Open kundtx opened 1 year ago
G10 Jinnan He: Nice try! It's really interesting that some adder-based modules can extract features like CNN. But I wonder is there any accuracy lost compared to CNN? If they can greatly simplify the computation, what is the trade-off between accuracy loss and complexity reduction? And how can we evaluate it?
@Caarinaaa G10 Jinnan He: Nice try! It's really interesting that some adder-based modules can extract features like CNN. But I wonder is there any accuracy lost compared to CNN? If they can greatly simplify the computation, what is the trade-off between accuracy loss and complexity reduction? And how can we evaluate it?
G28 Jiawei Gu: Good question,thank you. because the adder module replaces multiplication with addition, it can effectively reduce the amount of operations. However, this kind of addition and replacement still maintains the matching of image similarity features, so it still remains acceptable in some cases. Trade-offs are determined by factors such as specific tasks, model size or structure, which may require specific analysis of specific issues.
http://8.129.175.102/lfd2022fall-poster-session/28.html