INSight: Toward Practical Neuromorphic Computing
INsight is a neuromorphic computing system/architecture our lab is developing. The system aims at energy-efficient inference of practical, large-scale deep neural networks. Our goal is to evaluate AlexNet-scale neural networks(i.e., networks with several tens million parameters) in real-time within 1W power envelope using the conventional CMOS process for smartphones, wearables, Internet-of-Things(IoT) devices, etc.
Inspired by the brain
The neuromorphic chip we are developing consists of neurons and synapses. The neurons interconnected through synapses can perform visual/speech recognition, natural language processing, etc as our right brain does. Unlike existing custom accelerators for neural networks, a specific location of the chip is dedicated to a specific function, which makes our system more brain-like.
V1 in the brain and Layer1 in the chip actually play the same role: they detect edges from pixels.
The brain figure is from Nature Reviews Neuroscience 8, 379-391 (May 2007)
Zero memory access
CPUs, GPUs and many custom processors need off-chip memories to run. However, the seperation of processing elements and memories leads to processing bottlenecks and large energy consumption. Our neuromorphic chip does not require off-chip memories and stores all the parameters on chip. In our system, data just flow from the the input pins of the chip to the output pins continuously in the feed-forward fashion. This allows our system to scale and to achieve ultra-low energy consumption. Existing processors did not employ this architecture because the large number of parameters of practical neural networks requires several hundred megabytes storage/memory.
Compressing neural networks
INsight has a unique architecture owing to our software technology that simplifies a given neural network for the execution on neuromorphic architectures. The size of neural networks can be reduced up to 100X without noticeable loss of the performance.
The green dots shows the final performance of a neural network after compression. Even if the neural network is simplified highly, the performance is maintained.
Affordable
The prototypes of INsight system have been implemented by a very small team. The development cost is not several tens millions like other neuromorphic computing systems. This becomes possible because we fully leverage existing IC design methodologies. While we do not have fancy asynchronous, event-driven circuits, our results are competitive. Also, our system is compatible to existing technology such as traditional CMOS cameras.
Throughput and enery-efficiency
We are marching toward our goal: running AlexNet-scale neural networks in the neuromorphic fashion. Here are some results. We have achieved unparalleled throughput and unprecedented energy-efficiecy using off-the-shelf FPGAs, and these improvements will continue since the proposed system does not suffer from the Von Neumann bottleneck.
For more technical details, please refer to our arxiv paper.
Demos
Hand-written digit recognition is performed on INsight architecture. While we used a small, educational FPGA (Xilinx Artix7 100T),
we achieved 97.64% accuracy. The current version easily pass 99% on this FPGA.
CIFAR10 natural image classification is performed on INsight architecture. We achieved 78% accuracy on the FPGA (Xilinx Kintex7 325T).
We are finally running AlexNet in our neuromorphic computing system (Xilinx Virtex7 2000T).
Contacts
Prof. Jaeyong Chung (jychung at inu.ac.kr, http://esc.inu.ac.kr)
News
(2017/4) "Neuromorphic Computing System Finally Runs AlexNet!", Invited talk at KAIST
(2016/12) "Toward Neuromorphic Execution of Deep Learning Models", Invited talk at Inha University
(2016/11) "Toward Neuromorphic Execution of Deep Learning Models", Invited talk at Neuromorphic Hardware Lecture, IKEEE
(2016/9) "Toward Neuromorphic Execution of Deep Learning Models", Invited talk at RF/Analog Workshop, Korea
(2016/4) "Toward Neuromorphic Execution of Deep Learning Models", Invited talk at POSTECH
(2015/9) Xilinx Daily Blog introduced our project
(2015/8) DoD (Department of Defense) S&T news bulletin introduced our project
(2015/8) MIT Technology Review found our paper interesting
Publication
6. Jaeyong Chung, Taehwan Shin, and Joon-Sung Yang, "Simplifying Deep Neural Networks for Neuromorphic Architectures", IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD) (accepted)
5. Yongshin Kang, Joon-Sung Yang, and Jaeyong Chung, "Weight Partitioning for Dynamic Fixed-Point Neuromorphic Computing Systems", IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD) (accepted)
4. Yongshin Kang, Seban Kim, Taehwan Shin, and Jaeyong Chung, “Running Convolutional Layers of AlexNet in Neuromorphic Computing System”, Proc. IEEE Design Automation & Test in Europe Conference (DATE), March 2017 (Accepted, U-booth Demo)
3. Seban Kim and Jaeyong Chung, “Synthesis of Activation-Parallel Convolution Structures for Neuromorphic Architectures”, Proc. IEEE Design Automation & Test in Europe Conference (DATE), March 2017 (Accepted for Presentation) (Acceptance Rate: 193/794 = 24%)
2. Jaeyong Chung and Taehwan Shin, “Simplifying Deep Neural Networks for Neuromorphic Architectures”, Proc. IEEE/ACM Design Automation Conference (DAC), June 2016 (Acceptance Rate: 152/876 = 17%)
1. Taehwan Shin, Yongshin Kang, Seungho Yang, Seban Kim, and Jaeyong Chung, “Live Demonstration: Real-Time Image Classification on a Neuromorphic Computing System with Zero Off-Chip Memory Access”, Proc. IEEE International Symposium on Circuits and Systems (ISCAS), May 2016
Sponsors
