Quantization
A technique that shrinks AI models by reducing the precision of numbers they use, making them faster and smaller with little accuracy loss.
In Plain English
Quantization is like switching from a high-resolution photo to a lower-resolution version that still looks good but takes up less space. AI models store enormous numbers to make predictions, and those numbers are often stored with high precision (many decimal places). Quantization rounds these numbers to simpler, shorter forms—for example, using whole numbers instead of decimals—which cuts the model's size dramatically and speeds it up. The trick is doing this without losing the model's ability to make accurate predictions. This matters because smaller, faster models can run on phones, smart devices, and edge hardware that can't handle huge models.
💡Real-World Example
A fitness watch uses AI to detect your heart rhythm on the device itself, rather than sending data to the cloud. The heart-monitoring model had to be quantized so it could fit in the watch's limited memory and run without draining the battery in hours. By carefully reducing the precision of the model's numbers, engineers got it to run on the wrist device while still accurately detecting irregular heartbeats.
Related Terms
What did you think of our explanation?