Uncategorized – OU-Tulsa Lab of Image and Information Processing

Uncategorized

Kernelizing transformer

July 4, 2020 by samuel.cheng@ou.edu - No comments

Transformers are RNNs (paper ) and (video):

Remarks:

Kernalized transformer is not on par with the original transformer but can be much faster (x1000 times for some applications)
Kernalized transformer can be modelled as RNN and can help further speed up inference time.

Uncategorized

Object-Centric Learning with Slot Attention

July 2, 2020 by samuel.cheng@ou.edu - No comments

Video and paper.

tldr. but the routing mechanism here seems to be quite similar with the capsule one. The authors emphasize that a slot should learn not just one type of object. It seems that the main trick is to first route image feature to slots. Slots will then be train to fit not just one type of objects like capsule network.

Uncategorized

Set distribution networks

June 30, 2020 by samuel.cheng@ou.edu - No comments

A work (video, pdf) from Apple to generate set with energy based model.

Uncategorized

Big Self-Supervised Models are Strong Semi-Supervised Learners

June 21, 2020 by samuel.cheng@ou.edu - No comments

A nice video explanation of the paper .

The proposed learning paradigm:

Self-supervised pretraining
Supervised finetuning
Distillation: train a student to learn the output of the teacher rather than the true label.

It seems to have a rather counterintuitive conclusion. Labeled data do not always help. Or too many labeled data used during training does not help.

Uncategorized

Quaternion explanation

May 20, 2020 by samuel.cheng@ou.edu - No comments

A wonderful explanation of why $q p q^{-1}$ is rotation of $p$ , where $q$ is most easily interpreted with the complex-number like structure of $\cos(\theta/2) + \sin(\theta/2)(a i + b j + c k)$ with $a i + bj +ck$ normalized (i.e., $a^2+b^2+c^2=1$ ). Note that $q^{-1}$ is then just $latex \cos(\theta/2) - \sin(\theta/2)(a i + b j + c k)$ (since $(a i + b j + c k)^2=-a^2 - b^2 -c^2 = -1$ just like the imaginary $i$ for complex number).

Basically quarternion is 4D stereographic projection back to 3D space. When multiply by a quarternion $\cos(\alpha) + \sin(\alpha) {\bf v}$ from the left, points parallel to $\bf v$ will be translate along $\bf v$ and points penpendicular to $\bf v$ will rotate around $\bf v$ (following the right hand rule). Similarly, when multiplying from the right by the same quarternion, points parallel to $\bf v$ will be translated in the same direction, but points perpendicular to $\bf v$ will be rotated in the opposite direction (following left hand rule). So if we multiple $q$ and $q^{-1}$ from both left and right hand sides, the translation effect will be exactly cancelled, and the the rotation will be doubled. That is why the angle in $q$ should be halved.