AITowards Data Scienceabout 2 hours ago
Neural Network Learns Digits From Noise
8 min read
Researchers at Anthropic and UC Berkeley show a student network learns handwritten digits from pure random static by copying teacher outputs. The student reached 89.3% accuracy on the MNIST test set, crossing 50% by step 468 out of 2,340 steps, but only when starting from the same random weights as the teacher.
Level
Hype check
Tap to vote and see what everyone thinks.