learn/model families
Convolutional neural network
A convolutional neural network is a learning system built for data laid out on a grid, such as an image, that finds small local patterns and reuses the same pattern detectors everywhere.
What it is
Many kinds of scientific data sit on a grid: the pixels of a photograph, the slices of a three-dimensional scan, a plot of signal strength against time or frequency. A convolutional neural network, often shortened to CNN, works on such grids. It slides a small window across the data and, at each position, applies the same set of weighted sums, called a filter. One filter might respond to an edge, another to a blob, another to a diagonal streak. Because the same filter is used at every position, the network does not have to learn separately what an edge looks like in each corner of the image.
The outputs of those filters form a new grid, and further layers of filters run over that, so later layers respond to combinations of earlier patterns and to larger regions. Between layers the network usually shrinks the grid by keeping only the strongest response in each small patch, a step called pooling, and passes values through a simple rule that discards negatives. Whole blocks of layers are often stacked with shortcut connections that let information skip ahead, which makes very deep networks easier to train. At the end, the network can report a class, a position, or an outline.
Why scientists use it
Scientists reach for a CNN when the thing they want to read off the data is a shape or a texture rather than a number in a table. Deciding whether a smudge in a scan is a particle, or whether a trace is a real signal or an instrumental wobble, is work that people can do by eye but cannot easily write down as rules. A CNN is a way of learning those rules from examples that have already been labelled by hand.
It stands in, then, for the trained eye of an experienced observer, applied at a speed and volume no person could match. Modern instruments produce more images, spectra and scans than anyone can inspect. A CNN also tends to need fewer examples than a network that treats every pixel as unrelated, because reusing filters across positions builds in an assumption that is true of most physical data: a feature means much the same thing wherever it appears.
In the record
Limits
A CNN only learns what its labelled examples contain. Faced with an instrument, sample or sky it never saw in training, it may still report a confident answer, and it gives no warning that it is guessing. Its assumption that local patterns matter most makes it weaker when the meaning of the data depends on long-range relationships. It finds correlations, not causes, and can lock on to incidental features of how the data were collected. It also rarely explains itself, so its outputs usually need checking against independent measurements.
Written by AIxSci: the background from general knowledge, the examples from checked records.
By field
- Convolutional neural networks in structural biology and protein science
- Convolutional neural networks in materials and computational chemistry
- Convolutional neural networks in astronomy and astrophysics
- Convolutional neural networks in weather, climate and oceans
Named models
Every article
aix-00283
Deep learning forecast of El Niño suggests Indian and Atlantic oceans extend warning time
aix-00284
Neural network turns hourly satellite heat images into Gulf Stream current maps
aix-00287
Satellite image dataset extended to four channels for typhoon strength prediction
aix-00294
Neural network trained on simple equations spots tipping points in real data
aix-00298
Generative AI model forecasts UK rainfall 90 minutes ahead from radar images
aix-00299
Neural network forecasts weather straight from raw satellite and station observations
aix-00278
Neural networks sort millions of stars to build a young-star candidate catalogue
aix-00274
Neural network trained on simulated data hunts solar wave trains in radio spectra
aix-00219
Neural network supplies the sideways flows needed to measure the quiet Sun's magnetic energy
aix-00221
Neural network measures rotation periods for Kepler's main-sequence stars
aix-00223
Neural networks weigh cluster X-ray maps to estimate galaxy cluster masses
aix-00215
Neural network ranks gravitational wave triggers in search for neutron star mergers
aix-00245
Neural network trained on survey data retuned to spot transients for a liquid mirror telescope
aix-00248
Image classifiers flag a gas kink in a young star's disc, pointing to a planet
aix-00251
Astronomers map the mass of galaxy cluster Abell 2744 using lensed images
aix-00254
Neural networks read a fast radio burst's dispersion straight from its spectrum
aix-00209
Neural network turns Kaguya camera images into global maps of lunar surface chemistry
aix-00153
Neural networks pick amino acid sequences for short peptides in protein binding sites
