Faculty of Science (ETDs)

Permanent URI for this community

https://hdl.handle.net/10539/37933

Browse

Now showing 1 - 20 of 37

A fully-decentralised general-sum approach for multi-agent reinforcement learning using minimal modelling
(University of the Witwatersrand, Johannesburg, 2023-08) Kruger, Marcel Matthew Anthony; Rosman, Benjamin; James, Steven; Shipton, Jarrod
Multi-agent reinforcement learning is a prominent area of research in machine learning, extending reinforcement learning to scenarios where multiple agents concurrently learn and interact within the same environment. Most existing methods rely on centralisation during training, while others employ agent modelling. In contrast, we propose a novel method that adapts the role of entropy to assist in fully-decentralised training without explicitly modelling other agents using additional information to which most centralised methods assume access. We augment entropy to encourage more deterministic agents, and instead, we let the non-stationarity inherent in MARL serve as a mode for exploration. We empirically evaluate the performance of our method across five distinct environments, each representing unique challenges. Our assessment encompasses both cooperative and competitive cases. Our findings indicate that the approach of penalising entropy, rather than rewarding it, enables agents to perform at least as well as the prevailing standard of entropy maximisation. Moreover, our alternative approach achieves several of the original objectives of entropy regularisation in reinforcement learning, such as increased sample efficiency and potentially better final rewards. Whilst entropy has a significant role, our results in the competitive case indicate that position bias is still a considerable challenge.
An integrated approach for detecting and monitoring the Campuloclinium macrocephalum (Less) DC using the MaxEnt and machine learning models in the Cradle Nature Reserve, South Africa
(University of the Witwatersrand, Johannesburg, 2024) Makobe, Benjamin; Mhangara, Paidamoyo
The invasion of ecosystems by invasive plants is considered as one of the major human- induced global environmental change. The uncontrolled expansion of invasive alien plants is gaining international attention, and remote sensing technology is adopted to accurately detect and monitor the spread of invasive plants locally and globally. The Greater Cradle nature reserve is a world heritage site and intense research site for archaeology and paleontology.It was accorded the world status by the United Nations Educational, Scientific and Cultural Organizations (UNESCO) in 1991 due to its variety of biodiversity present and carries information of significance about the evolution of mankind. The invasion of Campuloclinium macrocephalum (pompom weed) at the Cradle nature reserve is downgrading the world status accorded to the site, lowers the grazing capacity for game animals and replaces the native vegetation. This research study explored the capability of Sentinel-2A multispectral imagery in mapping the spatial distribution of pompom weed at the nature reserve between 2019 and 2024. The non-parametric classification models, support vector machine (SVM) and random forests (RF) were evaluated to accurately detect, and discriminate pompom weed against the co-existing land cover types. Additionally, the species distribution modelling MaxEnt Entropy was incorporated to model spatial distribution and pompom weed habitat suitability. The findings indicates that SVM yielded 44% and 50.7% spatial coverage of pompom weed at the nature reserve in 2019 and 2024, respectively. Whereas, the RF model indicates that the spatial coverage of pompom weed was 31.1% and 39.3% in 2019 and 2024, respectively. The MaxEnt model identified both soil and rainfall as the most important environmental factors in fostering the aggressive proliferation of pompom weed at nature reserves. The MaxEnt predictive model obtained an area under curve score of 0.94, indicating outstanding prediction model performance. SVM and RF models had classification accuracy above 75%, indicating that they could distinguish pompom weeds from existing land cover types. The preliminary results of this study call for attention in using predictive models in predicting current and future spatial distribution of invasive weeds, for effective eradication control and environmental management.
Applying Machine Learning to Model South Africa’s Equity Market Index Price Performance
(University of the Witwatersrand, Johannesburg, 2023-07) Nokeri, Tshepo Chris; Mulaudzi, Rudzani; Ajoodha, Ritesh
Policymakers typically use statistical multivariate forecasting models to forecast the reaction of stock market returns to changing economic activities. However, these models frequently result in subpar performance due to inflexibility and incompetence in modeling non-linear relationships. Emerging research suggests that machine learning models can better handle data from non-linear dynamic systems and yield outstanding model performance. This research compared the performance of machine learning models to the performance of the benchmark model (the vector autoregressive model) when forecasting the reaction of stock market returns to changing economic activities in South Africa. The vector autoregressive model was used to forecast the reaction of stock market returns. It achieved a mean absolute percentage error (MAPE) value of 0.0084. Machine learning models were used to forecast the reaction of stock market returns. The lowest MAPE value was 0.0051. The machine learning model trained on low economic data dimensions performed 65% better than the benchmark model. Machine learning models also identified key economic activities when forecasting the reaction of stock market returns. Most research focused on whole features, few models for comparison, and barely focused on how different feature subsets and reduced dimensionality change model performance, a limitation this research addresses when considering the number of experiments. This research considered various experiments, i.e., different feature subsets and data dimensions, to determine whether machine learning models perform better than the benchmark model when forecasting the reaction of stock market returns to changing economic activities in South Africa.
