Fast and Scalable Training of Semisupervised CRFs for Activity Recognition
Maryam Mahdaviani, University of British Columbia and Tanzeem Choudhury, Intel Research
Poster ID: T48

m le Collecting labeled data is expensive in activity recognition and also in many other domains b ro P







Solution A fast and efficient semisupervised  training algorithm for features selection and ve(n( y )) parameter estimation in CRFs
i

yi
xi1 x
2 i

ve(n( yi ))

ve(n( yi ))

yi
xi1 xi2 xiN

ve(n( yi ))

L(w) = - log p( yi | ve(xi, n( yi)) + 
l l

t

t +t
l

u

i=1

i=t +1

Hu (i)

x

N i

labeled data

unlabeled data

Objective function minimizes the sum of local negative loglikelihoods given soft evidence from  neighbors plus entropy of unlabeled data Results Our method outperforms other proposed semisupervised  CRF training approaches in  realworld activity recognition tasks and is an order of magnitude faster to train