Patent
Activation function calculation unit for quantized LSTM
Zheng Yong,Chen Zhujia,Shu Yi +2 more
- 21 Apr 2020
TL;DR: In this paper, the activation function calculation unit for quantized LSTM is characterized in terms of a sigmoid function and a tanh function, and the activation functions are used to approximate an activation function.
read more
Abstract: The invention relates to an activation function calculation unit for quantized LSTM, and the activation function calculation unit is characterized in that the activation function calculation unit comprises a sigmoid function calculation unit and a tanh function calculation unit; the tanh function calculation unit comprises a first complement code solving unit, a second complement code solving unit, an amplitude comparator, an arithmetic operation unit, a first selector and a second selector. According to the activation function calculation unit for the quantized LSTM provided by the invention,after the operand for the LSTM network is quantized, the operand bit width of the activation function calculation unit is determined according to the network precision, the noise margin and the bit width requirement of the offset coefficient in the linear function for the operand. For the operands with the bit width of 8 bits, a linear function expression used for approximating an activation function is determined. A corresponding calculation unit is designed for the linear function expression. A booth algorithm is used, and simple shift addition is used to quickly calculate the multiplication with a fixed coefficient.
read more
Chat with Paper
AI Agents for this Paper
Find similar papers on Google Scholar, PubMed and Arxiv
Write a critical review of this paper
Analyze citations of this paper to find unaddressed research gaps
References
Patent
Multi-computing-unit coarse-grained reconfigurable system and method for recurrent neural network
Wang Chen,Xu Xinyan +1 more
- 31 May 2017
TL;DR: In this paper, a multi-computing unit coarse-grained reconfigurable system and method for a recurrent neural network LSTM (long short-term memory) is presented, which comprises multi-matrix product bias and calculation arrays, activation fitting calculation arrays and vector quantity calculation arrays.
13
Patent
Design method of hardware accelerator based on LSTM recursive neural network algorithm on FPGA platform
Li Xi,Xuehai Zhou,Chao Wang,Chen Xianglan +3 more
- 29 May 2018
TL;DR: In this article, a method for accelerating an LSTM neural network algorithm on an FPGA platform is presented. But the method comprises the following steps: a Tensorflow pair is constructed, and parameters of the neural network are trained; the parameters of LSTMs are compressed by adopting a compression means, and the problem that storage resources of the FPGAs are insufficient is solved; according to the prediction process of the compressed LstM network, a calculation part suitable for running on the field-programmable gate array platform is determined; accordingto the determined calculation
9
Low Complex & High Accuracy Computation Approximations to Enable On-Device RNN Applications
Sirish Kumar Pasupuleti,Raj Narayana Gadde,Vasanthakumar Rajagopal,Ashok Vishnoi,N Chandra Sekhar,R Chandra Kumar,Narasinga Rao Miniskar +6 more
- 26 May 2019
TL;DR: The proposed approximations eliminate multiplications, divisions and exponential operations by replacing them with simple arithmetic operations to significantly reduce the computation requirements without any perceivable loss of functional accuracy and reduce the memory size and bandwidth requirements.
7
•Posted Content
FINN-L: Library Extensions and Design Trade-off Analysis for Variable Precision LSTM Networks on FPGAs
Vladimir Rybalkin,Alessandro Pappalardo,Muhammad Mohsin Ghaffar,Giulio Gambardella,Norbert Wehn,Michaela Blott +5 more
TL;DR: In this article, the authors present an FPGA-based accelerator for BiLSTM neural network designed for optical character recognition, along with numerous other experimental proof points for a Zynq UltraScale+ XCZU7EV MPSoC within the given design space.
3
Patent
A configurable neural network activation function implementation device
Che Deliang,Li Na +1 more
- 28 May 2019
TL;DR: In this article, a configurable neural network activation function implementation device is presented, which consists of a controller, a symbol judgment module, a range detection module, an address generator, a parameter register, a floating point multiplier and a phase operand arithmetic unit.
2