Assessment of the Environmental Authorisation Processes and Mining Right Applications for Improved Environmental Outcomes
(University of the Witwatersrand, Johannesburg, 2023) Antoniades, Maria; Watson, Ingrid
This study investigates alignment of South African mining right and environmental authorisation application processes to determine their adequacy in catering for optimised early mine planning seeking to achieve enhanced environmental outcomes. First the legislative requirements for mining right and environmental authorisation applications are evaluated. Results are critically analysed, including an assessment of process alignments and disjunctions. Secondly, integration of the application study processes in practice are investigated. The practical implications of the application requirements are qualitatively examined through key informant and case study analysis. It is shown that integrated planning is not a legislated requirement nor readily adopted by proponents. Environmental planning conforms to technical outputs as tick-box exercises rather than being iterative and co-operative. Workstreams misalignments result in poor planning to the detriment of environmental outcomes. Finally, a practical guidance is presented for early integrated study processes aimed at meaningful project design through parallel planning to optimise environmental results.
Development of TileCoM firmware and software for the off-detector electronics of the ATLAS Tile Calorimeter at the HL-LHC
(University of the Witwatersrand, Johannesburg, 2023-08) Gololo, Mpho Gift Doctor; Argos, Fernando Carrio; Mellado, Bruce
In 2010 the LHC started to operate as the energy frontier particle accelerator in the world, situated close to Geneva and 100 m below the French and Swiss border in a circular tunnel of 27 km. The HL-LHC which is an upgrade of the LHC is envisioned to maximize the instantaneous luminosity of L = 1 × 1034 cm−2s −1 by a factor of 5 to fully exploit the physics potential at the energy frontier. During 10 years of operation, an improved TDAQ system architecture will have the capability to accommodate the trigger rates and the amount of data generated from the HL-LHC. TileCal is the ATLAS central hadronic calorimeter, a sampling calorimeter with iron as passive medium and plastic scintillator tiles as active medium. The ATLAS TileCal Phase-II upgrades will prepare the ATLAS experiment for the HL-LHC and includes new requirements in terms of radiation levels, an increase in data bandwidth, and clock distribution. To meet the requirements of the HL-LHC, a completely new readout electronics is designed to support the data acquisition system of TileCal. As part of the new readout electronics, this thesis is focused on the design of the TileCoM and Tile GbE Switch. The Tile GbE Switch PCB is manufactured by two South African companies, Trax Interconnect and Jemstech. The PCBs are fully func tional and have been integrated with new readout electronics. Three main function alities are implemented on the TileCoM in software and firmware implementation as key elements of the TDAQ and DCS of the ATLAS TileCal at the HL-LHC. The TileCoM and Tile GbE Switch are successfully integrated with the ATLAS Phase II TileCal upgrade electronics. This is achieved by successful remote control and monitoring of the ATLAS TileCal Phase-II upgrade electronics. This thesis shows monitoring results based on voltage, current and other parameters.
Diastereoselective conjugate addition reactions using diverse nucleophiles on a variety of Morita-Baylis-Hillman (MBH) adducts
(University of the Witwatersrand, Johannesburg, 2023-09) Bhom, Nafisa; Bode, Moira L.
The Morita-Baylis-Hillman (MBH) reaction involves the formation of a new carbon-carbon bond, generating an MBH adduct. These MBH adducts are multi-functional molecules, which can be used as synthons for the generation of complex and diverse compounds. The first part of the work described here involved the synthesis of a series of diverse ester and nitrile MBH adducts obtained as racemic mixtures. The MBH adducts were protected using different protecting groups, which could potentially control the diastereoselectivity and the formation of alternative products in the subsequent conjugate addition reaction. Conjugate addition reactions were performed on the protected MBH adducts using different nucleophiles to obtain the product as diastereomers. These reactions were monitored to detect whether diastereomers were obtained or not. The diastereomeric ratios obtained using different substrates, protecting groups and nucleophiles were determined. The best diastereomeric ratio was 3:1, obtained for the piperidine and benzylamine addition on the TBDMS protected nitrile adducts 192a/b and 196a/b. The addition of sulfur nucleophiles gave the conjugate addition product only and the addition of nitrogen nucleophiles gave both conjugate addition and allylic substitution products. It was found that the protecting groups did not have an effect on the diastereomeric ratio obtained, nor on the formation of alternative products. The last step performed in the sequence was the deprotection of the conjugate addition products. The configuration of the major and the minor diastereomers were determined, the major product was assigned as the syn diastereomer. The major:minor diastereomeric ratio for compound 208a/b was 3:1 and for compound 209a/b, a ratio of 2:1 was obtained. The next part of the work involved the synthesis of MBH adducts with amide as the electron withdrawing group. The originally proposed route involved the synthesis of MBH esters and their conversion into amides. The conjugate addition reactions were attempted on these amide adducts, but were unsuccessful. A number of alternative routes were attempted for the synthesis of amide adducts and conjugate addition products resulting from these adducts. From all the alternative routes, the best route was the originally proposed route.
