People

Each node in a neural network is trained to activate for a specific region in the input domain. Any training samples that fall within this domain are therefore implicitly clustered together. Recent work has highlighted the importance of these clusters during the training process but has not yet investigated their evolution during training. Towards this goal, we train several ReLU-activated MLPs on a simple classification task (MNIST) and show that a consistent training process emerges: (1) sample clusters initially increase in size and then decrease as training progresses, (2) the size of sample clusters in the first layer decreases more rapidly than in deeper layers, (3) binary node activations, especially of nodes in deeper layers, become more sensitive to class membership as training progresses, (4) individual nodes remain poor predictors of class membership, even if accurate when applied as a group. We report on the detail of these findings and interpret them from the perspective of a high-dimensional clustering process.

@{402,
  author = {Daniël Haasbroek, Marelie Davel},
  title = {Exploring neural network training dynamics through binary node activations},
  abstract = {Each node in a neural network is trained to activate for a specific region in the input domain. Any training samples that fall within this domain are therefore implicitly clustered together. Recent work has highlighted the importance of these clusters during the training process but has not yet investigated their evolution during training. Towards this goal, we train several ReLU-activated MLPs on a simple classification task (MNIST) and show that a consistent training process emerges: (1) sample clusters initially increase in size and then decrease as training progresses, (2) the size of sample clusters in the first layer decreases more rapidly than in deeper layers, (3) binary node activations, especially of nodes in deeper layers, become more sensitive to class membership as training progresses, (4) individual nodes remain poor predictors of class membership, even if accurate when applied as a group. We report on the detail of these findings and interpret them from the perspective of a high-dimensional clustering process.},
  year = {2020},
  journal = {Southern African Conference for Artificial Intelligence Research},
  pages = {304-320},
  month = {22/02/2021 - 26/02/2021},
  address = {South Africa},
  isbn = {978-0-620-89373-2},
  url = {https://sacair.org.za/proceedings/},
}

When training neural networks as classifiers, it is common to observe an increase in average test loss while still maintaining or improving the overall classification accuracy on the same dataset. In spite of the ubiquity of this phenomenon, it has not been well studied and is often dismissively attributed to an increase in borderline correct classifications. We present an empirical investigation that shows how this phenomenon is actually a result of the differential manner by which test samples are processed. In essence: test loss does not increase overall, but only for a small minority of samples. Large representational capacities allow losses to decrease for the vast majority of test samples at the cost of extreme increases for others. This effect seems to be mainly caused by increased parameter values relating to the correctly processed sample features. Our findings contribute to the practical understanding of a common behaviour of deep neural networks. We also discuss the implications of this work for network optimisation and generalisation.

@{401,
  author = {Arthur Venter, Marthinus Theunissen, Marelie Davel},
  title = {Pre-interpolation loss behavior in neural networks},
  abstract = {When training neural networks as classifiers, it is common to observe an increase in average test loss while still maintaining or improving the overall classification accuracy on the same dataset. In spite of the ubiquity of this phenomenon, it has not been well studied and is often dismissively attributed to an increase in borderline correct classifications. We present an empirical investigation that shows how this phenomenon is actually a result of the differential manner by which test samples are processed. In essence: test loss does not increase overall, but only for a small minority of samples. Large representational capacities allow losses to decrease for the vast majority of test samples at the cost of extreme increases for others. This effect seems to be mainly caused by increased parameter values relating to the correctly processed sample features. Our findings contribute to the practical understanding of a common behaviour of deep neural networks. We also discuss the implications of this work for network optimisation and generalisation.},
  year = {2020},
  journal = {Southern African Conference for Artificial Intelligence Research},
  pages = {296-309},
  month = {22/02/2021 - 26/02/2021},
  publisher = {Springer},
  address = {South Africa},
  isbn = {978-3-030-66151-9},
  doi = {https://doi.org/10.1007/978-3-030-66151-9_19},
}

Although Convolutional Neural Networks (CNNs) are widely used, their translation invariance (ability to deal with translated inputs) is still subject to some controversy. We explore this question using translation-sensitivity maps to quantify how sensitive a standard CNN is to a translated input. We propose the use of cosine similarity as sensitivity metric over Euclidean distance, and discuss the importance of restricting the dimensionality of either of these metrics when comparing architectures. Our main focus is to investigate the effect of different architectural components of a standard CNN on that network’s sensitivity to translation. By varying convolutional kernel sizes and amounts of zero padding, we control the size of the feature maps produced, allowing us to quantify the extent to which these elements influence translation invariance. We also measure translation invariance at different locations within the CNN to determine the extent to which convolutional and fully connected layers, respectively, contribute to the translation invariance of a CNN as a whole. Our analysis indicates that both convolutional kernel size and feature map size have a systematic influence on translation invariance. We also see that convolutional layers contribute less than expected to translation invariance, when not specifically forced to do so.

