FPGAs are a compelling substrate for supporting machine learning inference. Tools such as High-Level Synthesis and hls4ml can shorten the development cycle for deploying ML algorithms on FPGAs, but can struggle to han...
详细信息
ISBN:
(纸本)9798400713965
FPGAs are a compelling substrate for supporting machine learning inference. Tools such as High-Level Synthesis and hls4ml can shorten the development cycle for deploying ML algorithms on FPGAs, but can struggle to handle the large on-chip storage needed for many of these models. In particular the high BRAM usage found in many of these flows can cause Place & Route failures during synthesis. In this paper we propose using a Simulated-Annealing based flow to perform BRAM-aware quantization. This approach trades off inference accuracy with BRAM usage, to provide a high-quality inference engine that still meets on-chip resource constraints. We demonstrate this flow for Transformer-based machine learning algorithms, which include Flash Attention in a Stream-based Dataflow architecture. Our system imposes minimal accuracy drops, yet can reduce BRAM usage by 20%-50%, and improve power efficiency by 264%-812% compared to existing Transformer-based accelerators on FPGAs
The c-axis permittivity of 1T-TaS2 - a quasi-2D charge-density-wave material - changes upon illumination due to light-induced reorganization of CDW stacking. Here we probe the mechanism of this reorganization and find...
详细信息
We report a flexible temperature sensor based on Ti02 photonics that shows double the sensitivity compared to silicon photonics. This high sensitivity and biocompatibility pave the way towards point-of-care temperatur...
详细信息
Accurate weekly electricity load prediction is of utmost importance for electricity providers to ensure uninterrupted power supply to customers. This study applies an Artificial Neural Network (ANN) to achieve precise...
Accurate weekly electricity load prediction is of utmost importance for electricity providers to ensure uninterrupted power supply to customers. This study applies an Artificial Neural Network (ANN) to achieve precise weekly electricity load prediction. The dataset used for the ANN model consists of three months’ worth of data, including daily workload profiles, holiday work profiles, temperature, and humidity. For model training, 90% of the data is utilized with the Levenberg-Marquardt algorithm, while the remaining 10% is used for testing. The Mean Average Percentage Error (MAPE) is employed as the error metric. Based on the test results, the weekly load prediction error rate using ANN is determined to be 1.78% based on the MAPE value.
We demonstrate a dynamically tunable plasmonic metasurface enabled by light-tunable optical constants of a quantum material - 1T-TaS2. We observe a relative reflectance change of 10% under low-intensity incoherent ill...
详细信息
We report a flexible temperature sensor based on TiO2 photonics that shows double the sensitivity compared to silicon photonics. This high sensitivity and biocompatibility pave the way towards point-of-care temperatur...
详细信息
This article proposes an intelligent platform for monitoring students' steps on their way to school until they leave the school to their homes. This platform can identify students and notify those responsible and ...
详细信息
This paper proposes an analytical target modifi-cation for linear robust model predictive control strategies in order to deal with time-varying references defined by dynamic signal targets. The new approach can be dir...
详细信息
ISBN:
(数字)9798350382655
ISBN:
(纸本)9798350382662
This paper proposes an analytical target modifi-cation for linear robust model predictive control strategies in order to deal with time-varying references defined by dynamic signal targets. The new approach can be directly integrated to linear robust model predictive control algorithms that achieve piecewise constant reference tracking if recursive feasibility is ensured for any set-point. The main contribution is to present a direct analytical approach that provides a potentially improved steady-state tracking error performance with the same computation complexity of the original MPC for tracking piecewise constant reference. A simulation case study based on the trajectory tracking control of a quadrotor is used to illustrate the usefulness of the new analytical target modification layer.
In this paper, the overall health index of underground cable system is determined using Fuzzy Logic and Scoring and Weighting Average methods. The relevant data of 73 feeders has been collected in the prepared evaluat...
详细信息
Attendance systems have become more modern, and one of the biometric systems without physical contact is face recognition. However, many face-based attendance systems still carry out attendance individually and cannot...
详细信息
ISBN:
(数字)9798350376968
ISBN:
(纸本)9798350376975
Attendance systems have become more modern, and one of the biometric systems without physical contact is face recognition. However, many face-based attendance systems still carry out attendance individually and cannot detect multiple faces simultaneously. In addition, capturing facial data in real-time is still a challenge because the relatively large distance between the camera and the individual reduces the ability to recognize faces. The general solution is to use super-resolution to generate better-quality faces while maintaining the main facial recognition features. One technique still being researched is super-resolution generative adversarial networks (SRGAN). SRGAN can enlarge the resolution of captured images and maintain image quality sufficient for face recognition. The attendance system can be easily integrated into edge devices such as the Jetson Nano. This paper proposes automatic and effective attendance systems with the super-resolution technique to detect and recognize faces in low-resolution input. The experimental results show that using face data capture with a resolution of 40 × 40 pixels and a four-fold magnification results in a resolution of 160 × 160 pixels. Combining Face SRGAN with FaceNet architecture as the basis of face recognition can achieve an accuracy rate of 78.19% and an F1-Score of 81.13% with an average processing time of 1.61 seconds per frame on a PC and 14.55 seconds per frame on a Jetson Nano at an average of face recognition per frame of as many as up to 8 faces simultaneously.
暂无评论