Machine-learning experiments · 2017
On Calibration of Modern Neural Networks
Guo C, Pleiss G, Sun Y, Weinberger KQ
What it found
Modern neural networks are often over-confident: the confidence they report does not match how often they are right. A simple adjustment after training, temperature scaling, fixed much of this on most datasets tested.
Where the evidence stops
Image and text classification benchmarks, not advice to people.
How it shapes Pursue
Pursue
A model's own confidence score cannot be shown as certainty. Pursue never invents a confidence number.
A design input, not evidence that Pursue works.
Cited in
- 01How should an AI say it isn't sure?Plainly, and without inventing a confidence number. Modern neural networks are often over-confident (Guo et al., 2017), and a system allowed to decline uncertain cases makes far fewer errors on the ones it answers (Geifman and El-Yaniv, 2017); a 2019 review led from the University of Cambridge found that admitting uncertainty does not necessarily cost trust.Trust and uncertainty
- 02Is there an app that notices when my day is slipping and suggests one next step?Tre-Up, an iOS app in development, is being built for exactly that, and it is not released yet: its AI, Pursue, is designed to notice when your focus or energy shifts and to answer with one small next step, one question, or nothing. Until it ships, your phone's focus modes and scheduled notification summaries cover part of the job.Timing and receptivity