@{400,
  author = {Johannes Myburgh, Coenraad Mouton, Marelie Davel},
  title = {Tracking translation invariance in CNNs},
  abstract = {Although Convolutional Neural Networks (CNNs) are widely used, their translation invariance (ability to deal with translated inputs) is still subject to some controversy. We explore this question using translation-sensitivity maps to quantify how sensitive a standard CNN is to a translated input. We propose the use of cosine similarity as sensitivity metric over Euclidean distance, and discuss the importance of restricting the dimensionality of either of these metrics when comparing architectures. Our main focus is to investigate the effect of different architectural components of a standard CNN on that network’s sensitivity to translation. By varying convolutional kernel sizes and amounts of zero padding, we control the size of the feature maps produced, allowing us to quantify the extent to which these elements influence translation invariance. We also measure translation invariance at different locations within the CNN to determine the extent to which convolutional and fully connected layers, respectively, contribute to the translation invariance of a CNN as a whole. Our analysis indicates that both convolutional kernel size and feature map size have a systematic influence on translation invariance. We also see that convolutional layers contribute less than expected to translation invariance, when not specifically forced to do so.},
  year = {2020},
  journal = {Southern African Conference for Artificial Intelligence Research},
  pages = {282-295},
  month = {22/02/2021 - 26/02/2021},
  publisher = {Springer},
  address = {South Africa},
  isbn = {978-3-030-66151-9},
  doi = {https://doi.org/10.1007/978-3-030-66151-9_18},
}

Convolutional Neural Networks have become the standard for image classification tasks, however, these architectures are not invariant to translations of the input image. This lack of invariance is attributed to the use of stride which subsamples the input, resulting in a loss of information, and fully connected layers which lack spatial reasoning. We show that stride can greatly benefit translation invariance given that it is combined with sufficient similarity between neighbouring pixels, a characteristic which we refer to as local homogeneity. We also observe that this characteristic is dataset-specific and dictates the relationship between pooling kernel size and stride required for translation invariance. Furthermore we find that a trade-off exists between generalization and translation invariance in the case of pooling kernel size, as larger kernel sizes lead to better invariance but poorer generalization. Finally we explore the efficacy of other solutions proposed, namely global average pooling, anti-aliasing, and data augmentation, both empirically and through the lens of local homogeneity.

@{399,
  author = {Coenraad Mouton, Johannes Myburgh, Marelie Davel},
  title = {Stride and translation invariance in CNNs},
  abstract = {Convolutional Neural Networks have become the standard for image classification tasks, however, these architectures are not invariant to translations of the input image. This lack of invariance is attributed to the use of stride which subsamples the input, resulting in a loss of information, and fully connected layers which lack spatial reasoning. We show that stride can greatly benefit translation invariance given that it is combined with sufficient similarity between neighbouring pixels, a characteristic which we refer to as local homogeneity. We also observe that this characteristic is dataset-specific and dictates the relationship between pooling kernel size and stride required for translation invariance. Furthermore we find that a trade-off exists between generalization and translation invariance in the case of pooling kernel size, as larger kernel sizes lead to better invariance but poorer generalization. Finally we explore the efficacy of other solutions proposed, namely global average pooling, anti-aliasing, and data augmentation, both empirically and through the lens of local homogeneity.},
  year = {2020},
  journal = {Southern African Conference for Artificial Intelligence Research},
  pages = {267-281},
  month = {22/02/2021 - 26/02/2021},
  publisher = {Springer},
  address = {South Africa},
  isbn = {978-3-030-66151-9},
  doi = {https://doi.org/10.1007/978-3-030-66151-9_17},
}

The understanding of generalisation in machine learning is in a state of flux, in part due to the ability of deep learning models to interpolate noisy training data and still perform appropriately on out-of-sample data, thereby contradicting long-held intuitions about the bias-variance trade off in learning. We expand upon relevant existing work by discussing local attributes of neural network training within the context of a relatively simple framework.We describe how various types of noise can be compensated for within the proposed framework in order to allow the deep learning model to generalise in spite of interpolating spurious function descriptors. Empirically,we support our postulates with experiments involving overparameterised multilayer perceptrons and controlled training data noise. The main insights are that deep learning models are optimised for training data modularly, with different regions in the function space dedicated to fitting distinct types of sample information. Additionally,we show that models tend to fit uncorrupted samples first. Based on this finding, we propose a conjecture to explain an observed instance of the epoch-wise double-descent phenomenon. Our findings suggest that the notion of model capacity needs to be modified to consider the distributed way training data is fitted across sub-units.

@article{394,
  author = {Marthinus Theunissen, Marelie Davel, Etienne Barnard},
  title = {Benign interpolation of noise in deep learning},
  abstract = {The understanding of generalisation in machine learning is in a state of flux, in part due to the ability of deep learning models to interpolate noisy training data and still perform appropriately on out-of-sample data, thereby contradicting long-held intuitions about the bias-variance trade off in learning. We expand upon relevant existing work by discussing local attributes of neural network training within the context of a relatively simple framework.We describe how various types of noise can be compensated for within the proposed framework in order to allow the deep learning model to generalise in spite of interpolating spurious function descriptors. Empirically,we support our postulates with experiments involving overparameterised multilayer perceptrons and controlled training data noise. The main insights are that deep learning models are optimised for training data modularly, with different regions in the function space dedicated to fitting distinct types of sample information. Additionally,we show that models tend to fit uncorrupted samples first. Based on this finding, we propose a conjecture to explain an observed instance of the epoch-wise double-descent phenomenon. Our findings suggest that the notion of model capacity needs to be modified to consider the distributed way training data is fitted across sub-units.},
  year = {2020},
  journal = {South African Computer Journal},
  volume = {32},
  pages = {80-101},
  issue = {2},
  publisher = {South African Institute of Computer Scientists and Information Technologists},
  isbn = {ISSN: 1015-7999; E:2313-7835},
  doi = {https://doi.org/10.18489/sacj.v32i2.833},
}