Discover and read the best of Twitter Threads about #CVPR
Most recents (3)
I am excited to announce that I have successfully defended my PhD and have published my PhD thesis on βLearning with Differentiable Algorithmsβ. π
arxiv.org/abs/2209.00616
In the thesis, I explore how we can make discrete structures like algorithms differentiable. [1/13]
arxiv.org/abs/2209.00616
In the thesis, I explore how we can make discrete structures like algorithms differentiable. [1/13]
By making algorithms differentiable, we can integrate them end-to-end into neural network machine learning architectures. For example, we can continuously relax sorting (github.com/Felix-Petersenβ¦) for learning to rank. [2/13]
π’ In our #ACMMM21 paper, we highlight issues with training and evaluation of π°πΏπΌππ± π°πΌππ»ππΆπ»π΄ deep networks. π§΅π
For far too long, π°πΏπΌππ± π°πΌππ»ππΆπ»π΄ works in #CVPR, #AAAI, #ICCV, #NeurIPS have reported only MAE, but not standard deviation.
Looking at MAE and standard deviation from MAE, a very grim picture emerges. E.g. Imagine a SOTA net with MAE 71.7 but deviation is a whopping 376.4 !
What's the difference between recent works in semi-supervised learning? Key: consistency training (UDA, FixMatch, ReMixMatch, ICT, etc.) and self-training (#NoisyStudent), see photo below and my recent talk (bit.ly/thangluong-tal⦠with bonus slides on #MeenaBot!).
Our UDA work (arxiv.org/abs/1904.12848) proposes the use of strong augmentation (RandAugment) which subsequent works (FixMatch, NoisyStudent) follow. UDA uses soft pseudo-labels whereas FixMatch uses hard ones after "weak" augmentation in consistency training.
It is also important to note that adding noise to the student, training equal-or-larger students, and iterative self-training is a novel combination that defines the success of #NoisyStudent on ImageNet (arxiv.org/abs/1911.04252, to appear in #CVPR).
