arxiv: https://arxiv.org/abs/2103.03206 key points use learnable latent array with a fixed size as query in transformer module in perceiver architecture raw inputs can be long since it will be use as key and value in transformer module do cross attention to fusion input data into latent array. this allows perceiver to Read more…
arxiv: https://arxiv.org/abs/2008.13367 key points another anchor-free point based object detection network introduce new loss, varifocal loss which is a forked version from focal loss. Makes some changes from focal loss to compensate positive/negative imbalance futher. instead of prediction classification and IOU score separately, this work predicts a single scalar which Read more…
arxiv: https://arxiv.org/pdf/1703.06868.pdf key points arbitrary style transfer in real time use adaptive instance normalization(AdaIN) layers which aligns the mean and variance of content features allows to control content-style trade-off, style interpolation, color/spatial control previous works optimization approach using backprop of network to minimize style loss and content loss. This can be Read more…
https://towardsdatascience.com/advances-in-few-shot-learning-a-guided-tour-36bc10a68b77 great brief summary over matching networks, prototypical networks, model-agnostic meta-learning(MAML). well, the first two topics are well explained but MAML section needs a lot more thinking to understand. Also, I think MAML is closer to the topic of meta-learning rather than few shot learning. https://arxiv.org/pdf/2008.06365.pdf “An Overview of Deep Read more…
key point compared to resnet, the residual blocks are upgraded to have multiple “paths” or as the paper puts it “cardinality” which can be treated as another model architecture design hyper parameter. resnext architectures that have sufficient cardinality shows improved performance tldr: use improved residual blocks compared to resnet Different Read more…
good simple explanation and figure: https://www.kaggle.com/residentmario/cosine-annealed-warm-restart-learning-schedulers
arxiv: https://arxiv.org/pdf/1905.11946.pdf key point propose ‘compound scaling method’ which scales all width/depth/resolution together which is an efficient scaling method that can be applied to any existing structure introduce a new family of baseline structure called ‘EfficientNets’. The very smallest baseline structure was found by authors through NAS, and then the Read more…
https://arxiv.org/pdf/1611.08408.pdf key point use adversarial net to distinguish if input masks are from segmentation net or ground truth. this allows to add auxiliary loss, like below: where mce: multi class cross-entropy bce: binary cross entropy s(): segmentation network a(): adversarial network y: ground truth x: input s(x): segmentation network output Read more…
https://arxiv.org/abs/1912.07768 This work suggests that surrogate data need not be drawn from the original data distribution.This paper investigates the question of whether we can train a data-generating network that can produce synthetic data that effectively and efficiently teaches a target task to a learner propose new method to create synthetic Read more…