Dissolution of non-functionalized and functionalized nanomaterials in simulated biological and environmental fluids
(University of the Witwatersrand, Johannesburg, 2023-06) Mbanga, Odwa; Gulumian, Mary; Cukrowska, Ewa
The incorporation of nanoparticles in consumer products is exponentially high, however, research into their behaviour in biological and environmental surroundings is still very limited. In the present study, the static system and the continuous flow-through dissolution protocols were utilized to evaluate and elucidate the dissolution behaviour of gold, silver, and titanium dioxide nanoparticles. The behaviour of these particles was studied in a range of artificial physiological fluids and environmental media, to obtain a more precise comprehension of how they would react in the human body and the environment. The biodurability and persistence were estimated by calculating the dissolution kinetics of the nanoparticles in artificial physiological fluids and environmental media. The details of the current research are described as follows: An investigation into the dissolution of non-functionalized and functionalized gold nanoparticles was conducted as the first component of the research, examining the effect of surface functionalization on dissolution. The study determined the dissolution rates of functionalized and non-functionalized gold nanoparticles. Dissolution was observed to be significantly higher in acidic media than in alkaline media. The nanoparticle surface modification, particle aggregation, and chemical composition of the simulated fluid significantly affected the dissolution rate. It was concluded that gold nanoparticles are biodurable and have the potential to cause long-term health effect as well as high environmental persistency. This work has been published in the Journal of Nanoparticle Research and is presented in this thesis as Paper 1. Silver nanoparticles were also included in this study because they have many applications and industrial purposes. Therefore, their risk assessment was also of utmost importance. The results indicated that silver nanoparticle solubility was influenced by the alkalinity and acidity of artificial media. Low pH values and high ionic strength encouraged silver nanoparticle dissolution and accelerated the dissolution rate. The agglomeration state and reactivity of the particles changed upon exposure to simulated fluids, though their shape remained the same. The fast dissolution rates in most fluids indicated that the release of silver ions would cause short-term effects. This work has been published in Toxicology Reports and has been presented in this thesis as Paper 2. Although titanium dioxide nanoparticles are insoluble and undergo negligible dissolution, it was of utmost importance to investigate their behaviour in biological and environmental surroundings. This is as a result of the incorporation of these particles in everyday consumer products, in the nanosized range which raises concerns about their safety. Therefore, in Paper 3 presented in this thesis the dissolution kinetics of titanium dioxide nanoparticles in simulated body fluids representative of the lungs, stomach, blood plasma and media representing the aquatic ecosystem were investigated to anticipate how they behave in vivo. This work has been published in Toxicology In Vitro and presented in this thesis as Paper 3. The results indicated that titanium dioxide nanoparticles were very insoluble, and their dissolution was limited in all simulated fluids. Acidic media such as the synthetic stomach fluids were most successful in dissolving the particles, while alkaline media had lower dissolution. High ionic strength seawater also had a higher dissolution rate than freshwater. The dissolution rates of the particles were low, and their half-times were long. The results indicated that these particles could potentially cause health issues in the long term, as well as remain unchanged in the environment. This work has been published in Toxicology In Vitro and presented in this thesis as Paper 3. The last component of the research compared the dissolution kinetics of gold, silver and titanium dioxide nanoparticles through the use of the continuous flow-through system. The findings indicated that titanium dioxide nanoparticles were the most biodurable and persistent, followed by gold and silver nanoparticles. Therefore, it was suggested that product developers should use the OECD's guidelines for testing before releasing their product to the market to ensure its safety. This work has been published in Nanomaterials MDPI and presented in this thesis as Paper 4.
Distance measures, independence number and chromatic number
(University of the Witwatersrand, Johannesburg, 2023-03) Moholane, Letlhogonolo; Jonck, Betsie; Mukwembi, Simon
There are numerous parameters in graph theory. In this dissertation, we pay a special attention to average distance, independence number, average eccentricity, order and the chromatic number of a graph. In 1975, Doyle and Graver proved an upper bound on the average distance with respect to the order of the graph. This gave rise to studies that focus on upper and lower bounds on average distance in terms of other graph parameters. Approximately, three decades after Doyle and Graver proved their result, Dankelmann, Goddard, and Swart in 2004 produced a study that gave an upper bound on average eccentricity in terms of minimum degree and order of the graph, initiating studies that focus on giving bounds on average eccentricity with respect to other known graph parameters. In this dissertation, we investigate bounds on average eccentricity and on average distance. We give upper bounds on average eccentricity in terms of independence number of the graph and order of the graph. Then, we present bounds on average eccentricity when order and chromatic number of the graph are prescribed. The second part of the dissertation is dedicated to presenting upper bounds on average distance with respect to independence number and order of the graph, and again, in terms of chromatic number and order of the graph.
