Stochastic spiking neural networks based on ferroelectric memory devices
(2021) FYSM30 20211Solid State Physics
Department of Physics
- Abstract
- In the last decade the world has seen a massive explosion of activity in the ma chine learning field. While the benefits of machine learning are ever growing and more impressive, the drawbacks of the technology are becoming clear and can hardly be ignored. The problem with a lot of machine learning setups is that they require massive amount of resources such as computational power, data storage, training time and they consume immense amount of electrical power in the process. Our work attempts to ad dress these issues. We used stochastic spiking neural networks (SNN) to recognise hand written digits. The SNNs have stochastic binary synapses which intend to simulate the behaviour of nanoscaled ferroelectric tunnel junction (FTJ) memory... (More)
- In the last decade the world has seen a massive explosion of activity in the ma chine learning field. While the benefits of machine learning are ever growing and more impressive, the drawbacks of the technology are becoming clear and can hardly be ignored. The problem with a lot of machine learning setups is that they require massive amount of resources such as computational power, data storage, training time and they consume immense amount of electrical power in the process. Our work attempts to ad dress these issues. We used stochastic spiking neural networks (SNN) to recognise hand written digits. The SNNs have stochastic binary synapses which intend to simulate the behaviour of nanoscaled ferroelectric tunnel junction (FTJ) memory devices. The work presented here could pave the way for neuromorphic circuits which could drastically improve the speed and power consumption of machine learning algorithms in the future. Using a simple unsupervised learning scheme on the MNIST data set we report 77 % accuracy similar to regular analog synapses. Furthermore, we explore additional parameters, such as network size, stochastic switching probability level and addition of additional network layers. We further provide comparison to analog network learning rate, some insight into the challenges with the stochastic networks faced and we show that a stochastic network could be realised in real devices using a ferroelectric device switching model. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9057773
- author
- Sulinskas, Karolis LU
- supervisor
-
- Mattias Borg LU
- Jonas Johansson LU
- organization
- course
- FYSM30 20211
- year
- 2021
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- SNN, FTJ, Stochastic SNN, Spiking neural network, Binary SNN, SNN simulation
- language
- English
- id
- 9057773
- date added to LUP
- 2021-06-23 11:09:00
- date last changed
- 2021-06-23 11:09:00
@misc{9057773, abstract = {{In the last decade the world has seen a massive explosion of activity in the ma chine learning field. While the benefits of machine learning are ever growing and more impressive, the drawbacks of the technology are becoming clear and can hardly be ignored. The problem with a lot of machine learning setups is that they require massive amount of resources such as computational power, data storage, training time and they consume immense amount of electrical power in the process. Our work attempts to ad dress these issues. We used stochastic spiking neural networks (SNN) to recognise hand written digits. The SNNs have stochastic binary synapses which intend to simulate the behaviour of nanoscaled ferroelectric tunnel junction (FTJ) memory devices. The work presented here could pave the way for neuromorphic circuits which could drastically improve the speed and power consumption of machine learning algorithms in the future. Using a simple unsupervised learning scheme on the MNIST data set we report 77 % accuracy similar to regular analog synapses. Furthermore, we explore additional parameters, such as network size, stochastic switching probability level and addition of additional network layers. We further provide comparison to analog network learning rate, some insight into the challenges with the stochastic networks faced and we show that a stochastic network could be realised in real devices using a ferroelectric device switching model.}}, author = {{Sulinskas, Karolis}}, language = {{eng}}, note = {{Student Paper}}, title = {{Stochastic spiking neural networks based on ferroelectric memory devices}}, year = {{2021}}, }