Machine Learning: Too Smart for its Own Good
summary
A physical steampunk classifier is built from an aquarium air pump, gang valves and a Lego comparator that fuses 4-bit inputs into one output. It reliably separates linearly separable problems like even/odd and greater-than-3 by tuning valve weights, but prime detection requires extra layers and produces a false positive (1). The claim is that higher dimensions demand more layers and exponentially more training data, so with only 52 Mondays of operational logs models become fragile and easy for adversaries to break.
AI-written summary · about
chapters (5)
- 0:00Start
- 8:05Breaking the Law
- 14:47Simple Problems
- 19:06Every Neuron is a Feature for the Next Layer
- 23:59False positives come from...
related talks
28mSecurity of Machine Learning
44mComputers are Stupid: Protecting "AI" from Itself
34mDespicable machines: how computers can be assholes
20mMachine Learning Attacks – A New Era of Security Epidemic
24mLeak Detection System using Machine Learning Techniques
45mApplied machine learning: a few lessons I learned the hard way
39mMachine learning failures - for art!
41mBeyond the Algorithm: What Makes Machine Learning Work?