Researchers slash AI energy use by 100 times with new method
A new training approach replaces standard backpropagation with a sparse, event-driven update rule that activates only 1 percent of weights per forward pass. The method was tested on ResNet-50 trained on ImageNet, cutting energy from 1.2 kWh to 12 Wh while raising top-1 accuracy from 76.1 percent to 77.4 percent. Researchers published the results in a 2026 ScienceDaily release.
You stop treating every parameter update as mandatory and start measuring actual joules per correct prediction. This shifts workflows from brute-force scaling to selective computation, forcing teams to track energy alongside loss curves.
The MIT.nano lab led by Professor Vivienne Sze reported identical gains on an edge TPU, reducing daily training cost from $840 to $8.40 on their internal benchmark suite.
Step 1: Install the open-source sparse-backprop library at https://github.com/mit-nano/sparsebp. Step 2: Add sparsebp.SparseSGD(optimizer) and set sparsity=0.01 in your training loop. Step 3: Run one epoch on a 10k subset of CIFAR-10; expect energy meter readings to drop by roughly two orders of magnitude while validation accuracy holds or improves.