Anna Sergeevna Bosman

Savvas Panagiotou, Anna S. Bosman

Low-light image enhancement (LLIE) techniques attempt to increase the visibility of images captured in low-light scenarios. However, as a result of enhancement, a variety of image degradations such as noise and color bias are revealed. Furthermore, each particular LLIE approach may introduce a different form of flaw within its enhanced results. To combat these image degradations, post-processing denoisers have widely been used, which often yield oversmoothed results lacking detail. We propose using a diffusion model as a post-processing approach, and we introduce Low-light Post-processing Diffusion Model (LPDM) in order to model the conditional distribution between under-exposed and normally-exposed images. We apply LPDM in a manner which avoids the computationally expensive generative reverse process of typical diffusion models, and post-process images in one pass through LPDM. Extensive experiments demonstrate that our approach outperforms competing post-processing denoisers by increasing the perceptual quality of enhanced low-light images on a variety of challenging low-light datasets. Source code is available at https://github.com/savvaki/LPDM.

Anna Sergeevna Bosman, Andries Engelbrecht, Marde Helbig

Activation functions play a significant role in neural network design by enabling non-linearity. The choice of activation function was previously shown to influence the properties of the resulting loss landscape. Understanding the relationship between activation functions and loss landscape properties is important for neural architecture and training algorithm design. This study empirically investigates neural network loss landscapes associated with hyperbolic tangent, rectified linear unit, and exponential linear unit activation functions. Rectified linear unit is shown to yield the most convex loss landscape, and exponential linear unit is shown to yield the least flat loss landscape, and to exhibit superior generalisation performance. The presence of wide and narrow valleys in the loss landscape is established for all activation functions, and the narrow valleys are shown to correlate with saturated neurons and implicitly regularised network configurations.

Arné Schreuder, Anna Bosman, Andries Engelbrecht et al.

The process of training feedforward neural networks (FFNNs) can benefit from an automated process where the best heuristic to train the network is sought out automatically by means of a high-level probabilistic-based heuristic. This research introduces a novel population-based Bayesian hyper-heuristic (BHH) that is used to train feedforward neural networks (FFNNs). The performance of the BHH is compared to that of ten popular low-level heuristics, each with different search behaviours. The chosen heuristic pool consists of classic gradient-based heuristics as well as meta-heuristics (MHs). The empirical process is executed on fourteen datasets consisting of classification and regression problems with varying characteristics. The BHH is shown to be able to train FFNNs well and provide an automated method for finding the best heuristic to train the FFNNs at various stages of the training process.

Matthew C. Dickson, Anna S. Bosman, Katherine M. Malan

Loss functions play an important role in the training of artificial neural networks (ANNs), and can affect the generalisation ability of the ANN model, among other properties. Specifically, it has been shown that the cross entropy and sum squared error loss functions result in different training dynamics, and exhibit different properties that are complementary to one another. It has previously been suggested that a hybrid of the entropy and sum squared error loss functions could combine the advantages of the two functions, while limiting their disadvantages. The effectiveness of such hybrid loss functions is investigated in this study. It is shown that hybridisation of the two loss functions improves the generalisation ability of the ANNs on all problems considered. The hybrid loss function that starts training with the sum squared error loss function and later switches to the cross entropy error loss function is shown to either perform the best on average, or to not be significantly different than the best loss function tested for all problems considered. This study shows that the minima discovered by the sum squared error loss function can be further exploited by switching to cross entropy error loss function. It can thus be concluded that hybridisation of the two loss functions could lead to better performance in ANNs.

Thamsanqa Mlotshwa, Heinrich van Deventer, Anna Sergeevna Bosman

In supervised machine learning, the choice of loss function implicitly assumes a particular noise distribution over the data. For example, the frequently used mean squared error (MSE) loss assumes a Gaussian noise distribution. The choice of loss function during training and testing affects the performance of artificial neural networks (ANNs). It is known that MSE may yield substandard performance in the presence of outliers. The Cauchy loss function (CLF) assumes a Cauchy noise distribution, and is therefore potentially better suited for data with outliers. This papers aims to determine the extent of robustness and generalisability of the CLF as compared to MSE. CLF and MSE are assessed on a few handcrafted regression problems, and a real-world regression problem with artificially simulated outliers, in the context of ANN training. CLF yielded results that were either comparable to or better than the results yielded by MSE, with a few notable exceptions.

