BY Prof. Alan F. Smeaton
Computer vision and multimedia information processing have made extreme progress within the last decade and many tasks in these domains can be done with the same level of accuracy as if done by humans, or better. This is because we can leverage the benefits of having huge amounts of data available for training, we have an enormous increase in the amount of computer processing and emergence through using GPUs in parallel, and we have seen the evolution of machine learning as a suite of techniques to combine data and processing and deliver accurate vision-based systems.
While issues of bias, explainability, compute cost and carbon footprint in training each remain as hurdles, there is now such momentum in the neural networks approach to visual information processing that these hurdles will be overcome in time.
So now that we can process some forms of visual information better than humans what do we use this processing for ? Sometimes we use for smartphone and desktop image searching which we were not able to do before, but how many of us actually do that ? We can use this processing in autonomous vehicle navigation but that is a narrow and niche application as are security applications, searching CCTV for example, and healthcare diagnostics. If we examine the landscape of our research area we find that the greatest use of what we have developed – training data and neural networks and machine learning – is not in computer vision but elsewhere. There are now many other application areas that have jumped onto machine learning and call it “AI”. Some of these are a mis-match for neural information processing, many of them over-reach and over-claim what they can do, and only a small number are genuinely successful.
The history of technology development is littered with examples where innovation in one area trickles into secondary areas and benefits the secondary more than the primary. Sometimes it is planned like that, like motor racing, and sometimes it is accidental. While we can congratulate ourselves on our collective achievements in vision information processing over the last decade we have not yet found truly compelling applications for our own techniques in our own field. In this presentation I will survey the landscape of our vision applications, highlighting how we have not yet applied our own techniques to our most compelling needs, supporting our own personal memory.
Prof. Alan F. Smeaton
Professor, Founding Director
Alan F. Smeaton completed his PhD at University College Dublin and moved to Dublin City University where he has been a Full Professor of Computing since 1997. He has been Executive Dean of Faculty and Head of the School of Computing (twice). His research output includes more than 700 publications with more than 18,800 citations and a h-index of 68. He is an elected member of the Royal Irish Academy and an Academy Gold Medal Winner, a Fellow of the IEEE and winner of several awards and prizes including the Mark Everingham prize, the Niwa-Takayanagi prize and the Strix award. He is a founding Director of the Insight Centre for Data Analytics which has 400 researchers across 4 Universities in Ireland and is the largest non-capital public research award ever given in Ireland. In 2018 he has stepped back from his executive role in directing this in order to concentrate on teaching and research. He is a regular contributor to Irish and international media with several media “pieces” each year on radio, TV, news and online, in Ireland and internationally and on topics usually around Artificial Intelligence.
This web page was designed with Mobirise