC: A Model Checker for Weak Memory Models

GenMC is an LLVM-based state-of-the-art stateless model checker for concurrent C/C++ programs. Its modular infrastructure allows it to support complex memory models, such as RC11 and IMM, and makes it easy to extend to support further axiomatic memory models.In this paper, we discuss the overall architecture of the tool and how it can be extended to support additional memory models, programming languages, and/or synchronization primitives. To demonstrate the point, we have extended the tool with support for the Linux kernel memory model (LKMM), synchronization barriers, POSIX I/O system calls, and better error detection capabilities.

Question Answering Chatbot Using Memory Networks

Manipulation of a large amount of data becomes a very tedious task. Hence, the authors took the approach of memory networks for the implementation of the chatbot. Traditionally, the LSTM model was used to implement chatbots and QA systems. But the LSTM failed to store relevant information when given a longer information set. On the contrary, the memory networks have an additional memory component with it. This can help in storing long information for further use which is greatly advantageous for the QA and chatbot systems as compared to LSTM. The authors trained and tested their model over Facebook's bAbi dataset which consists of several tasks and has questions regarding each task to retrieve the accuracy of the model. On the pedestal of that dataset, they have presented the accuracy for every task in their study with memory networks.

A solution to the overdamping problem when simulating dust–gas mixtures with smoothed particle hydrodynamics

ABSTRACT We present a fix to the overdamping problem found by Laibe & Price when simulating strongly coupled dust–gas mixtures using two different sets of particles using smoothed particle hydrodynamics. Our solution is to compute the drag at the barycentre between gas and dust particle pairs when computing the drag force by reconstructing the velocity field, similar to the procedure in Godunov-type solvers. This fixes the overdamping problem at negligible computational cost, but with additional memory required to store velocity derivatives. We employ slope limiters to avoid spurious oscillations at shocks, finding the van Leer Monotonized Central limiter most effective.