The author first studies classification as a simplified policy-gradient problem.
Training for remaining time beats myopic accuracy and endless patience
A finite-horizon loss that accounts for the training budget improved ImageNet accuracy over cross-entropy, particularly with noisy labels.
Big Tech
Ian Osband
Google DeepMind
Research Digest··3 min read
Osband reframes supervised learning as sequentially allocating a limited update budget across examples.
Why this paper
From Google DeepMind
In one line
Truncating cross-entropy's infinite patience to a finite training horizon improves classification accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§