TRex22/MakeAISpeedy

AIExpo Africa 2022 - Poster Presentation # A Tool-kit Of Techniques To Make Computer Vision Models Easier To Train And Evaluate AIExpo Africa 2022 - Poster Presentation Poster at the event ## About me I am a senior software engineer at a Fin-tech called DocFox and a master’s candidate at the University of the Witwatersrand focusing on computer vision, explainable AI and self-driving cars. My blog: jasonchalom.com Email: contact@jasonchalom.com ## Thanks - My girlfriend :) - My master's supervisor: Dr Richard Klein - The Mathematical Sciences Support unit at the University of the Witwatersrand. - PRIME Lab (Located at the University of the Witwatersrand, specialising in Computer Vision) - Rail lab (raillab.org) # Abstract Machine learning techniques, especially in the area of computer vision, have opened a new frontier of research and industry - however the initial cost of entry into the field is a boundary, especially in developing economies. There are many techniques and approaches to optimising these work-flows in order to make them more computationally efficient, and therefore cheaper. Such techniques include traditional software engineering, data and signals processing, as well as more advanced optimisations to the model structures and training pipelines to make the process more efficient. These techniques have trade-offs and are task dependant but when combined appropriately can be a powerful tool-kit to improve computability of large work-flows. This poster evaluates a selection of techniques aimed at making computer vision easier to compute in memory usage, and time. # Introduction Artificial intelligence (AI) is a fundamental part of fourth industrial revolution (4IR). There is a lot of opportunity for innovation. However the cost, and access to AI - especially in computer vision - is a limiting factor in developing economies. AI on the edge is also limited by the practicality of the AI being used and the constraints of those systems. Many AI projects and state of …