Why Testing Is Not Proof: Neural Network Certification, Part 1
April the Siberian cat helps us follow one image from ordinary robustness to adversarial attacks, incomplete testing, and neural network certification.
Recent posts
Making Reliable Research Agents Accessible
How PIRA helps research agents stay focused, use strong evidence, make safe changes, and remember important work.
What ICLR 2026 Is Really About
My read of ICLR 2026: foundation models still dominate, but reliability, agents, evaluation, and security increasingly define the conference.
Chinese-dedede-Corrector: A Conservative 的地得 Checker
A rule-based tool for detecting likely misuse of 的 and suggesting 地 or 得.
Two Foundational Papers on Adversarial Examples
How early adversarial-example research moved from discovery to a fast attack and adversarial training.