aix-00152
Neural networks trained to find the cutting site on serpin proteins
aix-00150
Deep learning sorts crystal structures for single-molecule magnet behaviour
aix-00167
Neural network scans eleven years of night-sky images for mesospheric wave fronts
aix-00168
Deep learning reads cryo-EM maps alongside AlphaFold3 models to build protein structures
aix-00171
Neural network fills gaps in protein maps from raw diffraction data
aix-00138
AI-labelled protein pairs used to test how well function predictors handle unknown proteins
aix-00136
Diffusion model generates synthetic images of the Sun sorted by flare strength
aix-00134
Meta-learning lets a cosmology emulator adapt to new galaxy depth maps quickly
aix-00132
Five codes measured galaxy shapes in simulated Euclid images, one using neural networks
aix-00181
Protein language model embeddings used to predict succinylation sites in proteins
aix-00183
Neural network predicts nanostructure optics to design a two-colour laser collimator
aix-00184
Machine learning scores short protein stretches that mark proteins for destruction
aix-00190
A transformer model segments cryo-electron tomograms from three viewing directions at once
aix-00192
Study tests whether AI clean-up of cryo-EM maps harms ligand density
aix-00129
Neural networks sort supernovae and estimate their distances from brightness alone
aix-00128
Phage proteins drawn as pictures, then sorted by image-recognition networks
aix-00125
Classifying stars by type from a single wide-band telescope image
aix-00123
Neural networks swap places with black hole image simulations, both directions
aix-00121
Machine learning sorts mitochondrial proteins into three compartments from sequence alone
aix-00118
Neural networks replace a slow statistical fit to find solar oscillations in Hα images
aix-00106
Volunteer labels train a model to find asteroid trails in Hubble archive images
aix-00104
Neural networks sort 400,000 JWST galaxies to trace when spirals and spheroids appeared
aix-00100
Astronomers test pre-trained image models for sifting telescope alerts
aix-00089
Neural networks trained on atomic surroundings score how mutations change proteins
aix-00086
Neural networks trained on simulated radio bursts sort real bursts by shape
aix-00082
Neural networks turn simulated cluster mass maps into X-ray and SZ images
aix-00079
Changing the training penalty improves how a network spots beta-sheets in blurry protein maps
aix-00077
Neural networks sift 812,118 quasar spectra for hidden gravitational lenses
aix-00076
Neural network strips mains-power hum from LIGO data in about a second
aix-00073
Neural networks sort eleven years of Mars plasma data into three regions
aix-00065
Model predicts how fast an enzyme works on a substrate, with uncertainty attached
aix-00056
Deep learning tool predicts how single mutations weaken protein partnerships
aix-00055
Neural networks sift 315,000 galaxy images for rare polar ring systems
aix-00049
Eight machine learning methods sorted supernova gravitational wave signals by nuclear physics model
aix-00043
Neural network maps cosmic voids, walls and filaments from sparse galaxy tracers
aix-00042
Two unsupervised methods compress fast radio burst spectra into a handful of numbers
aix-00041
Astronomers hunt dust-buried young star clusters in eleven nearby galaxies
aix-00039
Machine learning predicts whether a dying massive star explodes, from its density profile
aix-00036
Four million simulated X-ray patterns used to test 21 crystal-symmetry classifiers
aix-00032
Neural network turns galaxy shape distortions into maps of invisible cluster mass
aix-00031
Keck spectra pin down distances for six gravitational lenses found by a neural network
aix-00030
Software spotted a nearby supernova and booked a telescope within minutes
aix-00029
Cryo-EM maps show water shifting in actin filaments as they age
aix-00026
Algorithm picks starting powders and temperatures for making inorganic materials
aix-00021
Neural network reads a million star spectra to measure temperature and chemistry
aix-00020
Neural network sifts Kepler signals and validates 301 new exoplanets
aix-00019
Neural network sorts six hundred thousand noise blips in LIGO gravitational-wave data
aix-00017
Neural network picks bright exploding stars from sky survey alerts and books telescope time
aix-00013
Neural network segments 650 battery particles to measure their detachment after cycling
aix-00010
Robotic lab with machine learning proposes and runs 353 inorganic synthesis experiments
aix-00008
Neural network locates single protein particles in cryo-electron microscope images
aix-00005
Cryo-tomography maps the myosin filament inside relaxed mouse heart muscle
aix-00003
Researchers map the sixteen-part Commander complex using crystals, cryo-EM and AlphaFold2
aix-00002
Deep learning turns raw NMR spectra into protein structures without human intervention
aix-00001