Weakly Supervised Learning

A radiologist looks at an abdominal scan and sees not just organs but their spatial relationships: which kidney is larger, where the tumor lies relative to a vessel. It's like reading a map, where it's not enough to know where a city usually stands; you also need to measure the distance between two points on the ground.

A protein engineer is like a chef trying to make a dish that is tasty, healthy, and cheap all at once. Change one ingredient, and you risk ruining another. In the past, computer assistants could only evaluate protein properties one at a time, missing how they are connected.

You're looking for fireflies in a dark forest. Instead of photographing the whole forest every second and then poring over the images, you carry a small device that spots the flashes on its own and cuts out only a circle around each one. The rest, the blackness, isn't saved. That's exactly how the new sensor for scientific cameras works.

Under an electron microscope, a protein looks like a series of blurry shadows — hundreds of thousands of images of the same molecule, but from different angles and with heavy noise. Previously, from these shadows, one first assembled a volumetric 'haze'—a map of how electric charge is distributed inside the protein.

In the ICU, the decision of how much fluid to give a sepsis patient and what dose of a vasopressor to administer is made every hour. A mistake in either direction can be fatal. But how do you know which decision would have been best? You could take thousands of patient records and ask a program to "replay" them, trying different actions. That's exactly what the authors did.

Inside a piece of aluminum, atoms are arranged in neat rows, like bricks in a wall. But sometimes a row ends before reaching the edge, and a "step" appears in the crystal, which physicists call a dislocation. These defects determine how strong, ductile, and how a metal behaves under load.

A tumor in a scan is like a tiny bright dot on a vast dark canvas. That dot is what the doctor needs, not a perfectly smooth background. But when a neural network learns to turn a regular CT scan into something resembling a PET scan, it often tries to make everything look nice, and ends up smearing that dot.

Two scans of the same brain taken at different times are like photographs of the same house before and after renovation. To see what changed, you overlay them and look at the differences. But an MRI scanner is not a perfect camera: it adds its own distortions, as if the photo were taken through slightly warped glass.

A doctor spends years learning to recognize diseases from brain scans. First, they excel at spotting signs of aging, then they master tumors, and suddenly they start confusing what they knew before. Many neural networks behave much the same way: new knowledge overwrites the old.

A glucose meter that doesn't prick your finger but just looks at your wrist: that's the dream for millions of people with diabetes. It's the promise of algorithms that extract glucose levels from the ordinary pulse signal, the photoplethysmogram (the same one that measures heart rate in smartwatches).

Inside every cell, things are constantly meeting and parting. Some pairs hold tight, like Velcro, while others barely touch, like pollen on the wind. These fleeting touches, where the binding strength is comparable to a grain of salt dissolving in a glass of water, govern many of the cell's signals and decisions.

The proton beam is like a sniper's bullet: it delivers most of its energy at a precisely set point, barely touching tissues on the way in. But over 4–6 weeks of head and neck tumor treatment, the target shifts: the patient loses weight, the tumor shrinks, posture changes. The sniper aimed at one spot, but the target has already moved.

Assembling the puzzle of the Universe, astronomers are used to every new detail fitting perfectly into the already completed picture. But sometimes there are pieces that seem cut from another jigsaw. Twenty thousand active galactic nuclei, and eleven percent of them the neural network confidently rejects as alien.

4 610 universes — that's exactly how many cosmological models astrophysicists needed to show that the familiar Gaussian mask no longer works for next-generation data. Future sky surveys, such as the Euclid mission and the Vera C.

In 2019, the LIGO facility recorded a gravitational waves, a ripple in the fabric of spacetime, whose time-frequency map resembled a strange comb. Each such signal from a black hole merger leaves a unique pattern, like a fingerprint.

To capture every detail of a statue's shape, you could break it into millions of tiny cubes and describe each one — slow, heavy, and your computer's memory would groan at the seams. That's how 3D medical scans are usually handled: a CT or MRI is a huge array of such cubes (voxels), and to create or improve an image, the machine has to plow through all of them.

Ordinary AI looks at medical images like puzzles made of unfamiliar pieces: it considers any two unmatched pictures completely different. It's like trying to tell twins apart without knowing they're related. In medicine, however, patients with the same condition look nearly identical on MRI, and the algorithm mistakes 'relatives'.

Reconstructing a full picture from fragments is like discerning an orchestra's melody in a noisy hall. Such challenges arise in photometry, where a star's brightness is recovered from faint twinkling. Traditional methods need precise noise information, which is almost always unknown.

Radar, like a bat in a cave, sends out a pulse and listens for the echo. But a microscopic imbalance in the receiver is unavoidable: it scrambles frequencies and creates a phantom twin for each target, like a second, slightly distorted echo. Usually, these phantoms just get in the way and are cleaned up beforehand. In this new work, ghost echoes become an ally.

