draft diagram for our solution

[suryagutta]

According to Colorado, most people will read the paper; they will glance at the first figure in the paper. We need to come up with some explainer diagrams and share with the group for feedback.

Based on the discussion in the team meeting (03/13 8:00 am), first we need to come up with a rough sketch of the current architecture. No need to focus on graphic creation.

[tchken]

• Draft to be confirmed

[suryagutta]

Good illustrations in https://generallyintelligent.ai/understanding-self-supervised-contrastive-learning.html for contrastive learning frameworks including MoCo-v2.

[suryagutta]

Is it called "squeeze layer" for both? In one convolution layer, we are making it 3 from 2 channels.

[suryagutta]

Main visualizations for the paper: https://docs.google.com/presentation/d/17CQJykk605wl_nkwAuWmFnTjzKO689GbVS2H2yma9eQ/edit#slide=id.gc554c97402_0_19

[tchken]

called squeeze layer is because both are cross channel pooling (2 channels squeeze to 256 x256 x1, and expand to 256x256x3) and after that it's an output of 3 so a 3x3 also applied could called it expansion layer. im not sure if we need to explicit 3x3 in the diagram after 1x1 conv, or just call the conv layer as dimension filter.

Question to Colorado -- do we need to explicit the 3x3 as the output of the channel is expand to 3? Or, simplify call it as dimension filter as whole and the explain in the context of the paper?

[taeil]

Main visualizations for the paper: https://docs.google.com/presentation/d/17CQJykk605wl_nkwAuWmFnTjzKO689GbVS2H2yma9eQ/edit#slide=id.gc554c97402_0_19

can we make the input image multiple layers? 2 and 10?

[tchken]

yea that's a good idea, like showing the 2 vs 10 channels

[suryagutta]

Updated the graphics with channel information.