normalizing flows

A translator learns a language not from a dictionary but from living conversations: they hear fragments of phrases, street noise, different accents. Out of that chaos they assemble rules that are written in no book.

Holding in mind the exact position of a pen as it's slowly moved is a routine task for the brain. For artificial neural networks, it's nearly impossible. Standard recurrent networks like GRU and LSTM tend to break continuous motion into rigid steps: the pen doesn't glide smoothly but jumps from point to point.

A basin of wet sand. You press in one corner, and water starts moving. But how exactly? Instead of an even circle, it might spread in tongues, linger in some spots, and rush through others. Underground, the same thing happens: oil, water, or gas migrate through rocks not by the textbook, but following the chaos of fractures and heterogeneities.

The optical sky background is a cosmic palimpsest, where light from billions of unresolved galaxies is layered, each layer corresponding to a different epoch. Until now we could only distinguish the brightest "letters"—individual stars and nearby galaxies.

Imagine a cookbook that gets rewritten from generation to generation. One day, the borscht recipe receives a secret ingredient from a pastry chef, and from that moment, the taste remembers not only ancestors but also a neighbor. In the Standard Model of evolution, genes behave the same way.

When an unmanned aerial vehicle learns from an experienced pilot, it memorizes not individual maneuvers but entire safe corridors — the trajectories along which the human mentor moved. This approach is called imitation learning: the machine mimics expert demonstrations.

A child exploring an unfamiliar apartment crawls from object to object and remembers what is visible from each corner. This is how the brain builds a mental map — an internal model of space. The gradCSCG algorithm works on the same principle: it receives a video stream from a camera and independently draws a map of the room, relying only on the changing images.

A snowflake grows, an oil slick spreads, a crowd moves — each case has its own equations. But the end result is often the same: shape, speed, and pattern follow a simple scheme that repeats across scales. This is universality — when a common law hides behind the details.

Gravitational microlensing is a fascinating cosmic effect predicted by Einstein and developed by Zwicky. A massive body passing between us and a distant star bends its light like a lens, creating characteristic magnifications and distortions.

The Universe is not silent: it is filled with a continuous gravitational-wave hum. In a decade, the LISA space observatory will tune its "ears" to the millihertz range, and the richest source will be our Galaxy — tens of thousands of white dwarf pairs whirling in a frantic dance.

X-ray detectors work like honeycombs: particles flash here and there in hexagonal cells. Conventional algorithms stretch these combs onto a square grid, creating ghostly polarization artifacts. Scientists used a neural network that lives by the rules of the hive — it processes the hexagonal structure directly, preserving the natural pattern.

Generative neural networks that create images typically rely on digital randomness. But a team of researchers built a system where randomness is born not in code, but in a tiny optical cavity filled with an organic dye.

A handful of sand falls onto a table: each grain rolls down until it rests in a hollow. That's how a potential works: the table's surface tells the grains where to go. But in a river with whirlpools, there's no single basin everything flows into. Water swirls, returns, and meanders. For such systems, the usual map of 'pits and hills' doesn't apply.

Population growth from a distance looks like a train picking up speed: it crawls at first, then gains momentum. But look closely at the data, and the picture blurs. One measurement shows a dozen individuals, another a hundred. Classical models wipe away this scatter like annoying smudges on a lens. They assume the noise is uniform and simply ignore it.

An artist paints a portrait stroke by stroke, and then someone offers them a flash camera — one click, and it's done. That's roughly the shift from multi-step models to one-step models in generating images of galaxies. Typical neural networks for image synthesis work like a watercolorist: they apply a layer, blur it, add details.

At a party, guests don't just chat in pairs; they gather in circles of shared interests: three discussing movies, four debating sports. The usual acquaintance map, like a phone's contact list, can't capture that: it only sees one-on-one links.

Quantum nothingness resembles water in a pot just before boiling. From time to time, a bubble spontaneously swells up — but not of steam, but of an entire universe. Such a bubble doesn't collapse; instead, it inflates into spacetime, triggering its own Big Bang. The only question is how often this happens.

Water in a river flows around rocks, air bends around an airplane wing: these motions are described by equations that are hard to solve even on a powerful computer. Neural networks can approximate such solutions, but they often err when data is collected at only a few points or is noisy.

A neuron in the brain is like a person handing out business cards at a crowded conference. At first, making acquaintances is easy, but as the address book fills up, each new contact requires more effort: there's little room left. The authors based their model on this 'crowding' rule: the more incoming connections a neuron already has, the harder it is to form another one.

An octopus tentacle bends, grabs, and stretches — all without a single bone. The secret is that inside it, sections that contract alternate with sections that remain passive. Scientists have long tried to replicate this trick in artificial materials, but until now they couldn't precisely dictate how a material would change shape. Now researchers have proposed 3D-printing elastic threads from two materials at once.

Previously, generative models for drug discovery worked like an artist painting without choosing a canvas. They learned from existing molecules and simply produced similar ones, without considering how the space of possible compounds is organized. SpaceGFN changes the approach: now chemical space is a construction set where the user specifies the parts and the assembly rules.

The proton is not a smooth ball but a tangle of quarks and gluons in constant motion, exchanging energy. Finding out exactly how energy, pressure, and spin are distributed inside it is impossible to do directly: measurements yield only sparse, noisy points. It's like trying to guess a face from a few blurry photos taken from different angles.

On a table lies a handful of round magnets. As long as no one touches them, they lie in disarray. But send a short current pulse through them, and each disk flips: the north pole swaps with the south. Neighboring disks push each other with their fields, like people in a crowded queue, and the whole handful arranges into a pattern that depends on which way the magnetic wind blew and how warm it is around.

Water in a stream bends around rocks, but the math of that motion hides in a black box: that's how ordinary neural networks work. They give an answer but don't say why. Engineers can't trust a prediction if they can't see its reasoning. Fluid dynamics now has a transparent model.

Glasses with flat lenses show a symmetric world: rise and fall, heat and cold look the same. For decades, climatologists have viewed El Niño through such glasses, using linear inverse models. These captured the rhythm of oscillations but missed that the warm phase of El Niño is typically sharper and stronger than the cold La Niña.

An interpreter at a conference hears the noise of the hall but keeps translating. They don't repeat every word; instead, they find an intermediate language where complex phrases become simple, yet meaning is preserved.

A translator who learns to correct their own slips works more accurately. In the world of forecasting noisy processes, a new method takes on this role: a reservoir computer that adds a random correction to its prediction at each step.

A translator who jumps at every rustle in the hall starts mixing up words by the end of the meeting. Their confidence in being right grows, while their accuracy drops. That is roughly how poorly calibrated probabilistic forecasts behave: the model claims a high probability for an event, but reality disagrees.