Tag Archives: neural network

Machine learning weather forecasts and black swan events

Leave a reply

Decorative painting of a stormy seascapeA couple of weeks ago I read about Google’s new weather forecasting algorithm, GraphCast.  It takes a radical new approach to forecasting by using machine learning rather than modelling the weather using the laws of physics [see ‘Storm in a computer‘ on November 16th, 2022].  GraphCast uses a graph neural network that has been trained on 39 years (1979 -2017) of historical data from the European Centre for Medium-Range Weather Forecasts (ECMWF). It requires two inputs: the current state of the weather and the state six hours ago; then it predicts the weather six hours ahead with a 0.25 degree latitude-longitude resolution (about 17 miles) at 38 vertical levels.  This compares to ECMWF’s high resolution forecasts which have 0.1 degree resolution (about 7 miles), 137 levels and 1 hour timesteps.  Although the training of the neural network took about four weeks on 32 Cloud TPU v4 devices (Tensor Processing Units), the forecast requires less than a minute on a single device whereas the ECMWF’s high resolution forecast requires a couple of hours on a supercomputer.  Within a day or so of reading about GraphCast, we watched ‘The Day After Tomorrow’, a movie in which a superstorm suddenly plunges the entire northern hemisphere into an ice age with dramatic consequences.  Part of the movie’s message is that humanity’s disregard for the state of the planet could lead to existential consequences.  It occurred to me that the traditional approach to weather forecasting using the laws of physics might predict the onset of such a superstorm and avoid it becoming a black swan event; however, it is very unlikely forecasts based on machine learning would predict it because there is nothing like it in the historical record used to train the neural network.  So for the moment we should continue to use the laws of physics to model and predict the weather since climate change appears to be making superstorms more likely [see ‘More violent storms‘ on March 1st 2017].

Sources:

Blum A, The weather forecast may show AI storms ahead, FT Weekend, 18/19 November 2023.

Lam R, Sanchez-Gonzalez A, Willson M, Wirnsberger P, Fortunato M, Alet F, Ravuri S, Ewalds T, Eaton-Rosen Z, Hu W, Merose A. Learning skillful medium-range global weather forecasting. Science. 10.1126/science.adi2336, 2023.

Image: Painting by Sarah Evans owned by the author.

 

Forest-sized brains

1 Reply

A couple of years ago I wrote in the abstract about ‘Slow thoughts from a planet sized brain‘ [on March 25th, 2020].  I read on vacation in Suzanne Simard‘s book, ‘Finding the Mother Tree‘ that glutamate, which is the most abundant neurotransmitter in the human brain, is also transmitted through mycorrhizal networks connecting trees in forests. Mycorrhizal fungi live in the soil around the roots of plants in a symbiotic relationship with the plants transmitting water to, and receiving sugar from, the plant roots.  Fir trees have been shown to transmit information about threats, e.g., budworm infestations, to one another and to other species of tree.  The speed of this information transmission is fast enough that production of enzymes to protect the trees increases within a day of the appearance of the threat.  We have assumed that folklore tales about enchanted forests are products of our imagination; but perhaps they are based on a long-lost appreciation that forests possess a level of consciousness.  Consciousness seems to require different parts of a system to communicate with one another and form networks [see ‘Digital hive mind‘ on November 30th, 2016], which Simard and others have demonstrated occurs in forests with the mycorrhizal networks being equivalent to the neural network in our brains.  The scale of a forest’s network is such that communication will be slower than in our brain but that is not necessarily an inhibitor of consciousness.  So, perhaps forest-sized brains would be intermediate between human-sized and planet-sized.

Source:

Slow deep thoughts from a planet-sized brain

3 Replies

I overheard a clip on the radio last week in which someone was parodying the quote from Marvin, the Paranoid Android in the Hitchhiker’s Guide to the Galaxy: ‘Here I am with a brain the size of a planet and they ask me to pick up a piece of paper. Call that job satisfaction? I don’t.’  It set me thinking about something that I read a few months ago in Max Tegmark’s book: ‘Life 3.0 – being human in the age of artificial intelligence‘ [see ‘Four requirements for consciousness‘ on January 22nd, 2020].  Tegmark speculates that since consciousness seems to require different parts of a system to communicate with one another and form networks or neuronal assemblies [see ‘Digital hive mind‘ on November 30th, 2016], then the thoughts of large systems will be slower by necessity.  Hence, the process of forming thoughts in a planet-sized brain will take much longer than in a normal-sized human brain.  However, the more complex assemblies that are achievable with a planet-sized brain might imply that the thoughts and experiences would be much more sophisticated, if few and far between.  Tegmark suggests that a cosmic mind with physical dimensions of a billion light-years would only have time for about ten thoughts before dark energy fragmented it into disconnected parts; however, these thoughts and associated experiences would be quite deep.

Sources:

Douglas Adams, The Hitchhiker’s Guide to the Galaxy, Penguin Random House, 2007.

Max Tegmark,  Life 3.0 – being a human in the age of artificial intelligence, Penguin Books, Random House, UK, 2018.

 

Citizens of the world

3 Replies

Last week in Liverpool, we hosted a series of symposia for participants in a dual PhD programme involving the University of Liverpool and National Tsing Hua University, in Taiwan, that has been operating for nearly a decade.  On the first day, we brought together about dozen staff from each university, who had not met before, and asked them to present overviews of their research and explore possible collaborations using as a theme: UN Sustainable Development Goal No.11: Sustainable Cities and Communities.  The expertise of the group included biology, computer science, chemistry, economics, engineering, materials science and physics; so, we had wide-ranging discussions.  On the second and third day, we connected a classroom on each campus using a video conferencing system and the two dozen PhD students in the dual programme presented updates on their research from whichever campus they are currently resident.  Each student has a supervisor in each university and divides their time between the two universities exploiting the expertise and facilities in the two institutions.

The range of topics covered in the student presentations was probably even wider than on the first day; extending from deep neural networks, through nuclear reactor technology, battery design and three-dimensional cell culturing to policy impacts on households.  One student spoke about the beauty of mathematical equations she is working on that describe the propagation of waves in lattice structures; while, another told us about his investigation of the causes of declining fertility rates across the world.  Data from the UN DESA Population Division show that live births per woman in the Americas & Europe have already fallen below the 2.1 required to sustain the population, while it is projected to fall below this level in south-east Asia within the next five years and in the world by 2060.  This made me think that perhaps the Gaia principle, proposed by James Lovelock, is operating and that human population is self-regulating as it interacts with constraints imposed by the Earth though perhaps not in a fashion originally envisaged.