Isak Potgieter, Christopher W. Cleghorn, Anna S. Bosman

This study investigates the use of local optima network (LON) analysis, a derivative of the fitness landscape of candidate solutions, to characterise and visualise the neural architecture space. The search space of feedforward neural network architectures with up to three layers, each with up to 10 neurons, is fully enumerated by evaluating trained model performance on a selection of data sets. Extracted LONs, while heterogeneous across data sets, all exhibit simple global structures, with single global funnels in all cases but one. These results yield early indication that LONs may provide a viable paradigm by which to analyse and optimise neural architectures.

Heinrich van Deventer, Pieter Janse van Rensburg, Anna Bosman

Neural networks have been criticised for their inability to perform continual learning due to catastrophic forgetting and rapid unlearning of a past concept when a new concept is introduced. Catastrophic forgetting can be alleviated by specifically designed models and training techniques. This paper outlines a novel Spline Additive Model (SAM). SAM exhibits intrinsic memory retention with sufficient expressive power for many practical tasks, but is not a universal function approximator. SAM is extended with the Kolmogorov-Arnold representation theorem to a novel universal function approximator, called the Kolmogorov-Arnold Spline Additive Model - KASAM. The memory retention, expressive power and limitations of SAM and KASAM are illustrated analytically and empirically. SAM exhibited robust but imperfect memory retention, with small regions of overlapping interference in sequential learning tasks. KASAM exhibited greater susceptibility to catastrophic forgetting. KASAM in combination with pseudo-rehearsal training techniques exhibited superior performance in regression tasks and memory retention.

Hermann Tapamo, Anna Bosman, James Maina et al.

Flexible road pavements deteriorate primarily due to traffic and adverse environmental conditions. Cracking is the most common deterioration mechanism; the surveying thereof is typically conducted manually using internationally defined classification standards. In South Africa, the use of high-definition video images has been introduced, which allows for safer road surveying. However, surveying is still a tedious manual process. Automation of the detection of defects such as cracks would allow for faster analysis of road networks and potentially reduce human bias and error. This study performs a comparison of six state-of-the-art convolutional neural network models for the purpose of crack detection. The models are pretrained on the ImageNet dataset, and fine-tuned using a new real-world binary crack dataset consisting of 14000 samples. The effects of dataset augmentation are also investigated. Of the six models trained, five achieved accuracy above 97%. The highest recorded accuracy was 98%, achieved by the ResNet and VGG16 models. The dataset is available at the following URL: https://zenodo.org/record/7795975

Jarrod Goschen, Anna Sergeevna Bosman, Stefan Gruner

Ongoing progress in computational intelligence (CI) has led to an increased desire to apply CI techniques for the purpose of improving software engineering processes, particularly software testing. Existing state-of-the-art automated software testing techniques focus on utilising search algorithms to discover input values that achieve high execution path coverage. These algorithms are trained on the same code that they intend to test, requiring instrumentation and lengthy search times to test each software component. This paper outlines a novel genetic programming framework, where the evolved solutions are not input values, but micro-programs that can repeatedly generate input values to efficiently explore a software component's input parameter domain. We also argue that our approach can be generalised such as to be applied to many different software systems, and is thus not specific to merely the particular software component on which it was trained.

Christiaan P. Opperman, Anna S. Bosman, Katherine M. Malan

Despite huge successes on a wide range of tasks, neural networks are known to sometimes struggle to generalise to unseen data. Many approaches have been proposed over the years to promote the generalisation ability of neural networks, collectively known as regularisation techniques. These are used as common practice under the assumption that any regularisation added to the pipeline would result in a performance improvement. In this study, we investigate whether this assumption holds in practice. First, we provide a broad review of regularisation techniques, including modern theories such as double descent. We propose a taxonomy of methods under four broad categories, namely: (1) data-based strategies, (2) architecture strategies, (3) training strategies, and (4) loss function strategies. Notably, we highlight the contradictions and correspondences between the approaches in these broad classes. Further, we perform an empirical comparison of the various regularisation techniques on classification tasks for ten numerical and image datasets applied to the multi-layer perceptron and convolutional neural network architectures. Results show that the efficacy of regularisation is dataset-dependent. For example, the use of a regularisation term only improved performance on numeric datasets, whereas batch normalisation improved performance on image datasets only. Generalisation is crucial to machine learning; thus, understanding the effects of applying regularisation techniques, and considering the connections between them is essential to the appropriate use of these methods in practice.

