(Poster ID W20) Edwin V. Bonilla, Kian Ming A. Chai, Christopher K. I. Williams School of Informatics, University of Edinburgh We propose a model based on a GP prior that directly induces correlations between tasks specified by a task-similarity matrix K f . 1. The Model: f (x)fm (x ) yi = K fm k x (x, x )


Multi-task Gaussian Process Prediction

3. Cancellation of Transfer: Noiseless observations + grid:
f1 f2 f3 x1 x2 x* x3 x4

N (f (xi ),  2 ).

Observations in one task affect predictions on the others.


4. Results: Compiler task, school data
f3

f1

f2

f1

f2

f3
0.12

NO TRANSFER TASK-DESCRIPTOR FREE-FORM

y1

y2

y3

y1

y2

y3
MAE

0.1 0.08 0.06 0.04

Other approaches 2. Predictions: ¯ f (x ) =

Our approach (k  kx )T -1 y,  f x
f

0.02 0 16 32 64 128

with 

= K K +D I

N