poster - å ¯æcat.phys.s.u-tokyo.ac.jp/~zliu/slides/neurips2019_poster...Title Microsoft...
Transcript of poster - å ¯æcat.phys.s.u-tokyo.ac.jp/~zliu/slides/neurips2019_poster...Title Microsoft...
Deep Gamblers: Learning to Abstain with Portfolio TheoryLiu Ziyin Liu (UTokyo), Zhikang T. Wang(UTokyo), Paul Liang(CMU), Ruslan salakhutdinov (CMU), Louis-Philippe Morency (CMU), Masahito Ueda (UTokyo),
Classification and the Inadequacy of 𝑛𝑙𝑙 lossWant to find: 𝜃 = arg max Pr(𝑌|𝜃)
In practice, minimize 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑙𝑜𝑔 𝑙𝑜𝑠𝑠(𝑛𝑙𝑙 loss): min − log 𝑝(𝑌|𝜃)
The proposed method: the gambler’s loss
max 𝐸 log 𝑆 = max 𝑝 log(𝑜 𝑏 + 𝑏 )
SOTA Performance…
Surprising Benefit:-Training with gambler’s loss reduces overfit-Improved performance when noisy label is present
The Learned Representation is Better Separable:
Toy Example: Image Rotation..
Toy Example: Identifying Disconfident Images..
Intuition: Prediction as Horse RaceHorse Race with Reservation
𝑚 horses Betting strategy: ∑ 𝑏 → ∑ 𝑏 Chance of winning: 𝑝Payoff if we bet on the winning horse: 𝑜Return after winning: 𝑆 = 𝑜 𝑏 → 𝑜 𝑏 + 𝑏
Objective: maximize doubling rate:
max 𝑊 = max 𝐸 log 𝑆 = max 𝑝 log(𝑜 𝑏 + 𝑏 )
Classification Problem = Betting problem with Reservation with 𝑜 = 1, 𝑏 = 0
Classification Problem ≤ Betting problem with Reservation