Selectivity considered harmful: evaluating the causal impact of class selectivity in DNNs

• URL: http://arxiv.org/abs/2003.01262v3
• Date: Wed, 14 Oct 2020 17:31:24 GMT
• Title: Selectivity considered harmful: evaluating the causal impact of class selectivity in DNNs
• Authors: Matthew L. Leavitt and Ari Morcos
• Abstract summary: We investigate the causal impact of class selectivity on network function by directly regularizing for or against class selectivity. Using this regularizer to reduce class selectivity across units in convolutional neural networks increased test accuracy by over 2% for ResNet18 trained on Tiny ImageNet. For ResNet20 trained on CIFAR10 we could reduce class selectivity by a factor of 2.5 with no impact on test accuracy, and reduce it nearly to zero with only a small ($sim$2%) drop in test accuracy.
• Score: 7.360807642941714
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: The properties of individual neurons are often analyzed in order to understand the biological and artificial neural networks in which they're embedded. Class selectivity-typically defined as how different a neuron's responses are across different classes of stimuli or data samples-is commonly used for this purpose. However, it remains an open question whether it is necessary and/or sufficient for deep neural networks (DNNs) to learn class selectivity in individual units. We investigated the causal impact of class selectivity on network function by directly regularizing for or against class selectivity. Using this regularizer to reduce class selectivity across units in convolutional neural networks increased test accuracy by over 2% for ResNet18 trained on Tiny ImageNet. For ResNet20 trained on CIFAR10 we could reduce class selectivity by a factor of 2.5 with no impact on test accuracy, and reduce it nearly to zero with only a small ($\sim$2%) drop in test accuracy. In contrast, regularizing to increase class selectivity significantly decreased test accuracy across all models and datasets. These results indicate that class selectivity in individual units is neither sufficient nor strictly necessary, and can even impair DNN performance. They also encourage caution when focusing on the properties of single units as representative of the mechanisms by which DNNs function.

Related papers

• Unveiling the Power of Sparse Neural Networks for Feature Selection [60.50319755984697]
  Sparse Neural Networks (SNNs) have emerged as powerful tools for efficient feature selection. We show that SNNs trained with dynamic sparse training (DST) algorithms can achieve, on average, more than $50%$ memory and $55%$ FLOPs reduction. Our findings show that feature selection with SNNs trained with DST algorithms can achieve, on average, more than $50%$ memory and $55%$ FLOPs reduction.
  arXiv  Detail & Related papers  (2024-08-08T16:48:33Z)
• Neuroevolutionary algorithms driven by neuron coverage metrics for semi-supervised classification [60.60571130467197]
  In some machine learning applications the availability of labeled instances for supervised classification is limited while unlabeled instances are abundant. We introduce neuroevolutionary approaches that exploit unlabeled instances by using neuron coverage metrics computed on the neural network architecture encoded by each candidate solution.
  arXiv  Detail & Related papers  (2023-03-05T23:38:44Z)
• Much Easier Said Than Done: Falsifying the Causal Relevance of Linear Decoding Methods [1.3999481573773074]
  Linear classifier probes identify highly selective units as the most important for network function. In spite of the absence of ablation effects for selective neurons, linear decoding methods can be effectively used to interpret network function. More specifically, we find that an interaction between selectivity and the average activity of the unit better predicts ablation performance deficits for groups of units in AlexNet, VGG16, MobileNetV2, and ResNet101.
  arXiv  Detail & Related papers  (2022-11-08T16:43:02Z)
• Efficient and Robust Classification for Sparse Attacks [34.48667992227529]
  We consider perturbations bounded by the $ell$--norm, which have been shown as effective attacks in the domains of image-recognition, natural language processing, and malware-detection. We propose a novel defense method that consists of "truncation" and "adrial training" Motivated by the insights we obtain, we extend these components to neural network classifiers.
  arXiv  Detail & Related papers  (2022-01-23T21:18:17Z)
• Debiased Graph Neural Networks with Agnostic Label Selection Bias [59.61301255860836]
  Most existing Graph Neural Networks (GNNs) are proposed without considering the selection bias in data. We propose a novel Debiased Graph Neural Networks (DGNN) with a differentiated decorrelation regularizer. Our proposed model outperforms the state-of-the-art methods and DGNN is a flexible framework to enhance existing GNNs.
  arXiv  Detail & Related papers  (2022-01-19T16:50:29Z)
• Provably Training Neural Network Classifiers under Fairness Constraints [70.64045590577318]
  We show that overparametrized neural networks could meet the constraints. Key ingredient of building a fair neural network classifier is establishing no-regret analysis for neural networks.
  arXiv  Detail & Related papers  (2020-12-30T18:46:50Z)
• Linking average- and worst-case perturbation robustness via class selectivity and dimensionality [7.360807642941714]
  We investigate whether class selectivity confers robustness (or vulnerability) to perturbations of input data. We found that networks regularized to have lower levels of class selectivity were more robust to average-case perturbations. In contrast, class selectivity increases robustness to multiple types of worst-case perturbations.
  arXiv  Detail & Related papers  (2020-10-14T00:45:29Z)
• On the relationship between class selectivity, dimensionality, and robustness [25.48362370177062]
  We investigate whether class selectivity confers robustness (or vulnerability) to perturbations of input data. We found that mean class selectivity predicts vulnerability to naturalistic corruptions. We found that class selectivity increases robustness to multiple types of gradient-based adversarial attacks.
  arXiv  Detail & Related papers  (2020-07-08T21:24:45Z)
• Prior choice affects ability of Bayesian neural networks to identify unknowns [0.0]
  We show that the choice of priors has a substantial impact on the ability of the model to confidently assign data to the correct class. We also show that testing alternative options can improve the performance of BNNs.
  arXiv  Detail & Related papers  (2020-05-11T10:32:47Z)
• Learning Class Regularized Features for Action Recognition [68.90994813947405]
  We introduce a novel method named Class Regularization that performs class-based regularization of layer activations. We show that using Class Regularization blocks in state-of-the-art CNN architectures for action recognition leads to systematic improvement gains of 1.8%, 1.2% and 1.4% on the Kinetics, UCF-101 and HMDB-51 datasets, respectively.
  arXiv  Detail & Related papers  (2020-02-07T07:27:49Z)
• Approximation and Non-parametric Estimation of ResNet-type Convolutional Neural Networks [52.972605601174955]
  We show a ResNet-type CNN can attain the minimax optimal error rates in important function classes. We derive approximation and estimation error rates of the aformentioned type of CNNs for the Barron and H"older classes.
  arXiv  Detail & Related papers  (2019-03-24T19:42:39Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.