Four short links: 10 December 2019
- The Hidden Worries of Facial Recognition Technology — excellent post by Tsinghua Professor Lao Dongyan, questioning Chinese authorities’ plans to put facial recognition into the Beijing subway, with some great rebuttals to common objections. First, some people may think that I am overthinking it, and I cannot appreciate and thank the government, as a father figure, for its protection and kindness. I can only say: forgive me, but I cannot accept this type of kindness.
- A Framework for Regulating Competition on the Internet (Ben Thompson) — an interesting framework, where regulators and aggregators are differentiated. Platforms are the most powerful economic and innovation engines in technology: they create the possibility for products that never existed previously and are the foundation for huge amounts of innovation. It is in the interest of society that there be more and larger platforms, not fewer and smaller. [… For aggregators,] regulatory priorities should be the opposite of platforms: given that aggregator power comes from controlling demand, regulators should look at the acquisition of other potential aggregators with extreme skepticism. At the same time, whatever an aggregator chooses to do on its own site or app is less important, because users and third parties can always go elsewhere, and if they don’t, that is because they are satisfied.
- Failure Modes in Machine Learning (Microsoft) — excellent round up of intentional failures (perturbation attack, poisoning attack, model inversion, membership inference, model stealing, reprogramming ML system, adversarial example in the physical domain, malicious ML provider recovering training data, attacking the ML supply chain, backdoor ML, and exploit software dependencies) and unintentional failures (reward hacking, side effects, distributional shifts, natural adversarial examples, common corruption, and incomplete testing). (via BoingBoing)
- O(n^2) (Bruce Dawson)– Dawson’s first law of computing: O(n^2) is the sweet spot of badly scaling algorithms: fast enough to make it into production, but slow enough to make things fall down once it gets there. [After some debugging work,] I found that WinMgmt.exe was executing roughly a branch instruction per cycle which meant that the loop (which I already knew was consuming most of the CPU time) was running extremely quickly, and the slowness was because it was executing hundreds of billions of times.