AI Dictionary › Fondamenti AI
Test di robustezza (Robustness Testing)
Robustness testing checks how well an AI model maintains reliable performance when the input changes unexpectedly, for example through typos, unusual phrasing or small variations from the training data. A robust model keeps giving sensible responses even under non-ideal conditions, while a fragile one can change behavior drastically in the face of small perturbations. It is an aspect of evaluation distinct from plain accuracy on standard cases.
To perform it, variants of the original input are generated, such as synonyms, rephrasings or artificial noise, and it is observed whether the model's response stays consistent and correct. Performance on the original input is then compared with performance on the variants, to quantify how sensitive the model is to change. Significant performance drops flag areas needing improvement.
It is used to validate models before production release, especially in contexts where real inputs are unpredictable, such as conversational assistants or decision-support systems. It also helps identify vulnerabilities exploitable by malicious inputs deliberately designed to fool the model.
It is a concept inherited from software engineering and control systems, where robustness has long referred to a system's ability to function correctly even under unforeseen or adverse conditions.
From our network
HSE Genius: AI for Safety Data Sheets
Extract SDS data, H phrases and ECHA compliance checks in seconds, powered by AI.
Visit hsegenius.com →From the Agora Intelligence blog
📱 Download the Android app (beta) iOS coming soon
Say what you mean. Get what you need.
Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.