On the Relationship between Self-Attention and Convolutional Layers
- Recent trends of incorporating attention mechanisms in vision have led researchers to reconsider the supremacy of convolutional layers as a primary building block. Beyond helping CNNs to handle long-range dependencies, Ramachandran et al. (2019) showed that attention can completely replace convolution and achieve state-of-the-art performance on vision tasks. This raises the question: do learned attention layers operate similarly to convolutional layers? This work provides evidence that attention layers can perform convolution and, indeed, they often learn to do so in practice. Specifically, we prove that a multi-head self-attention layer with sufficient number of heads is at least as powerful as any convolutional layer. Our numerical experiments then show that the phenomenon also occurs in practice, corroborating our analysis. Our code is publicly available.
- I’ve just started to think about how machines and humans could serve as different attention heads, which is why we concentrate into populations with shared features. Attention, given the right conditions, may be an emergent phenomena. Need to look at Kauffman.
- More Forward – done!
- Dedication – done
- Acknowledgements – started!
- Sometime between the end of the forward and meeting with Aaron, move over to the new template
7:00 – 4:30 ASRC PhD, BD, GOES
- Stampedes are a form of runaway attention, and precision/recall aid that process
- Starting on forward. Using the Arab Spring and GamerGate as the framing
- 11:00 VOLPE Meeting
- Pursuing the resilience proposal was well received. Next, go up and meet with the folks?
- Install card – done! Passed the smoke test
7:00 – 5:00 ASRC PhD, GOES
NLP highlights podcast
- Fix H3a-c – look at the heatmaps to see if there is some way of showing cell visitation as trustworthy, low border cells as safe, and stampede conditions as untrustworthy. Otherwise, use DTW
- Helpful information on Excel Histograms
Nomad, flocking, and stampeding heatmaps
- A border/core ratio explains this nicely. when border dwell time (BDT) > 1, dangerous stampede. When BDT = 1, then nomads, When BDT < 1, flocking.
- Updated the simulation results section. Now I need to update the conclusion hypothesis. – done!
Got my graphics card!
7:00- 4:00 ASRC PhD, GOES
- Finishing discussion – done
- Rolling in TACJ from introduction – done
- Adding conclusions – done
- Fix H3a-c
- Reimbursement for fall – done
- Mission Drive meeting (need to get time for dissertation and GSAW prep)
7:00 – 11:30, 3:00 – 5:00 PhD
- More slides. I think I’m going to try saving a snapshot of the PDF that I can highlight and annotate.
- That works, though every time I want to make an edit, I go back to the source material and forget to use the other pdf.
- Also, saving out the PDF using Acrobat really shrinks the file size, 50MB down to 2.7MB
- Finished Motivation and Introduction. Working on Background
- Nice bike ride to start the year off