Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userRethinking Positive Aggregation and Propagation of Gradients in Gradient-based Saliency Methods
67
0
0.0
(
0
)
Added by
Ashkan Khakzar
Publication date
2020
fields
Informatics Engineering
and research's language is
English
Ask ChatGPT about the research
No Arabic abstract
Saliency methods interpret the prediction of a neural network by showing the importance of input elements for that prediction. A popular family of saliency methods utilize gradient information. In this work, we empirically show that two approaches for handling the gradient information, namely positive aggregation, and positive propagation, break these methods. Though these methods reflect visually salient information in the input, they do not explain the model prediction anymore as the generated saliency maps are insensitive to the predicted output and are insensitive to model parameter randomization. Specifically for methods that aggregate the gradients of a chosen layer such as GradCAM++ and FullGrad, exclusively aggregating positive gradients is detrimental. We further support this by proposing several variants of aggregation methods with positive handling of gradient information. For methods that backpropagate gradient information such as LRP, RectGrad, and Guided Backpropagation, we show the destructive effect of exclusively propagating positive gradient information.
rate research
Read More
In recent years, gradient-based methods for solving bi-level optimization tasks have drawn a great deal of interest from the machine learning community. However, to calculate the gradient of the best response, existing research always relies on the singleton of the lower-level solution set (a.k.a., Lower-Level Singleton, LLS). In this work, by formulating bi-level models from an optimistic bi-level viewpoint, we first establish a novel Bi-level Descent Aggregation (BDA) framework, which aggregates hierarchical objectives of both upper level and lower level. The flexibility of our framework benefits from the embedded replaceable task-tailored iteration dynamics modules, thereby capturing a wide range of bi-level learning tasks. Theoretically, we derive a new methodology to prove the convergence of BDA framework without the LLS restriction. Besides, the new proof recipe we propose is also engaged to improve the convergence results of conventional gradient-based bi-level methods under the LLS simplification. Furthermore, we employ a one-stage technique to accelerate the back-propagation calculation in a numerical manner. Extensive experiments justify our theoretical results and demonstrate the superiority of the proposed algorithm for hyper-parameter optimization and meta-learning tasks.
We study a new aggregation operator for gradients coming from a mini-batch for stochastic gradient (SG) methods that allows a significant speed-up in the case of sparse optimization problems. We call this method AdaBatch and it only requires a few lines of code change compared to regular mini-batch SGD algorithms. We provide a theoretical insight to understand how this new class of algorithms is performing and show that it is equivalent to an implicit per-coordinate rescaling of the gradients, similarly to what Adagrad methods can do. In theory and in practice, this new aggregation allows to keep the same sample efficiency of SG methods while increasing the batch size. Experimentally, we also show that in the case of smooth convex optimization, our procedure can even obtain a better loss when increasing the batch size for a fixed number of samples. We then apply this new algorithm to obtain a parallelizable stochastic gradient method that is synchronous but allows speed-up on par with Hogwild! methods as convergence does not deteriorate with the increase of the batch size. The same approach can be used to make mini-batch provably efficient for variance-reduced SG methods such as SVRG.
Input perturbation methods occlude parts of an input to a function and measure the change in the functions output. Recently, input perturbation methods have been applied to generate and evaluate saliency maps from convolutional neural networks. In practice, neutral baseline images are used for the occlusion, such that the baseline images impact on the classification probability is minimal. However, in this paper we show that arguably neutral baseline images still impact the generated saliency maps and their evaluation with input perturbations. We also demonstrate that many choices of hyperparameters lead to the divergence of saliency maps generated by input perturbations. We experimentally reveal inconsistencies among a selection of input perturbation methods and find that they lack robustness for generating saliency maps and for evaluating saliency maps as saliency metrics.
We build a theoretical framework for designing and understanding practical meta-learning methods that integrates sophisticated formalizations of task-similarity with the extensive literature on online convex optimization and sequential prediction algorithms. Our approach enables the task-similarity to be learned adaptively, provides sharper transfer-risk bounds in the setting of statistical learning-to-learn, and leads to straightforward derivations of average-case regret bounds for efficient algorithms in settings where the task-environment changes dynamically or the tasks share a certain geometric structure. We use our theory to modify several popular meta-learning algorithms and improve their meta-test-time performance on standard problems in few-shot learning and federated learning.
Training sparse networks to converge to the same performance as dense neural architectures has proven to be elusive. Recent work suggests that initialization is the key. However, while this direction of research has had some success, focusing on initialization alone appears to be inadequate. In this paper, we take a broader view of training sparse networks and consider the role of regularization, optimization, and architecture choices on sparse models. We propose a simple experimental framework, Same Capacity Sparse vs Dense Comparison (SC-SDC), that allows for a fair comparison of sparse and dense networks. Furthermore, we propose a new measure of gradient flow, Effective Gradient Flow (EGF), that better correlates to performance in sparse networks. Using top-line metrics, SC-SDC and EGF, we show that default choices of optimizers, activation functions and regularizers used for dense networks can disadvantage sparse networks. Based upon these findings, we show that gradient flow in sparse networks can be improved by reconsidering aspects of the architecture design and the training regime. Our work suggests that initialization is only one piece of the puzzle and taking a wider view of tailoring optimization to sparse networks yields promising results.
suggested questions
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
