Smaller neural networks could cut the cost of training AI models

A research team developed a technique that uses a stripped-down version of a neural network to identify which training examples matter most, then combines insights from that smaller network back into the full one. This reduces the computational work needed to train AI models while maintaining or improving accuracy — meaning you need fewer computers running for fewer hours to get a working AI system.

Training modern AI models is expensive because it requires processing massive amounts of data; if you could identify the truly useful examples upfront and discard the noise, you'd cut hardware costs and energy use significantly, which matters as AI systems consume more electricity and become more central to business operations.