liveBook · Manning

Activation-pattern diversity

As explained in Quantization and Fast Inference

Activation-pattern diversity is the variety of numerical activation patterns elicited by different inputs to a model.

Activation-pattern diversity is the range of numerical activations produced by different inputs. Calibration data should cover the production patterns that matter; missing user groups, time periods, or rare edge cases can leave calibration ranges unrepresentative. section 4.2

Image-condition shifts

In the ResNet-18 measurements, Layer 4 activation maxima range from 12.64 for edge-enhanced images to 22.60 for high-contrast images. Calibrating on one condition alone can miss such variation.

How Quantization and Fast Inference covers it

How deep it goes
Explains sampling, coverage validation, and activation-range measurements, but does not detail quantization internals.
Examples
Compares ResNet-18 image conditions and lists language-model prompt categories for calibration.
What you can do after
Apply stratified sampling and activation checks to build and validate calibration coverage.
How it connects
Connects calibration data to post-training quantization, activation ranges, and production-distribution coverage.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Quantization and Fast Inference. The section links point to the parts of the book it draws on.