Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Design space exploration of hardware accelerated continual learning methods in convolutional neural networks

Perakis Emmanouil

Full record

---

Summary

Artificial Intelligence (AI) and Machine Learning (ML) have seen indisputable advancements over the years, spanning a large number of branches from medicine and industry related machinery to data analytics and Internet of Things. One way in which Machine Learning on the edge falters is to learn from new, never before seen data, without having access to the previous data. If left as it is, trying to learn new classes results in catastrophic forgetting. By training a classifier that is separated from the network's parameters the model can learn new tasks without forgetting previously learned ones, and do this at inference time. This is where Continual Learning, and more importantly to this thesis, Streaming Linear Discriminant Analysis comes into play. In this thesis, an accelerator for the previously mentioned method was fully implemented and downloaded on an Field Programmable Gate Array (FPGA) device and compared to other platforms such as modern CPUs and Graphical Processing Units (GPUs). This accelerator results in fixed point latency that is two orders of magnitude smaller than even GPUs and hundrends of times more energy efficient. The floating point latency speedup is a lot smaller but still comparable to modern devices, while retaining the energy efficiency.

Available Files

Services