Moseli Mots'oehli, Anna Sergeevna Bosman, Johan Pieter De Villiers

Algorithmic music composition is a way of composing musical pieces with minimal to no human intervention. While recurrent neural networks are traditionally applied to many sequence-to-sequence prediction tasks, including successful implementations of music composition, their standard supervised learning approach based on input-to-output mapping leads to a lack of note variety. These models can therefore be seen as potentially unsuitable for tasks such as music generation. Generative adversarial networks learn the generative distribution of data and lead to varied samples. This work implements and compares adversarial and non-adversarial training of recurrent neural network music composers on MIDI data. The resulting music samples are evaluated by human listeners, their preferences recorded. The evaluation indicates that adversarial training produces more aesthetically pleasing music.

Heinrich van Deventer, Anna Bosman

Artificial neural networks (ANNs), despite their universal function approximation capability and practical success, are subject to catastrophic forgetting. Catastrophic forgetting refers to the abrupt unlearning of a previous task when a new task is learned. It is an emergent phenomenon that hinders continual learning. Existing universal function approximation theorems for ANNs guarantee function approximation ability, but do not predict catastrophic forgetting. This paper presents a novel universal approximation theorem for multi-variable functions using only single-variable functions and exponential functions. Furthermore, we present ATLAS: a novel ANN architecture based on the new theorem. It is shown that ATLAS is a universal function approximator capable of some memory retention, and continual learning. The memory of ATLAS is imperfect, with some off-target effects during continual learning, but it is well-behaved and predictable. An efficient implementation of ATLAS is provided. Experiments are conducted to evaluate both the function approximation and memory retention capabilities of ATLAS.

Reinhard Booysen, Anna Sergeevna Bosman

Artificial neural network (NN) architecture design is a nontrivial and time-consuming task that often requires a high level of human expertise. Neural architecture search (NAS) serves to automate the design of NN architectures and has proven to be successful in automatically finding NN architectures that outperform those manually designed by human experts. NN architecture performance can be quantified based on multiple objectives, which include model accuracy and some NN architecture complexity objectives, among others. The majority of modern NAS methods that consider multiple objectives for NN architecture performance evaluation are concerned with automated feed forward NN architecture design, which leaves multi-objective automated recurrent neural network (RNN) architecture design unexplored. RNNs are important for modeling sequential datasets, and prominent within the natural language processing domain. It is often the case in real world implementations of machine learning and NNs that a reasonable trade-off is accepted for marginally reduced model accuracy in favour of lower computational resources demanded by the model. This paper proposes a multi-objective evolutionary algorithm-based RNN architecture search method. The proposed method relies on approximate network morphisms for RNN architecture complexity optimisation during evolution. The results show that the proposed method is capable of finding novel RNN architectures with comparable performance to state-of-the-art manually designed RNN architectures, but with reduced computational demand.

Heinrich van Deventer, Anna Sergeevna Bosman

Continual learning is the sequential learning of different tasks by a machine learning model. Continual learning is known to be hindered by catastrophic interference or forgetting, i.e. rapid unlearning of earlier learned tasks when new tasks are learned. Despite their practical success, artificial neural networks (ANNs) are prone to catastrophic interference. This study analyses how gradient descent and overlapping representations between distant input points lead to distal interference and catastrophic interference. Distal interference refers to the phenomenon where training a model on a subset of the domain leads to non-local changes on other subsets of the domain. This study shows that uniformly trainable models without distal interference must be exponentially large. A novel antisymmetric bounded exponential layer B-spline ANN architecture named ABEL-Spline is proposed that can approximate any continuous function, is uniformly trainable, has polynomial computational complexity, and provides some guarantees for distal interference. Experiments are presented to demonstrate the theoretical properties of ABEL-Splines. ABEL-Splines are also evaluated on benchmark regression problems. It is concluded that the weaker distal interference guarantees in ABEL-Splines are insufficient for model-only continual learning. It is conjectured that continual learning with polynomial complexity models requires augmentation of the training data or algorithm.

