@dmvaldman I still think that it had some explanations of this joke in its corpus, but at the very least it's able to extend basic variations of logic pretty well.
Excited to present our work "Squeezeformer: An Efficient Transformer for Automatic Speech Recognition" at NeurIPS 2022 in New Orleans tomorrow Thursday December 1st!
Stop by our poster at 11AM - 1 PM (CST) in Hall J Poster #620
https://t.co/Ku9rgHN6KX
#NeurIPS2022
@quocleix@tanmingxing This is really great work! I was wondering about a detail. I was looking at the interesting lite-R ASPP network design and was curious about the pooling layer's striding. Why is it set to output a different aspect ratio instead of being 1:1 with the input? Thanks!