Vision Transformers (ViTs) have shown competitive accuracy in image classification tasks compared with CNNs. Yet, they generally require much more data for model pre-training. Most of recent works thus are dedicated to designing more complex architec
tures or training methods to address the data-efficiency issue of ViTs. However, few of them explore improving the self-attention mechanism, a key factor distinguishing ViTs from CNNs. Different from existing works, we introduce a conceptually simple scheme, called refiner, to directly refine the self-attention maps of ViTs. Specifically, refiner explores attention expansion that projects the multi-head attention maps to a higher-dimensional space to promote their diversity. Further, refiner applies convolutions to augment local patterns of the attention maps, which we show is equivalent to a distributed local attention features are aggregated locally with learnable kernels and then globally aggregated with self-attention. Extensive experiments demonstrate that refiner works surprisingly well. Significantly, it enables ViTs to achieve 86% top-1 classification accuracy on ImageNet with only 81M parameters.
Continual learning tackles the setting of learning different tasks sequentially. Despite the lots of previous solutions, most of them still suffer significant forgetting or expensive memory cost. In this work, targeted at these problems, we first stu
dy the continual learning process through the lens of information theory and observe that forgetting of a model stems from the loss of emph{information gain} on its parameters from the previous tasks when learning a new task. From this viewpoint, we then propose a novel continual learning approach called Bit-Level Information Preserving (BLIP) that preserves the information gain on model parameters through updating the parameters at the bit level, which can be conveniently implemented with parameter quantization. More specifically, BLIP first trains a neural network with weight quantization on the new incoming task and then estimates information gain on each parameter provided by the task data to determine the bits to be frozen to prevent forgetting. We conduct extensive experiments ranging from classification tasks to reinforcement learning tasks, and the results show that our method produces better or on par results comparing to previous state-of-the-arts. Indeed, BLIP achieves close to zero forgetting while only requiring constant memory overheads throughout continual learning.
The vulnerability to slight input perturbations is a worrying yet intriguing property of deep neural networks (DNNs). Despite many previous works studying the reason behind such adversarial behavior, the relationship between the generalization perfor
mance and adversarial behavior of DNNs is still little understood. In this work, we reveal such relation by introducing a metric characterizing the generalization performance of a DNN. The metric can be disentangled into an information-theoretic non-robust component, responsible for adversarial behavior, and a robust component. Then, we show by experiments that current DNNs rely heavily on optimizing the non-robust component in achieving decent performance. We also demonstrate that current state-of-the-art adversarial training algorithms indeed try to robustify the DNNs by preventing them from using the non-robust component to distinguish samples from different categories. Also, based on our findings, we take a step forward and point out the possible direction for achieving decent standard performance and adversarial robustness simultaneously. We believe that our theory could further inspire the community to make more interesting discoveries about the relationship between standard generalization and adversarial generalization of deep learning models.
In this work, we propose a novel technique to boost training efficiency of a neural network. Our work is based on an excellent idea that whitening the inputs of neural networks can achieve a fast convergence speed. Given the well-known fact that inde
pendent components must be whitened, we introduce a novel Independent-Component (IC) layer before each weight layer, whose inputs would be made more independent. However, determining independent components is a computationally intensive task. To overcome this challenge, we propose to implement an IC layer by combining two popular techniques, Batch Normalization and Dropout, in a new manner that we can rigorously prove that Dropout can quadratically reduce the mutual information and linearly reduce the correlation between any pair of neurons with respect to the dropout layer parameter $p$. As demonstrated experimentally, the IC layer consistently outperforms the baseline approaches with more stable training process, faster convergence speed and better convergence limit on CIFAR10/100 and ILSVRC2012 datasets. The implementation of our IC layer makes us rethink the common practices in the design of neural networks. For example, we should not place Batch Normalization before ReLU since the non-negative responses of ReLU will make the weight layer updated in a suboptimal way, and we can achieve better performance by combining Batch Normalization and Dropout together as an IC layer.
