ICACT20240037 Question.2
Questioner: niuxiaoling@caict.ac.cn 2024-02-07 오전 10:06:55 |
ICACT20240037 Answer.2
Answer by Auhor lemon@etri.re.kr 2024-02-07 오전 10:06:55
Chrome  Click!! |
| Question 2. Can you explain the source of the dataset you used in your paper and the results of the experimental comparison? thank you. |
Thank you for your question. The source of the data set used in the paper is indicated in the paper, and can be downloaded and used for free as data that is publicly available on the Internet.
Comparing the experimental results of the MNIST and Fashion-MNIST datasets, we were able to reduce the number of wires while maintaining greater accuracy in MNIST. The reason is that MNIST has a slightly lower classification difficulty than Fashion-MNIST, so accuracy is maintained even when the number of synapses that can contain information is reduced. |
ICACT20240037 Question.3
Questioner: franklinlu888@outlook.com 2024-02-07 오전 10:34:40 |
ICACT20240037 Answer.3
Answer by Auhor lemon@etri.re.kr 2024-02-07 오전 10:34:40
Chrome Click!! |
| Good afternoon, I am the chair of Session 2B, Bing-Yuh Lu from China, and welcome you to ICACT2024.
Your artwork mentioned the reduction of the wiring in the digital circuit! That is good. However, the tradeoff between precision and cost is an art. What is your considerations of the tradeoff? Thank you!
|
Thank you for your question. If the number of synapses is greatly reduced in the network structure proposed in this paper, accuracy decreases, as shown in the experimental results. However, depending on the classification difficulty of the dataset used in SNN, the number of synapses can be reduced without loss of accuracy. And even if there is a slight decrease in accuracy compared to the network using FCL, using the proposed ACL can be sufficiently considered in terms of hardware implementation of SNN because it reduces the number of accuracy synapses. |
ICACT20240037 Question.4
Questioner: htzustc@mail.ustc.edu.cn 2024-02-07 오전 10:49:37 |
ICACT20240037 Answer.4
Answer by Auhor lemon@etri.re.kr 2024-02-07 오전 10:49:37
Chrome Click!! |
| Could you explain how to transform input data(e.g.images in MNIST dataset) into discrete-time spikes or pulses used in SNN? |
Thank you for your question.
In this paper, the rate-coding technique, which is widely used to convert image datasets into spike signals, was used to convert the MNIST dataset into a discrete-time spike signal, which is the input of the SNN.
Rate-coding is a coding technique that generates a number of spikes proportional to the intensity of the pixel, and the generated spikes follow a Poisson distribution in discrete time. |