Effects and consequences of natural and artificial light at night on small mammals in peri-urban Johannesburg, South Africa
(University of the Witwatersrand, Johannesburg, 2024) Oosthuizen, Tasha; Pillay, Neville; Oosthuizen, Marietjie
Studies investigating artificial light at night (ALAN) have increased over recent years. However, research examining the influence of ALAN on southern African small mammal species are lacking and even information on their basic biology is scarce. To close this knowledge gap, I investigated the effect of ALAN on different facets of animal behaviour in African small mammals. Firstly, I evaluated the impact of the natural (lunar cycle) and ALAN on the community composition and species abundance in two populations of small mammals. I chose two field sites: one facing Johannesburg (exposed to ALAN) and one facing away. I conducted mark-recapture trapping to ascertain the occurrence and abundance of small mammals. The Light site had both a higher species composition and a higher animal abundance when compared to the composition and abundance of the Dark site. The lunar cycle had an effect; on nights with a full moon, the species composition and animal abundance of both study sites declined, while on new moon nights, the opposite occurred, with an increase in both the species composition and abundance on the Light and Dark sites. The absence of a negative ALAN effect on the Light site can potentially be ascribed to the availability of microhabitats for small mammals to escape illumination, leaving them seemingly unaffected. Next, I assessed the locomotor activity of three species of commonly occurring rodents on the study area, one crepuscular (19 single-striped grass mice, Lemniscomys rosalia), one species with reportedly variable activity (19 angoni vlei rats, Otomys angoniensis) and one nocturnal (19 southern multimammate mice, Mastomys coucha). They were captured at a different location than the mark-recapture study sites and tested in captivity under natural (exposed to natural light and temperature changes), laboratory (standard laboratory conditions; 12h light:12h dark and constant temperature) and ALAN treatments. Lemniscomys rosalia exhibited crepuscular activity under all three experimental treatments, Otomys angoniensis was mostly nocturnal with some diurnal activity. The temporal activity profiles of the two species that showed some activity during the light hours were unaffected by ALAN. Mastomys coucha displayed strictly nocturnal activity during the natural and laboratory treatments, but during ALAN treatments the temporal activity profiles of some animals shifted so that they were active during the start of the day. Lemniscomys rosalia and O. angoniensis were more active under the natural treatment, whilst M. coucha was more active in the laboratory treatment. When exposed to 2 Lux ALAN presented remotely, there was no effect on the level of activity in O. angoniensis, L. rosalia showed a reduction of about 20% in its activity, whereas M. coucha reduced its activity by more than 50%. Finally, I studied how ALAN impacted the foraging behaviour of the three species under four treatments (during the day, at night, 2 Lux ALAN and 10 Lux ALAN). Foraging behaviour differed in the three species under different light conditions. Lemniscomys rosalia was risk-averse when feeding during the diurnal and nocturnal (no light at night) treatments. Otomys angoniensis showed irregular responses in their foraging behaviour under all foraging treatments. Mastomys coucha showed no differences when feeding under any of the nocturnal treatments, but it was inactive under the diurnal treatment. Overall, my study revealed that the effect of ALAN is not similar for all small mammalian species and appear to depend on both the spatial and temporal niches that the different species occupy. Strictly nocturnal animals seem to be affected the most, whereas animals that are active during the day showed lesser responses. Given the rapid increase in urbanisation and anthropogenic disturbances, more and more species are exposed to ALAN. Species that prefer darker, more secluded habitats appear to be more vulnerable and at higher risk of local extinctions as a result of disturbances, such as ALAN and habitat transformation. My study highlights that ALAN affects both nocturnal and diurnal rodents to the extent that it can have fitness consequences, including changed active times, foraging efficiency, movement patterns and susceptibility to predation. Finally, the disruption of rodent behaviour can have cascading effects for ecosystems and my study also emphasises the importance of safeguarding our night skies to protect biodiversity.
Evaluation of the JSE’s environmental reporting requirements of South African listed companies
(University of the Witwatersrand, Johannesburg, 2024) Hariram, Viratha; Schwaibold, Ute
Through its mandatory environmental reporting requirements, the Johannesburg Stock Exchange (JSE) plays a pivotal role in the private sector to align to Environmental, Social and Governance frameworks and disclose information of a company’s environmental priorities and performance in addressing areas of concern. While there are attempts to safeguard the environment from damage and degradation, it is unclear if this framework is suitable at appropriately addressing the environmental areas of concern facing South Africa. In order to evaluate the JSE’s environmental requirements on listed companies, this study aimed to identify the local and global environmental priorities via the South African State of Environment Outlook Report (local view), National Development Plan (local view) and Sustainable Development Goals (global view) and thereafter determine if the Global Reporting Initiative, the only set of mandatory environmental requirements of the JSE, was adequately addressing the indicators they outlined. Furthermore, using a scoring system from zero to four, the study evaluated the alignment of ten randomly selected JSE listed companies against the requirements of the Global Reporting Initiative to note their level of adherence and alignment to the South African State of Environment Outlook Report, National Development Plan and Sustainable Development Goals. The results indicate that the Global Reporting Initiative (GRI) covers 73% of environmental priorities discussed in the South African State of Environment Outlook Report, National Development Plan and Sustainable Development Goals. Of the total 74 indicators of environmental concern identified from the three reports, the GRI did not require disclosure for 20 indicators. Only three indicators that were required to be disclosed by the GRI had scored a four on the rating scale due to it being an integrated disclosure that takes into account one or more other related environmental categories. The evaluation of the sampled listed company’s adherence to the GRI via their sustainability reports and integrated annual reports indicated that none of the companies had provided sufficient disclosures to meet the requirements of the GRI. 80% of the sampled listed companies made a strong alignment to the Sustainable Development Goals in their sustainability reports and / or integrated annual reports. A shortcoming of the JSE that was stated by all four sustainability experts interviewed was the lack of enforcement for listed companies to make quality and comprehensive disclosures or accountability expected from listed companies. There is no formal process of review or consequences for listed companies who do not comply.
