Pruning large neural networks while maintaining their performance is often desirable due to the reduced space and time complexity. In existing methods, pruning done within an iterative optimization procedure with either heuristically designed schedules or additional hyperparameters, undermining utility. this work, we present a new approach that prunes given network once at initialization prior training. To achieve this, introduce saliency criterion based on connection sensitivity identifies...

In continual learning (CL), an agent learns from a stream of tasks leveraging prior experience to transfer knowledge future tasks. It is ideal framework decrease the amount supervision in existing algorithms. But for successful transfer, learner needs remember how perform previous One way endow ability seen past store small memory, dubbed episodic that stores few examples and then replay these when training this work, we empirically analyze effectiveness very memory CL setup where each...

Real world applications of stereo depth estimation require models that are robust to dynamic variations in the environment. Even though deep learning based methods successful, they often fail generalize unseen environment, making them less suitable for practical such as autonomous driving. In this work, we introduce a ``learning-to-adapt'' framework enables continuously adapt new target domains an unsupervised manner. Specifically, our approach incorporates adaptation procedure into...

We introduce Retrieval Augmented Classification (RAC), a generic approach to augmenting standard image classification pipelines with an explicit retrieval module. RAC consists of base encoder fused parallel branch that queries non-parametric external memory pre-encoded images and associated text snippets. apply the problem long-tail demonstrate significant improvement over previous state-of-the-art on Places365-LT iNaturalist-2018 (14.5% 6.7% respectively), despite using only training...

Network pruning is a promising avenue for compressing deep neural networks. A typical approach to starts by training model and then removing redundant parameters while minimizing the impact on what learned. Alternatively, recent shows that can be done at initialization prior training, based saliency criterion called connection sensitivity. However, it remains unclear exactly why an untrained, randomly initialized network effective. In this work, noting sensitivity as form of gradient, we...

Semi-supervised semantic segmentation methods use a small amount of clean pixel-level annotations to guide the interpretation larger quantity unlabelled image data. The challenges providing pixel-accurate at scale mean that labels are typically noisy, and this contaminates final results. In work, we propose an approach is robust label noise in annotated method uses two diverse learning groups with different network architectures effectively handle both images. Each group consists teacher...

Calibrating neural networks is of utmost importance when employing them in safety-critical applications where the downstream decision making depends on predicted probabilities. Measuring calibration error amounts to comparing two empirical distributions. In this work, we introduce a binning-free measure inspired by classical Kolmogorov-Smirnov (KS) statistical test which main idea compare respective cumulative probability From this, approximating distribution using differentiable function...

Incomplete tabular datasets are ubiquitous in many applications for a number of reasons such as human error data collection or privacy considerations.One would expect natural solution this is to utilize powerful generative models diffusion models, which have demonstrated great potential across image and continuous domains.However, vanilla often exhibit sensitivity initialized noises.This, along with the sparsity inherent domain, poses challenges thereby impacting robustness these...

We propose a deep generative model of humans in natural images which keeps 2D pose separated from other latent factors variation, such as background scene and clothing. In contrast to methods that learn models low-dimensional representations, e.g., segmentation masks skeletons, our single-stage end-to-end conditional-VAEGAN learns directly on the image space. The flexibility this approach allows sampling people with independent variations appearance. Moreover, it enables reconstruction...

Compressing large Neural Networks (NN) by quantizing the parameters, while maintaining performance is highly desirable due to reduced memory and time complexity. In this work, we cast NN quantization as a discrete labelling problem, examining relaxations, design an efficient iterative optimization procedure that involves stochastic gradient descent followed projection. We prove our simple projected approach is, in fact, equivalent proximal version of well-known mean-field method. These...

While widely acknowledged as highly effective in computer vision, multi-label MRFs with non-convex priors are difficult to optimize. To tackle this, we introduce an algorithm that iteratively approximates the original energy appropriately weighted surrogate is easier minimize. Our guarantees decreases at each iteration. In particular, consider scenario where global minimizer of can be obtained by a graph cut algorithm, and show our then lets us handle large variety priors. We demonstrate...

Neural network quantization has become increasingly popular due to efficient memory consumption and faster computation resulting from bitwise operations on the quantized networks. Even though they exhibit excellent generalization capabilities, their robustness properties are not well-understood. In this work, we systematically study of networks against gradient based adversarial attacks demonstrate that these models suffer vanishing issues show a fake sense robustness. By attributing poor...

Self-supervised learning (SSL) aims to produce useful feature representations without access any human-labeled data annotations. Due the success of recent SSL methods based on contrastive learning, such as SimCLR, this problem has gained popularity. Most current approaches append a parametrized projection head end some backbone network optimize InfoNCE objective and then discard learned after training. This raises fundamental question: Why is learnable required if we are it training? In...

The fully connected conditional random field (CRF) with Gaussian pairwise potentials has proven popular and effective for multi-class semantic segmentation. While the energy of a dense CRF can be minimized accurately using linear programming (LP) relaxation, state-of-the-art algorithm is too slow to useful in practice. To alleviate this deficiency, we introduce an efficient LP minimization CRFs. end, develop proximal framework, where dual each problem optimized via block coordinate descent....

Dense conditional random fields (CRFs) have become a popular framework for modeling several problems in computer vision such as stereo correspondence and multiclass semantic segmentation. By long-range interactions, dense CRFs provide labeling that captures finer detail than their sparse counterparts. Currently, the state-of-the-art algorithm performs mean-field inference using filter-based method but fails to strong theoretical guarantee on quality of solution. A question naturally arises...