The photo enhancement algorithm is like a navigator that step by step leads a blurry picture toward a sharp view. It removes noise and distortions, moving pixel by pixel along a familiar map. But if the image was taken in unfamiliar conditions (say, with unusual blur) the map doesn't match the terrain.

A single-pixel camera works like an artist creating a picture through a stencil with many holes. In each frame, only a portion of the light, modulated by a mask, a spatial light modulator, reaches the single sensor, and then the computer reconstructs the full image by assembling this mosaic with computational algorithms.

Automatic speech judges are algorithms that evaluate pronunciation and emotions. However, they often work like a food critic who delivers a verdict after reading the menu but not tasting the dish.

Urban concrete jungles turn drones into blinded bats. Cameras go blind around every corner, GPS gets lost among skyscrapers, and radio waves scatter into a hundred echoes.

Learning a foreign language is like tuning a musical instrument. The new technology works like a smart tuner: it takes your recording and a native speaker's reference speech, stretches or compresses them in time (like adjusting a string) to align them precisely, and evaluates three key parameters: clarity of sounds, rhythmic precision, and intonation.

The brain communicates in short electrical clicks. Previously, to learn this language, thousands of examples with translations from the experimenter were needed. MOJO takes a different path: like learning a language through immersion, it first listens to the 'stream of consciousness' without any hints and finds recurring patterns on its own.

A drop of urine under a microscope isn't a flat puddle but a multi-layered volume. Like a stack of transparent pages: if you focus on one, the neighboring ones blur. The microscope's light only shows one depth layer sharply. To see all the cells, a lab technician had to take a series of snapshots, manually adjusting focus — tedious and slow.

The standard approach to a certain class of machine learning problems is like a spring that's been stretched and then forgotten: standard methods artificially inflate the problem's size, turning computations into a black hole for memory and time.

The AI diagnostician learns like a med student before an exam. You have an X-ray and patient complaints. But sometimes, a ready-made doctor's report pops up in the training materials — the full answer. Then the neural network memorizes it, and test accuracy jumps to 98%. However, that's cheating: real skill isn't developed.

The data from a single CT scan is comparable to several feature-length films. TCAM-Diff bypasses this bottleneck: the program compresses the 3D scene into three shadow projections — like a sculptor reconstructing a solid statue from silhouettes on three walls.

Usually a CT scanner takes many X-ray images from different angles and assembles a 3D picture. But sometimes doctors only scan a small area (say, one lung) and the data cuts off. It's like a chef given a single spoonful of soup and asked to name the exact recipe for the entire pot.

When a giant star collapses into a supernova or two neutron spheres merge in a cosmic dance, trillions of neutrinos carry away the cataclysm's energy.

A raw digital photo resembles a mosaic: each pixel holds only one base color — red, green, or blue. To compress such a file, it's split into brightness (a detailed black-and-white image) and color components.

In a noisy room, a person picks out a desired voice thanks to two ears capturing the difference in sound arrival. A computer with several microphones does the same: it gets spatial cues — where each voice is coming from. Previously, these cues were used to focus on one direction, like a flashlight beam.

When you type 'salad' and the keyboard suggests 'Caesar,' behind the scenes millions of phones are jointly learning to predict words. They share not your messages, but only improvements to the common algorithm, like a coach who sends athletes only the tweaks to the plan: 'add 5 push-ups to yesterday’s routine.' Information entropy helps send only unexpected, valuable updates.

Researchers took 16 everyday activities — from washing dishes to squatting — and broke the movements down into elementary components, much like a spectroscope splits light into a spectrum. Comparing three approaches, they found an unexpected key to recognition: a single averaged snapshot is enough. Only nine joints out of dozens carried the crucial information — the rest hardly mattered.

The float in a toilet tank is an ideal controller: sensing the water level, it instantly brings the system into equilibrium without looking back at the past. Engineers applied this principle to neural networks.

A large neural network like Whisper masters hundreds of languages and resembles the standard model of the universe: everything is balanced. But training on a rare language disrupts this equilibrium — it turns into a black hole, pulling in all training resources. The network starts forgetting the other languages.

Predicting the properties of a new molecule is like guessing the flavor of an unfamiliar dish. The program works like an experienced chef: it “tastes” similar molecules whose properties are already known, and based on their “flavor,” it refines its guess.

For a long time, identifying quantum phases was like interrogating a witness: you had to reach every corner of the system to uncover its secret. In the era of Landau, order was tied to symmetry breaking and a local parameter — everything plain as day.

Machine learning has long ceased to be just a tool — it is becoming an investigator, independently fishing out the laws of nature from data. But how does a silicon mind come to understand?