Farming systems in South Africa beyond 2020: a scenario-based study, using systems analysis, of the connectivity between farming systems in the Vhembe district, Limpopo, South Africa
(University of the Witwatersrand, Johannesburg, 2024) Materechera-Mitochi, Fenji; Scholes, Mary
Agriculture is a significant contributor to the South African economy and overall development as it contributes to poverty reduction and food security. It is against this backdrop that agricultural development becomes a focus area for decision making amongst stakeholders, as it is directly linked to food systems. The traditional approach to agricultural production in the country has been one that views farming as mainly based on land ownership and yield in isolation from the broader context of the four drivers of production namely land, labour, capital and enterprise. The concept of farming systems provides a broader perspective on farming and encompasses the entire value chain for a commodity which includes production, management practices, marketing, value addition, financial resources, and policies. The South African agrarian structure is characterised by a dualism in which large-scale commercial farmers co-exist alongside small-scale farmers. This is a legacy of the apartheid system of governance. Large-scale commercial farmers, who are mostly capital intensive, have historically been regarded as the main drivers of national food security while small-scale farmers on the other hand are viewed as significant contributors to food security at a household level. Both farmers are therefore important contributors to the national agricultural economy. Research on the two types of farmers in the South African context is usually focused on the respective farmers’ approaches to production individually and does not consider them as joint ventures. This study was aimed at providing an alternative approach to viewing South Africa’s farming systems by evaluating current farming systems in the Vhembe district of Limpopo, South Africa, using systems analysis as a tool to highlight the connectivity of the interactions within and between them. The study also aimed to conceptualize scenarios for sustainable future farming systems in South Africa. The Vhembe district in the Limpopo province was chosen for the study because both largescale commercial and small-scale farmers occur and due to the favourable sub-tropical climate, the area has become a hub for the farming of numerous high value crops that contribute positively to the country’s agricultural economy. The study made use of a mixed methods approach that combined the analysis of primary data obtained from in-depth interviews and secondary data obtained from an agricultural database to identify and characterize large-scale commercial and small-scale farming systems in the Vhembe district. The study examined the drivers of production for three different commodities, macadamia nuts, mangos and avocado iii pears, the two types of farming systems and their connectivity. The study was grounded on the conceptual framework of systems thinking and used a systems analysis tool i.e., causal loop diagrams to analyse the connectivity between the two farming systems. Lastly, the study developed conceptual scenarios using a deductive scenario method to conceptualise scenarios for the future of the two farming systems and the different commodities. Key findings of the study showed that farming systems need to be understood through the lens of the four drivers of production. Land as a driver of production interacts with multiple other factors in shaping the management of a sustainable farming system. Examples of these factors include the link between land availability, ownership and farm size, decision-making and resource allocation tied to land management practices, and socio-economic considerations including the diversification of livelihoods by incorporating non-farm income and the farmers’ adaptability to uncertainties such as climate change. The findings also revealed that there are interconnections between the two types of farming systems presenting potential for enhanced production and commercial opportunities. The conceptual scenarios developed in the study and the systems thinking tool of causal loop diagrams proved to be valuable tools to inform decision making and policy development. The study’s main conclusion points to the potential of large-scale commercial and small-scale farming systems in South Africa operating as joint ventures in the future and enhancing the sustainability of agricultural production and livelihoods. It also recommends the use of systems thinking that includes social, financial and environmental values and impacts in decision making for agricultural development.
Generating Rich Image Descriptions from Localized Attention
(University of the Witwatersrand, Johannesburg, 2023-08) Poulton, David; Klein, Richard
The field of image captioning is constantly growing with swathes of new methodologies, performance leaps, datasets, and challenges. One new challenge is the task of long-text image description. While the vast majority of research has focused on short captions for images with only short phrases or sentences, new research and the recently released Localized Narratives dataset have pushed this to rich, paragraph length descriptions. In this work we perform additional research to grow the sub-field of long-text image descriptions and determine the viability of our new methods. We experiment with a variety of progressively more complex LSTM and Transformer-based approaches, utilising human-generated localised attention traces and image data to generate suitable captions, and evaluate these methods on a suite of common language evaluation metrics. We find that LSTM-based approaches are not well suited to the task, and under-perform Transformer-based implementations on our metric suite while also proving substantially more demanding to train. On the other hand, we find that our Transformer-based methods are well capable of generating captions with rich focus over all regions of the image and in a grammatically sound manner, with our most complex model outperforming existing approaches on our metric suite.
Generative Model Based Adversarial Defenses for Deepfake Detectors
(University of the Witwatersrand, Johannesburg, 2023-08) Kavilan Dhavan, Nair; Klein, Richard
Deepfake videos present a serious threat to society as they can be used to spread mis-information through social media. Convolutional Neural Networks (CNNs) have been effective in detecting deepfake videos, but they are vulnerable to adversarial attacks that can compromise their accuracy. This vulnerability can be exploited by deepfake creators to evade detection. In this study, we evaluate the effectiveness of two genera- tive adversarial defense mechanisms, APE-GAN and MagNet, in the context of deepfake detection. We use the FaceForensics++ dataset and a CNN victim model based on the XceptionNet architecture, which we attack using the iterative fast gradient sign method at two different levels of ✏, ✏ = 0.0001 and ✏ = 0.01. We find that both APE-GAN and MagNet can purify the adversarial images and restore the performance of the vic- tim model to within 10% of the model’s accuracy on benign fake inputs. However, these methods were less effective at restoring accuracy for adversarial real examples and were not able to significantly restore accuracy when the adversarial attack was aggressive (✏ = 0.01). We recommend that an adversarial defense method be used in conjunction with a deepfake detector to improve the accuracy of predictions. APE-GAN and MagNet are effective methods in the deepfake context, but their effectiveness is limited when the adversarial attack is aggressive.