Mia Gerber, Anna Sergeevna Bosman, Johan Pieter de Villiers

Automated machine learning (AutoML) is a research area focusing on using optimisation techniques to design machine learning (ML) algorithms, alleviating the need for a human to perform manual algorithm design. Real-time AutoML enables the design process to happen while the ML algorithm is being applied to a task. Real-time AutoML is an emerging research area, as such existing real-time AutoML techniques need improvement with respect to the quality of designs and time taken to create designs. To address these issues, this study proposes an Online Meta-learning for AutoML in Real-time (OnMAR) approach. Meta-learning gathers information about the optimisation process undertaken by the ML algorithm in the form of meta-features. Meta-features are used in conjunction with a meta-learner to optimise the optimisation process. The OnMAR approach uses a meta-learner to predict the accuracy of an ML design. If the accuracy predicted by the meta-learner is sufficient, the design is used, and if the predicted accuracy is low, an optimisation technique creates a new design. A genetic algorithm (GA) is the optimisation technique used as part of the OnMAR approach. Different meta-learners (k-nearest neighbours, random forest and XGBoost) are tested. The OnMAR approach is model-agnostic (i.e. not specific to a single real-time AutoML application) and therefore evaluated on three different real-time AutoML applications, namely: composing an image clustering algorithm, configuring the hyper-parameters of a convolutional neural network, and configuring a video classification pipeline. The OnMAR approach is effective, matching or outperforming existing real-time AutoML approaches, with the added benefit of a faster runtime.

Mamuku Mokuwe, Michael Burke, Anna Sergeevna Bosman

Saliency maps are often used in computer vision to provide intuitive interpretations of what input regions a model has used to produce a specific prediction. A number of approaches to saliency map generation are available, but most require access to model parameters. This work proposes an approach for saliency map generation for black-box models, where no access to model parameters is available, using a Bayesian optimisation sampling method. The approach aims to find the global salient image region responsible for a particular (black-box) model's prediction. This is achieved by a sampling-based approach to model perturbations that seeks to localise salient regions of an image to the black-box model. Results show that the proposed approach to saliency map generation outperforms grid-based perturbation approaches, and performs similarly to gradient-based approaches which require access to model parameters.

Anna Sergeevna Bosman, Andries Engelbrecht, Mardé Helbig

It has been argued in the past that high-dimensional neural networks do not exhibit local minima capable of trapping an optimisation algorithm. However, the relationship between loss surface modality and the neural architecture parameters, such as the number of hidden neurons per layer and the number of hidden layers, remains poorly understood. This study employs fitness landscape analysis to study the modality of neural network loss surfaces under various feed-forward architecture settings. An increase in the problem dimensionality is shown to yield a more searchable and more exploitable loss surface. An increase in the hidden layer width is shown to effectively reduce the number of local minima, and simplify the shape of the global attractor. An increase in the architecture depth is shown to sharpen the global attractor, thus making it more exploitable.

Anna Sergeevna Bosman, Andries Engelbrecht, Mardé Helbig

Quantification of the stationary points and the associated basins of attraction of neural network loss surfaces is an important step towards a better understanding of neural network loss surfaces at large. This work proposes a novel method to visualise basins of attraction together with the associated stationary points via gradient-based random sampling. The proposed technique is used to perform an empirical study of the loss surfaces generated by two different error metrics: quadratic loss and entropic loss. The empirical observations confirm the theoretical hypothesis regarding the nature of neural network attraction basins. Entropic loss is shown to exhibit stronger gradients and fewer stationary points than quadratic loss, indicating that entropic loss has a more searchable landscape. Quadratic loss is shown to be more resilient to overfitting than entropic loss. Both losses are shown to exhibit local minima, but the number of local minima is shown to decrease with an increase in dimensionality. Thus, the proposed visualisation technique successfully captures the local minima properties exhibited by the neural network loss surfaces, and can be used for the purpose of fitness landscape analysis of neural networks.

Gerard Jacques van Wyk, Anna Sergeevna Bosman

Convolutional neural network (CNN) architectures have traditionally been explored by human experts in a manual search process that is time-consuming and ineffectively explores the massive space of potential solutions. Neural architecture search (NAS) methods automatically search the space of neural network hyperparameters in order to find optimal task-specific architectures. NAS methods have discovered CNN architectures that achieve state-of-the-art performance in image classification among other tasks, however the application of NAS to image-to-image regression problems such as image restoration is sparse. This paper proposes a NAS method that performs computationally efficient evolutionary search of a minimally constrained network architecture search space. The performance of architectures discovered by the proposed method is evaluated on a variety of image restoration tasks applied to the ImageNet64x64 dataset, and compared with human-engineered CNN architectures. The best neural architectures discovered using only 2 GPU-hours of evolutionary search exhibit comparable performance to the human-engineered baseline architecture.

Kevin Louis de Jong, Anna Sergeevna Bosman

This paper proposes an efficient unsupervised method for detecting relevant changes between two temporally different images of the same scene. A convolutional neural network (CNN) for semantic segmentation is implemented to extract compressed image features, as well as to classify the detected changes into the correct semantic classes. A difference image is created using the feature map information generated by the CNN, without explicitly training on target difference images. Thus, the proposed change detection method is unsupervised, and can be performed using any CNN model pre-trained for semantic segmentation.