Recurrent models are becoming a popular choice for video enhancement tasks such as video denoising. In this work, we focus on their stability as dynamical systems and show that they tend to fail catastrophically at inference time on long video sequences. To address this issue, we (1) introduce a diagnostic tool which produces adversarial input sequences optimized to trigger instabilities and that can be interpreted as visualizations of spatio-temporal receptive fields, and (2) propose two approaches to enforce the stability of a model: constraining the spectral norm or constraining the stable rank of its convolutional layers. We then introduce Stable Rank Normalization of the Layers (SRNL), a new algorithm that enforces these constraints, and verify experimentally that it successfully results in stable recurrent video processing.
Recent studies have shown that skeletonization (pruning parameters) of networks at initialization provides all the practical benefits of sparsity both at inference and training time, while only marginally degrading their performance. However, we observe that beyond a certain level of sparsity (approx 95%), these approaches fail to preserve the network performance, and to our surprise, in many cases perform even worse than trivial random pruning. To this end, we propose to find a skeletonized network with maximum foresight connection sensitivity (FORCE). Intuitively, out of all possible sub-networks, we propose to find the one whose connections would have a maximum impact on the loss when perturbed. Our approximate solution to maximize the FORCE, progressively prunes connections of a given network at initialization. This allows parameters that were unimportant at earlier stages of skeletonization to become important at later stages. In many cases, our approach enables us to remove up to 99.9% parameters, while keeping networks trainable and providing significantly better performance than recent approaches. We demonstrate the effectiveness of our approach at various levels of sparsity (from medium to extreme) through extensive experiments and analysis. Code can be found in https://github.com/naver/force.
Evaluating Visual Dialogue, the task of answering a sequence of questions relating to a visual input, remains an open research challenge. The current evaluation scheme of the VisDial dataset computes the ranks of ground-truth answers in predefined candidate sets, which Massiceti et al. (2018) show can be susceptible to the exploitation of dataset biases. This scheme also does little to account for the different ways of expressing the same answer--an aspect of language that has been well studied in NLP. We propose a revised evaluation scheme for the VisDial dataset leveraging metrics from the NLP literature to measure consensus between answers generated by the model and a set of relevant answers. We construct these relevant answer sets using a simple and effective semi-supervised method based on correlation, which allows us to automatically extend and scale sparse relevance annotations from humans to the entire dataset. We release these sets and code for the revised evaluation scheme as DenseVisDial, and intend them to be an improvement to the dataset in the face of its existing constraints and design choices.
Miscalibration -- a mismatch between a model's confidence and its correctness -- of Deep Neural Networks (DNNs) makes their predictions hard to rely on. Ideally, we want networks to be accurate, calibrated and confident. We show that, as opposed to the standard cross-entropy loss, focal loss (Lin et al., 2017) allows us to learn models that are already very well calibrated. When combined with temperature scaling, whilst preserving accuracy, it yields state-of-the-art calibrated models. We provide a thorough analysis of the factors causing miscalibration, and use the insights we glean from this to justify the empirically excellent performance of focal loss. To facilitate the use of focal loss in practice, we also provide a principled approach to automatically select the hyperparameter involved in the loss function. We perform extensive experiments on a variety of computer vision and NLP datasets, and with a wide variety of network architectures, and show that our approach achieves state-of-the-art accuracy and calibration in almost all cases.
In continual learning, the learner faces a stream of data whose distribution changes over time. Modern neural networks are known to suffer under this setting, as they quickly forget previously acquired knowledge. To address such catastrophic forgetting, many continual learning methods implement different types of experience replay, re-learning on past data stored in a small buffer known as episodic memory. In this work, we complement experience replay with a new objective that we call anchoring, where the learner uses bilevel optimization to update its knowledge on the current task, while keeping intact the predictions on some anchor points of past tasks. These anchor points are learned using gradient-based optimization to maximize forgetting, which is approximated by fine-tuning the currently trained model on the episodic memory of past tasks. Experiments on several supervised learning benchmarks for continual learning demonstrate that our approach improves the standard experience replay in terms of both accuracy and forgetting metrics and for various sizes of episodic memories.
Quantizing large Neural Networks (NN) while maintaining the performance is highly desirable for resource-limited devices due to reduced memory and time complexity. NN quantization is usually formulated as a constrained optimization problem and optimized via a modified version of gradient descent. In this work, by interpreting the continuous parameters (unconstrained) as the dual of the quantized ones, we introduce a Mirror Descent (MD) framework (Bubeck (2015)) for NN quantization. Specifically, we provide conditions on the projections (i.e., mapping from continuous to quantized ones) which would enable us to derive valid mirror maps and in turn the respective MD updates. Furthermore, we discuss a numerically stable implementation of MD by storing an additional set of auxiliary dual variables (continuous). This update is strikingly analogous to the popular Straight Through Estimator (STE) based method which is typically viewed as a "trick" to avoid vanishing gradients issue but here we show that it is an implementation method for MD for certain projections. Our experiments on standard classification datasets (CIFAR-10/100, TinyImageNet) with convolutional and residual architectures show that our MD variants obtain fully-quantized networks with accuracies very close to the floating-point networks.