Mitigating severe over-parameterization in deep convolutional neural networks through forced feature abstraction and compression with an entropy-based heuristic

Convolutional Neural Networks (CNNs) such as ResNet-50, DenseNet-40 and ResNeXt-56 are severely over-parameterized, necessitating a consequent increase in the computational resources required for model training which scales exponentially for increments in model depth. In this paper, we propose an Entropy-Based Convolutional Layer Estimation (EBCLE) heuristic which is robust and simple, yet effective in resolving the problem of over-parameterization with regards to network depth of CNN model. The EBCLE heuristic employs a priori knowledge of the entropic data distribution of input datasets to determine an upper bound for convolutional network depth, beyond which identity transformations are prevalent offering insignificant contributions for enhancing model performance. Restricting depth redundancies by forcing feature compression and abstraction restricts over-parameterization while decreasing training time by 24.99% - 78.59% without degradation in model performance. We present empirical evidence to emphasize the relative effectiveness of broader, yet shallower models trained using the EBCLE heuristic, which maintains or outperforms baseline classification accuracies of narrower yet deeper models. The EBCLE heuristic is architecturally agnostic and EBCLE based CNN models restrict depth redundancies resulting in enhanced utilization of the available computational resources. The proposed EBCLE heuristic is a compelling technique for researchers to analytically justify their HyperParameter (HP) choices for CNNs. Empirical validation of the EBCLE heuristic in training CNN models was established on five benchmarking datasets (ImageNet32, CIFAR-10/100, STL-10, MNIST) and four network architectures (DenseNet, ResNet, ResNeXt and EfficientNet B0-B2) with appropriate statistical tests employed to infer any conclusive claims presented in this paper.

Examining and Mitigating Kernel Saturation in Convolutional Neural Networks using Negative Images

Neural saturation in Deep Neural Networks (DNNs) has been studied extensively, but remains relatively unexplored in Convolutional Neural Networks (CNNs). Understanding and alleviating the effects of convolutional kernel saturation is critical for enhancing CNN models classification accuracies. In this paper, we analyze the effect of convolutional kernel saturation in CNNs and propose a simple data augmentation technique to mitigate saturation and increase classification accuracy, by supplementing negative images to the training dataset. We hypothesize that greater semantic feature information can be extracted using negative images since they have the same structural information as standard images but differ in their data representations. Varied data representations decrease the probability of kernel saturation and thus increase the effectiveness of kernel weight updates. The two datasets selected to evaluate our hypothesis were CIFAR- 10 and STL-10 as they have similar image classes but differ in image resolutions thus making for a better understanding of the saturation phenomenon. MNIST dataset was used to highlight the ineffectiveness of the technique for linearly separable data. The ResNet CNN architecture was chosen since the skip connections in the network ensure the most important features contributing the most to classification accuracy are retained. Our results show that CNNs are indeed susceptible to convolutional kernel saturation and that supplementing negative images to the training dataset can offer a statistically significant increase in classification accuracies when compared against models trained on the original datasets. Our results present accuracy increases of 6.98% and 3.16% on the STL-10 and CIFAR-10 datasets respectively.

Examining convolutional feature extraction using Maximum Entropy (ME) and Signal-to-Noise Ratio (SNR) for image classification

Convolutional Neural Networks (CNNs) specialize in feature extraction rather than function mapping. In doing so they form complex internal hierarchical feature representations, the complexity of which gradually increases with a corresponding increment in neural network depth. In this paper, we examine the feature extraction capabilities of CNNs using Maximum Entropy (ME) and Signal-to-Noise Ratio (SNR) to validate the idea that, CNN models should be tailored for a given task and complexity of the input data. SNR and ME measures are used as they can accurately determine in the input dataset, the relative amount of signal information to the random noise and the maximum amount of information respectively. We use two well known benchmarking datasets, MNIST and CIFAR-10 to examine the information extraction and abstraction capabilities of CNNs. Through our experiments, we examine convolutional feature extraction and abstraction capabilities in CNNs and show that the classification accuracy or performance of CNNs is greatly dependent on the amount, complexity and quality of the signal information present in the input data. Furthermore, we show the effect of information overflow and underflow on CNN classification accuracies. Our hypothesis is that the feature extraction and abstraction capabilities of convolutional layers are limited and therefore, CNN models should be tailored to the input data by using appropriately sized CNNs based on the SNR and ME measures of the input dataset.