Hydrogeological assessments and investigation of inflow sources at Lumwana Copper Mine, Zambia
(University of the Witwatersrand, Johannesburg, 2023) Mbilima, Mike; Abiye, Tamiru
This Research Report presents results of integrated field and desktop-based hydrogeological investigations at the Lumwana Mine, Zambia. Groundwater occurrence in the mine poses challenges with effective mining operations and slope stability. The primary aim of this study was to establish the sources of groundwater inflows and to establish the nature of surface water and groundwater interaction within the Lumwana Mine hydro-geotechnical units. The Lumwana hydrogeological investigation has been achieved through the integration of multi-disciplinary data types, which include geology, structures, hydrochemistry, meteorological data (rainfall, temperature, humidity and evapotranspiration), environmental isotopes, dewatering pumping records, groundwater level monitoring, water temperature, general hydrogeological data and surface hydrology. The investigation has confirmed the presence of hydraulic connections between different surface water bodies such as dams, diversion channels, streams and open pit excavation, and has proven to be a useful approach in tracing the source of mine inflows. Rainfall, groundwater and surface water samples have similar δ18O and δ2H isotopic signatures thus lamenting the existence of a hydraulic link between groundwater and surface water. Recharge estimation through Water Table Fluctuation method (WTF) determined 8% of mean annual precipitation (MAP). The dominant hydrochemical facies are Ca-Mg-HCO3 and Ca-Mg-SO4. The local geology and geochemistry of the tailings are the main controllers of groundwater chemistry through rock-water interaction. The geology of the study area consists of older metamorphosed gneisses, schists, migmatites, amphibolites and granitoids. Integrated assessment of the Lumwana hydrogeological environment has enabled the development of the Lumwana Mine hydrogeological conceptual model. In the shallow, highly to moderately weathered zones, groundwater flows from south towards low topographic regions in the northwest mimicking the general topography. The hydraulic test conducted at Lumwana Mine has revealed the saprock units have higher hydraulic conductivity by several orders compared to the saprolites and the fresh bedrock, where groundwater flow is mainly controlled by the occurrence and distribution of the fracture network.
Improving audio-driven visual dubbing solutions using self-supervised generative adversarial networks
(University of the Witwatersrand, Johannesburg, 2023-09) Ranchod, Mayur; Klein, Richard
Audio-driven visual dubbing (ADVD) is the process of accepting a talking-face video, along with a dubbing audio segment, as inputs and producing a dubbed video such that the speaker appears to be uttering the dubbing audio. ADVD aims to address the language barrier inherent in the consumption of video-based content caused by the various languages in which videos may be presented. Specifically, a video may only be consumed by the audience that is familiar with the spoken language. Traditional solutions, such as subtitles and audio-dubbing, hinder the viewer’s experience by either obstructing the on-screen content or introducing an unpleasant discrepancy between the speaker’s mouth movements and the input dubbing audio, respectively. In contrast, ADVD strives to achieve a natural viewing experience by synchronizing the speaker’s mouth movements with the dubbing audio. A comprehensive survey of several ADVD solutions revealed that most existing solutions achieve satisfactory visual quality and lip-sync accuracy but are limited to low-resolution videos with frontal or near frontal faces. Since this is in sharp contrast to real-world videos, which are high-resolution and contain arbitrary head poses, we present one of the first ADVD solutions trained with high-resolution data and also introduce the first pose-invariant ADVD solution. Our results show that the presented solution achieves superior visual quality while also achieving high measures of lip-sync accuracy, consequently enabling the solution to achieve significantly improved results when applied to real-world videos.
Improving Semi-Supervised Learning Generative Adversarial Networks
(University of the Witwatersrand, Johannesburg, 2023-08) Moolla, Faheem; Bau, Hairong; Van Zyl, Terence
Generative Adversarial Networks (GANs) have shown remarkable potential in generating high-quality images, with semi-supervised GANs providing a high classification accuracy. In this study, an enhanced semi supervised GAN model is proposed wherein the generator of the GAN is replaced by a pre-trained decoder from a Variational Autoencoder. The model presented outperforms regular GAN and semi-supervised GAN models during the early stages of training, as it produces higher quality images. Our model demonstrated significant improvements in image quality across three datasets - namely the MNIST, Fashion MNIST, and CIFAR-10 datasets - as evidenced by higher accuracies obtained from a Convolutional Neural Network (CNN) trained on generated images, as well as superior inception scores. Additionally, our model prevented mode collapse and exhibited smaller oscillations in the discriminator and generator loss graphs compared to baseline models. The presented model also provided remarkably high levels of classification accuracy, by obtaining 99.32% on the MNIST dataset, 92.78% on the Fashion MNIST dataset, and 83.22% on the CIFAR-10 dataset. These scores are notably robust as they improved some of the classification accuracies obtained by two state-of-the-art models, indicating that the presented model is a significantly improved semi-supervised GAN model. However, despite the high classification accuracy for the CIFAR-10 dataset, a considerable drop in accuracy was observed when comparing generated images to real images for this dataset. This suggests that the quality of those generated images can be bettered and the presented model performs better with less complex datasets. Future work could explore techniques to enhance our model’s performance with more intricate datasets, ultimately expanding its applicability across various domains.
