select search filters
briefings
roundups & rapid reactions
before the headlines
Fiona fox's blog
Scientists comment on the Nobel Prize in Physics, awarded for discoveries and inventions in machine learning.
Dr Deepak Padmanabhan, Senior Lecturer, School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast (QUB), said:
“Think of humans and their behaviour. When we feel cold, we would like to turn on the heater. With each degree of temperature rise, we feel better. Until a point. After that, it becomes too hot, and we start wanting the temperature to be lower. This is a simple example of non-linearity in human life. Things don’t always move in the same direction, after a while, things start to reverse. Think of quality of life and earnings. If one is too poor, each extra euro in earnings makes a difference. This continues until we reach a point where quality of life saturates; it may even make us too comfortable that we start becoming lazy and prone to diseases such as obesity which could start taking a toll on quality of life. The dependency of quality of life on earnings is non-linear. Quality of life doesn’t depend just on earnings, but on a whole lot of things – most of it affect it non-linearly. So, it is not just one non-linear effect, but it is non-linearity upon non-linearity upon non-linearity that is embedded in human lives, or for that matter, any complex natural phenomena.
If computers are to model this kind of non-linear complexity to be able to effectively model human activities or to be useful to them in various ways, they ought to be able to deal with non-linearity abundantly and comfortably. The artificial neuron is a small mathematical device that can model non-linearity. It is inspired by the neuron in the brain, but only models it very parochially. Yet, several artificial neurons working together, one feeding into several others, can create the capacity to model complex non-linearity. This led to the wave of interest in deep learning – which is simply artificial neurons stacked together so the stacks are deep enough to model complex non-linearities. This, till date, remains the fundamental building block of the interest in AI today.
Hopfield used artificial neurons to create associative memory – the kind of memory which enables humans to create associations between events and use such associations for retrieval. On the other hand, Hinton popularized efficient and fast algorithms to ‘train’ networks of artificial neurons using data, so they are wired in ways that can then internalise complex phenomena encoded in data.”
Prof Rhodri Cusack, Thomas Mitchell Professor of Cognitive Neuroscience, Trinity College Dublin, said:
“Artificial neural networks were originally inspired by neuroscience, and there continues to be a flourishing interaction. Artificial neural networks have proven valuable models of processes in the brain, such as learning in human infants. This neuroscientific understanding is then inspiring new strategies for artificial neural networks. In short, machines are helping us understand ourselves, which in turn provides new avenues for technology. None of this would be possible without the seminal work of Hopfield and Hinton.”
Prof Peter Gallagher, Head of Astrophysics and Director of Dunsink Observatory, Dublin Institute for Advanced Studies (DIAS), said:
“Machine learning is transforming how researchers in space science and astrophysics are analysing and interpreting complex data sets from modern ground- and space-based scientific instruments.
For example, the Low Frequency Array (LOFAR) radio telescope produces huge amounts of data that we in the Dublin Institute for Advanced Studies and the Technological University of the Shannon, Athlone are analysing using machine learning algorithms. Machine learning allows us to automatically find and characterise large numbers of solar radio bursts that would be impossible to achieve by eye. The Irish LOFAR telescope (www.LOFAR.ie) is located at Birr Castle, Co. Offaly.
DIAS are also leading a Horizon Europe project called ARCAFF (www.ARCAFF.eu) that uses deep machine learning to characterise sunspots and to better forecast solar flares. Machine learning is critical to processing the large number of images that we receive to space missions, such as NASA’s Solar Dynamics Observatory, and then learning which solar features are importantly for predicting when a solar flare might occur.”
Declarations of interest
Prof Gallagher: none
Prof Cusack: none
Dr Padmanabhan: none