Dreaming to Distill: Data-free Knowledge Transfer via DeepInversion

We introduce DeepInversion, a new method for synthesizing images from the image distribution used to train a deep neural network. We 'invert' a trained network (teacher) to synthesize class-conditional input images starting from random noise, without using any additional information about the training dataset. Keeping the teacher fixed, our method optimizes the input while regularizing the distribution of intermediate feature maps using information stored in the batch normalization layers of the teacher. Further, we improve the diversity of synthesized images using Adaptive DeepInversion, which maximizes the Jensen-Shannon divergence between the teacher and student network logits. The resulting synthesized images from networks trained on the CIFAR-10 and ImageNet datasets demonstrate high fidelity and degree of realism, and help enable a new breed of data-free applications - ones that do not require any real images or labeled data. We demonstrate the applicability of our proposed method to three tasks of immense practical importance -- (i) data-free network pruning, (ii) data-free knowledge transfer, and (iii) data-free continual learning.

Cost Volume Pyramid Based Depth Inference for Multi-View Stereo

We propose a cost volume based neural network for depth inference from multi-view images. We demonstrate that building a cost volume pyramid in a coarse-to-fine manner instead of constructing a cost volume at a fixed resolution leads to a compact, lightweight network and allows us inferring high resolution depth maps to achieve better reconstruction results. To this end, a cost volume based on uniform sampling of fronto-parallel planes across entire depth range is first built at the coarsest resolution of an image. Given current depth estimate, new cost volumes are constructed iteratively on the pixelwise depth residual to perform depth map refinement. While sharing similar insight with Point-MVSNet as predicting and refining depth iteratively, we show that working on cost volume pyramid can lead to a more compact, yet efficient network structure compared with the Point-MVSNet on 3D points. We further provide detailed analyses of relation between (residual) depth sampling and image resolution, which serves as a principle for building compact cost volume pyramid. Experimental results on benchmark datasets show that our model can perform 6x faster and has similar performance as state-of-the-art methods.

VACL: Variance-Aware Cross-Layer Regularization for Pruning Deep Residual Networks

Improving weight sparsity is a common strategy for producing light-weight deep neural networks. However, pruning models with residual learning is more challenging. In this paper, we introduce Variance-Aware Cross-Layer (VACL), a novel approach to address this problem. VACL consists of two parts, a Cross-Layer grouping and a Variance Aware regularization. In Cross-Layer grouping the $i^{th}$ filters of layers connected by skip-connections are grouped into one regularization group. Then, the Variance-Aware regularization term takes into account both the first and second-order statistics of the connected layers to constrain the variance within a group. Our approach can effectively improve the structural sparsity of residual models. For CIFAR10, the proposed method reduces a ResNet model by up to 79.5% with no accuracy drop and reduces a ResNeXt model by up to 82% with less than 1% accuracy drop. For ImageNet, it yields a pruned ratio of up to 63.3% with less than 1% top-5 accuracy drop. Our experimental results show that the proposed approach significantly outperforms other state-of-the-art methods in terms of overall model size and accuracy.

Quadtree Generating Networks: Efficient Hierarchical Scene Parsing with Sparse Convolutions

Semantic segmentation with Convolutional Neural Networks is a memory-intensive task due to the high spatial resolution of feature maps and output predictions. In this paper, we present Quadtree Generating Networks (QGNs), a novel approach able to drastically reduce the memory footprint of modern semantic segmentation networks. The key idea is to use quadtrees to represent the predictions and target segmentation masks instead of dense pixel grids. Our quadtree representation enables hierarchical processing of an input image, with the most computationally demanding layers only being used at regions in the image containing boundaries between classes. In addition, given a trained model, our representation enables flexible inference schemes to trade-off accuracy and computational cost, allowing the network to adapt in constrained situations such as embedded devices. We demonstrate the benefits of our approach on the Cityscapes, SUN-RGBD and ADE20k datasets. On Cityscapes, we obtain an relative 3% mIoU improvement compared to a dilated network with similar memory consumption; and only receive a 3% relative mIoU drop compared to a large dilated network, while reducing memory consumption by over 4$\times$.