Imputation of missing values and the application of transfer machine learning to predict water quality in acid mine drainage treatment plants
(University of the Witwatersrand, Johannesburg, 2024) Hasrod, Taskeen
Access to clean water is one of the most difficult challenges of the 21st century. Natural unpolluted water bodies are becoming one of the most dramatically declining resources due to environmental pollution. In countries like South Africa which has a mining-centred economy, toxic pollution from mine tailing dumps and unused mines leach into the underground water table and contaminate it. This is known as Acid Mine Drainage (AMD) and poses a grave threat to humans, animals and the environment due to its toxic element and acidic content. It is, therefore, imperative that sustainable wastewater treatment procedures be put in place in order to decrease the toxicity of the AMD such that clean water may be recovered. An efficient circular economy is created in the process since original wastewater can be recycled to not only provide clean water, but also valuable byproducts such as sulphur (from the elevate sulphate content) and other important minerals. Traditional analytical chemistry methods used to measure sulphate are usually time-consuming, expensive and inefficient, thereby, leading to incomplete analytical results being reported. To address this, this study aimed at imputing missing values for sulphate concentrations in one AMD treatment plant dataset and then using that to conduct transfer learning to predict concentrations in two other AMD treatment plants datasets. The approach involved using historical water data and applying geochemical modelling as a thermodynamical tool to assess the water chemistry and conduct preliminary data cleaning. Based on this, Machine Learning (ML) was then used to predict the sulphate concentrations, thus, addressing limited data on this parameter in the datasets. With complete and accurate sulphate concentrations, it is possible to conduct further modelling and experimental work aimed at recovering important minerals such as octathiocane, S8 (a commercial form of sulphur), gypsum and metals. Historical data obtained from the three AMD treatment plants in Johannesburg, South Africa (viz., Central Rand, East Rand and West Rand) were obtained and the larger Central Rand dataset was split into smaller untreated AMD (Pump A and Pump B) subsets. Thermodynamic and solution equilibria aspects of the water were assessed using the PHREEQC geochemical modelling code. This served as a preliminary data cleanup step. Eight baseline as well as three ensemble machine learning regression models were trained on the Central Rand subsets and compared to each other to find the best performing model that was then used to conduct Transfer Learning (TL) onto the East Rand and West Rand datasets to predict their sulphate levels. The findings pointed to a high correlation of sulphate to temperature (°C), Total Dissolved Solids (mg/L) and most importantly, iron (mg/L). The linear correlation between iron and sulphate substantiated pyrite (FeS2) as their source following weathering. Water quality parameters were found to be dependent on factors such as weather and geography this was evident in the treated water that had quite different chemistry to that of the untreated AMD. Neutralisation agents used were based on those parameters, thus, further delineating the chemistry of the treated and untreated water. The best performing ML model was the Stacking Ensemble (SE) regressor trained on Pump B’s data and combined the best performing models namely, Linear Regressor (LR), Ridge Regressor (RD), K-Nearest Neighbours Regressor (KNNR), Decision Tree Regressor (DT), Extreme Gradient Boosting Regressor (XG), Random Forest Regressor (RF) and Multi-Layer Perceptron Artificial Neural Network Regressor (MLP) as the level 0 models and LR as the level 1 model. Level 0 consisted of training heterogenous base models to obtain the crucial features from the dataset. These individual predictions and features were then fed to a single meta-learner model in in the next layer (level 1) to generate a final prediction. The stacking ensemble model performed well and achieved Mean Squared Error (MSE) of 0.000011, Mean Absolute Error (MAE) of 0.002617 and R2 of 0.999737 in under 2 minutes. This model was selected to be used for TL to the East Rand and West Rand datasets. Ensemble methods (bagging, boosting and stacking) outperformed individual baseline models. However, when comparing stacking ensemble ML that combined all the baseline models with stacking ensemble ML that only combined the best performing models, it was found that there was no significant improvement in excluding bad models from the stack as long as the good models were included. In one case, it was actually beneficial to include the bad performing models. All models were trained in under 2 minutes which proved the benefit of using ML approaches compared to traditional approaches. The treated water data was highly uncorrelated such that model training was unsuccessful with the highest achievable R2 value being 0.14, thus, no treated water model was available for TL. TL was successfully conducted on the cleaned and modelled East Rand AMD dataset using the Central Rand (Pump B) stacking regressor and a high level of accuracy with respect to Mean Square Error (MSE), Mean Absolute Error (MAE) and R2 (MSE:0.00124, MAE:0.0290 and R2:0.963) between the predicted and true sulphate values was achieved. This was achieved despite a marked difference in the distributions between the Central Rand and East Rand datasets which further proved the power of utilizing ML for water data. TL was successful in imputing missing values in the West Rand dataset following prediction of sulphate levels in the cleaned and modelled West Rand AMD and treated water datasets. No true values for sulphate levels in the West Rand dataset were given, as such, accuracy comparisons could not be made. However, a general baseline idea of the amount of sulphate present in the West Rand treatment plant could now be understood. The sulphate levels in all three treatment plants (Central Rand, East Rand and West Rand) were found to greatly differ from each other with the Central Rand having the most normal distribution, the East Rand having the most precise distribution and the West Rand having the most variable distribution. Whilst the sulphate levels in the treated effluent waters could not be reliably predicted due to inherent issues (e.g., analytical inaccuracies and inconsistences) and poor correlations within the treated water datasets, sulphate levels in all three of the untreated AMD datasets were successfully predicted with a high degree of accuracy. This underpinned the observation made previously about the discrepancies between treated and untreated water. The study has shown that it is possible to impute missing values in one water dataset and use transfer learning to complete and consolidate another similar, but scarce, dataset(s). This approach has been lacking in the water industry, resulting in the reliance and use of traditional methods that are expensive and inadequate. This has caused water practitioners to abandon scarce datasets, thus, losing potentially valuable information that could be useful for water remediation and recovery of valuable resources from the water. As a spin off from the study, it has been indicated that automation of such data analysis is possible. This was achieved by developing a Graphical User Interface (GUI) for ease of use of the SE-ML model by those with little to no programming background nor ML knowledge e.g., the laboratory staff at the AMD treatment plants. This can also be used for teaching purposesin academia.
