Notes
🔍

Per-Channel Quantization

A neural network’s channels do not all use numbers in the same range. One output channel in a convolution layer might contain very small weights, while another contains much larger ones. Per-channel quantization preserves that difference when converting a model from floating point to compact integer values.

How it works
With ordinary per-tensor quantization, an entire weight tensor shares one scale and zero-point: the values used to map real numbers into INT8 integers. Per-channel quantization instead assigns a separate scale—usually one per output channel of a convolution or fully connected layer. For each channel, the runtime approximates a real weight w as an integer using that channel’s scale. A small-range channel gets a fine-grained scale; a large-range channel gets a wider one. It is like using a different ruler for each drawer in a toolbox rather than forcing every tool to fit one ruler.

Why accuracy improves
A single scale must be large enough to represent the largest value anywhere in the tensor. That can make small values round too aggressively, effectively erasing useful weight differences. Per-channel quantization reduces this error, especially in convolution layers whose filters have uneven ranges. It commonly delivers INT8 model accuracy much closer to the original FP32 model than per-tensor weight quantization.

What it means on edge hardware

  • A smart camera can run an INT8 object detector with less flash and RAM use while retaining better detection quality.
  • A wake-word model can use efficient integer kernels continuously without a large accuracy loss.
  • Each channel’s scale adds a small amount of metadata and arithmetic, but this cost is tiny beside the accuracy gained.
  • The deployment runtime and accelerator must support it. TensorFlow Lite, for example, commonly uses per-axis quantization for convolution weights; unsupported operators can force slower fallback execution.

Per-channel quantization assigns separate scale factors, and optionally zero-points, to each output channel of a weight tensor rather than using one range for the entire tensor. It preserves accuracy when channels have different value distributions while still storing and computing weights in low precision such as INT8. This improves the accuracy–size trade-off for edge models, especially convolutional networks, with minimal metadata overhead.

Imagine packing clothes for a trip: using one suitcase size for everything can crush delicate items or waste space around small ones. Per-channel quantization gives each part of an AI model its own better-fitting “packing scale.”

It lets the model use smaller, simpler numbers so it takes less storage and can run faster on a phone, camera, or tiny sensor. But instead of applying the same amount of simplification everywhere, it adjusts each channel—one stream of information inside the model—separately. That helps preserve accuracy, especially when different parts contain values of very different sizes. The result is a compact on-device AI model that is less likely to lose important details.