Less is More: An Exploration of Data Redundancy with Active Dataset Subsampling

Deep Neural Networks (DNNs) often rely on very large datasets for training. Given the large size of such datasets, it is conceivable that they contain certain samples that either do not contribute or negatively impact the DNN's performance. If there is a large number of such samples, subsampling the training dataset in a way that removes them could provide an effective solution to both improve performance and reduce training time. In this paper, we propose an approach called Active Dataset Subsampling (ADS), to identify favorable subsets within a dataset for training using ensemble based uncertainty estimation. When applied to three image classification benchmarks (CIFAR-10, CIFAR-100 and ImageNet) we find that there are low uncertainty subsets, which can be as large as 50% of the full dataset, that negatively impact performance. These subsets are identified and removed with ADS. We demonstrate that datasets obtained using ADS with a lightweight ResNet-18 ensemble remain effective when used to train deeper models like ResNet-101. Our results provide strong empirical evidence that using all the available data for training can hurt performance on large scale vision tasks.

ExpandNets: Exploiting Linear Redundancy to Train Small Networks

While very deep networks can achieve great performance, they are ill-suited to applications in resource-constrained environments. Knowledge transfer, which leverages a deep teacher network to train a given small network, has emerged as one of the most popular strategies to address this problem. In this paper, we introduce an alternative approach to training a given small network, based on the intuition that parameter redundancy facilitates learning. We propose to expand each linear layer of a small network into multiple linear layers, without adding any nonlinearity. As such, the resulting expanded network can be compressed back to the small one algebraically, but, as evidenced by our experiments, consistently outperforms training the small network from scratch. This strategy is orthogonal to knowledge transfer. We therefore further show on several standard benchmarks that, for any knowledge transfer technique, using our expanded network as student systematically improves over using the small network.

Large-Scale Visual Active Learning with Deep Probabilistic Ensembles

Annotating the right data for training deep neural networks is an important challenge. Active learning using uncertainty estimates from Bayesian Neural Networks (BNNs) could provide an effective solution to this. Despite being theoretically principled, BNNs require approximations to be applied to large-scale problems, where both performance and uncertainty estimation are crucial. In this paper, we introduce Deep Probabilistic Ensembles (DPEs), a scalable technique that uses a regularized ensemble to approximate a deep BNN. We conduct a series of large-scale visual active learning experiments to evaluate DPEs on classification with the CIFAR-10, CIFAR-100 and ImageNet datasets, and semantic segmentation with the BDD100k dataset. Our models require significantly less training data to achieve competitive performances, and steadily improve upon strong active learning baselines as the annotation budget is increased.

Deep Probabilistic Ensembles: Approximate Variational Inference through KL Regularization

In this paper, we introduce Deep Probabilistic Ensembles (DPEs), a scalable technique that uses a regularized ensemble to approximate a deep Bayesian Neural Network (BNN). We do so by incorporating a KL divergence penalty term into the training objective of an ensemble, derived from the evidence lower bound used in variational inference. We evaluate the uncertainty estimates obtained from our models for active learning on visual classification. Our approach steadily improves upon active learning baselines as the annotation budget is increased.

The Relevance of Bayesian Layer Positioning to Model Uncertainty in Deep Bayesian Active Learning

One of the main challenges of deep learning tools is their inability to capture model uncertainty. While Bayesian deep learning can be used to tackle the problem, Bayesian neural networks often require more time and computational power to train than deterministic networks. Our work explores whether fully Bayesian networks are needed to successfully capture model uncertainty. We vary the number and position of Bayesian layers in a network and compare their performance on active learning with the MNIST dataset. We found that we can fully capture the model uncertainty by using only a few Bayesian layers near the output of the network, combining the advantages of deterministic and Bayesian networks.

Effective Use of Synthetic Data for Urban Scene Semantic Segmentation

Training a deep network to perform semantic segmentation requires large amounts of labeled data. To alleviate the manual effort of annotating real images, researchers have investigated the use of synthetic data, which can be labeled automatically. Unfortunately, a network trained on synthetic data performs relatively poorly on real images. While this can be addressed by domain adaptation, existing methods all require having access to real images during training. In this paper, we introduce a drastically different way to handle synthetic images that does not require seeing any real images at training time. Our approach builds on the observation that foreground and background classes are not affected in the same manner by the domain shift, and thus should be treated differently. In particular, the former should be handled in a detection-based manner to better account for the fact that, while their texture in synthetic images is not photo-realistic, their shape looks natural. Our experiments evidence the effectiveness of our approach on Cityscapes and CamVid with models trained on synthetic data only.