Innovative surface, tunnel, and in-pit geophysical methods for mineral exploration and mine planning: case studies from the Bushveld Complex mines, South Africa
(University of the Witwatersrand, Johannesburg, 2023) Rapetsoa, Moyagabo Kenneth
Innovative geophysical methods were used to study the platinum group element mineralisation and their associated geological structures at Maseve and Tharisa mines, western Bushveld Complex. Four case studies are presented in this thesis that incorporate the use of in-mine or near mine geophysical methods for mineral exploration. The first one being in-mine seismic data acquired in 2020 at Maseve mine using cost-effective seismic source and sensors, followed by innovative seismic experiments acquired in 2022 at Maseve mine to evaluate the viability of using tunnel and surface experiments for mineral exploration in a noisy, logistically difficult mine environment. Thirdly, the 2021 integrated geophysical surveys conducted at Tharisa mine to image fractures that act as water pathways into the pit. Finally, integrated geophysical techniques are used to delineate boulders to enhance future mine planning and designs at Tharisa mine. The acquired geophysical data were processed using modern processing algorithms to enhance the target mineralization and complex geological structures in all the sites. In-mine reflection seismic datasets acquired in 2020 at Maseve mine proved useful as they provided optimum imaging of the economic Platinum Group Elements (PGEs) such as the Merensky Reef and Upper Group 2 chromitite layers (known as reefs). This is one of the few in-mine seismic experiments to have been conducted in South Africa for mineral exploration. In 2022, 2D reflection seismic profiles were acquired on surface above the Merensky Reef and Upper Group 2 chromitite, together with four 2D reflection seismic profiles acquired along the mine tunnel at ~ 550 m below the surface and tens of meters above known mineralisation: Merensky Reef and Upper Group 2 chromitite layer. Interpretation of the in-mine and surface seismic data were complemented by the use of 3D ray tracing numerical simulations to understand the distribution and out-of-plane reflectivity from the target mineralization. The 2022 Maseve reflection seismic data improved the imaging of geological structures and mineral deposits. The geophysical data acquired in 2021 at Tharisa mine demonstrated the importance of using near-surface integrated geophysical methods (magnetics, seismics, and electrical resistivity) with other datasets such as borehole logs and physical property measurements to understand the geophysical response of the mineral deposits. Ground magnetic data delineated a major dyke that was identified on the aeromagnetic data and geological mapping. Electrical resistivity tomography, on the other hand, identified linear low resistivity zones that differentiateiii fractured and undisturbed hard rock. Seismic methods were important for depth to bedrock imaging. Integration of geophysical methods was encouraged by the need to understand geological structures (e.g., faults, dykes, iron-rich ultramafic pegmatites, boulders) that can have impact on the efficiency, safety and costs of mining in South Africa. Moreover, this approach encourages the implementation of innovative geophysical surveys in brownfield sites for better mine design and planning, and to increase a life of mine (LoM)
Learning to adapt: domain adaptation with cycle-consistent generative adversarial networks
(University of the Witwatersrand, Johannesburg, 2023) Burke, Pierce William; Klein, Richard
Domain adaptation is a critical part of modern-day machine learning as many practitioners do not have the means to collect and label all the data they require reliably. Instead, they often turn to large online datasets to meet their data needs. However, this can often lead to a mismatch between the online dataset and the data they will encounter in their own problem. This is known as domain shift and plagues many different avenues of machine learning. From differences in data sources, changes in the underlying processes generating the data, or new unseen environments the models have yet to encounter. All these issues can lead to performance degradation. From the success in using Cycle-consistent Generative Adversarial Networks(CycleGAN) to learn unpaired image-to-image mappings, we propose a new method to help alleviate the issues caused by domain shifts in images. The proposed model incorporates an adversarial loss to encourage realistic-looking images in the target domain, a cycle-consistency loss to learn an unpaired image-to-image mapping, and a semantic loss from a task network to improve the generator’s performance. The task network is con-currently trained with the generators on the generated images to improve downstream task performance on adapted images. By utilizing the power of CycleGAN, we can learn to classify images in the target domain without any target domain labels. In this research, we show that our model is successful on various unsupervised domain adaptation (UDA) datasets and can alleviate domain shifts for different adaptation tasks, like classification or semantic segmentation. In our experiments on standard classification, we were able to bring the models performance to near oracle level accuracy on a variety of different classification datasets. The semantic segmentation experiments showed that our model could improve the performance on the target domain, but there is still room for further improvements. We also further analyze where our model performs well and where improvements can be made.
Leveraging Machine Learning in the Search for New Bosons at the LHC and Other Resulting Applications
(University of the Witwatersrand, Johannesburg, 2023-09) Stevenson, Finn David; Mellado, Bruce
This dissertation focuses on the use of semi-supervised machine learning for data generation in high-energy physics, specifically to aid in the search for new bosons at the Large Hadron Collider. The overarching physics analysis for this work involves the development of a generative machine learning model to assist in the search for resonances in the Zγ final state background data. A number of Variational Auto-encoder (VAE) derivatives are developed and trained to be able to generate a chosen Monte Carlo fast simulated dataset. These VAE derivatives are then evaluated using chosen metrics and plots to assess their performance in data generation. Overall, this work aims to demonstrate the utility of semi-supervised machine learning techniques in the search for new resonances in high-energy physics. Additionally, a resulting application of the use of machine learning in COVID-19 crisis management was also documented.

Browse

Browsing Faculty of Science (ETDs) by SDG "SDG-9: Industry, innovation and infrastructure"

Results Per Page

Sort Options