On Discriminative Learning of Prediction Uncertainty Vojtch Franc, - - PowerPoint PPT Presentation

on discriminative learning of prediction uncertainty
SMART_READER_LITE
LIVE PREVIEW

On Discriminative Learning of Prediction Uncertainty Vojtch Franc, - - PowerPoint PPT Presentation

On Discriminative Learning of Prediction Uncertainty Vojtch Franc, Daniel Pra Department of Cybernetics Faculty of Electrical Engineering Czech Technical University in Prague On Discriminative Learning of Prediction Uncertainty h ( x


slide-1
SLIDE 1

On Discriminative Learning of Prediction Uncertainty

Vojtěch Franc, Daniel Průša Department of Cybernetics Faculty of Electrical Engineering Czech Technical University in Prague

slide-2
SLIDE 2

Selective classifier: (h, c)(x) = h(x) with probability c(x) reject with probability 1 − c(x) where h: X → Y is a classifier and c: X → [0, 1] is a selection function

2/2

On Discriminative Learning of Prediction Uncertainty

slide-3
SLIDE 3

Selective classifier: (h, c)(x) = h(x) with probability c(x) reject with probability 1 − c(x) where h: X → Y is a classifier and c: X → [0, 1] is a selection function Example: Linear SVM h(x) = sign(φ(x), w + b) c(x) = [ [|φ(x), w + b| ≥ θ] ]

2/2

On Discriminative Learning of Prediction Uncertainty

slide-4
SLIDE 4

Selective classifier: (h, c)(x) = h(x) with probability c(x) reject with probability 1 − c(x) where h: X → Y is a classifier and c: X → [0, 1] is a selection function Example: Linear SVM h(x) = sign(φ(x), w + b) c(x) = [ [|φ(x), w + b| ≥ θ] ] Coverage: φ(c) = Ex∼p

  • c(x)
  • Selective risk:

RS(h, c) =

E(x,y)∼p

  • ℓ(y,h(x)) c(x)
  • φ(x)

20 40 60 80 100

coverage [%]

2 4 6 8

selective risk [%]

RS = 2.1%

SVM + distance to hyperplane

2/2

On Discriminative Learning of Prediction Uncertainty

slide-5
SLIDE 5

Selective classifier: (h, c)(x) = h(x) with probability c(x) reject with probability 1 − c(x) where h: X → Y is a classifier and c: X → [0, 1] is a selection function Example: Linear SVM h(x) = sign(φ(x), w + b) c(x) = [ [|φ(x), w + b| ≥ θ] ]

20 40 60 80 100

coverage [%]

2 4 6 8

selective risk [%]

RS = 2.1%

SVM + distance to hyperplane

In our paper we show: 1) What is the optimal c(x) 2) How to learn c(x) discriminatively

2/2

On Discriminative Learning of Prediction Uncertainty

slide-6
SLIDE 6

Selective classifier: (h, c)(x) = h(x) with probability c(x) reject with probability 1 − c(x) where h: X → Y is a classifier and c: X → [0, 1] is a selection function Example: Linear SVM h(x) = sign(φ(x), w + b) c(x) = [ [|φ(x), w + b| ≥ θ] ]

20 40 60 80 100

coverage [%]

2 4 6 8

selective risk [%]

RS = 2.1% RS = 0.2%

SVM + distance to hyperplane SVM + learned selection function

In our paper we show: 1) What is the optimal c(x) 2) How to learn c(x) discriminatively

2/2

On Discriminative Learning of Prediction Uncertainty