1 Multi-Task Semi-Supervised Underwater Mine Detection Lawrence Carin, Qiuhua Liu and Xuejun Liao Duke University Jason Stack Office of Naval Research.
Post on 08-Jan-2018
216 Views
Preview:
DESCRIPTION
Transcript
1
Multi-Task Semi-Supervised Underwater Mine Detection
Lawrence Carin, Qiuhua Liu and Xuejun LiaoDuke University
Jason StackOffice of Naval Research
Intra-Scene Context
What Analyst Processes Individual Signatures Processedby Supervised Classifiers
Message:
Analyst Places Classification of Any Given Item Within Context of All Items in the SceneSupervised Classifier Classifies Each Item in Isolation
Decision surface based on labeled data (supervised)
Decision surface based on labeled & Unlabeled data (semi-supervised)
Inter-Scene Context
8
Message
Humans are very good at exploiting context, both within a given scene and across multiple scenes
Intra-scene context: semi-supervised learning
Inter-scene context: multi-task and transfer learning
A major focus of machine learning these days
9
Data Manifold Representation Based on Markov Random Walks
Given X={x1, …,xN}, first construct a graph G=(X,W), with the affinity matrix W, where the (i, j)-th element of W is defined by a Gaussian kernel:
we consider a Markov transition matrix A, which defines a Markov random walk, where the (i, j)-th element:
gives the probability of walking from xi to xj by a single step.
The one-step Markov random work provides a local similarity measure between data points.
)2/exp( 22
ijiij xxw
N
k ik
ijij
w
wa
1
10
Semi-Supervised Multitask Learning(1/2)
Semi-supervised MTL: Given M partially labeled data manifolds, each defining a classification task, we propose a unified sharing structure to learn the M classifiers simultaneously.
The Sharing Prior: We consider M PNBC classifiers, parameterized by
The M classifiers are not independent
but coupled by a joint prior distribution:
,m....,2,1 Mm
M
mmmM pp
1111 ),..,|(),..,(
11
Semi-Supervised Multitask Learning(2/2)
With
The normal distributions indicates the meta-knowledge indicating how the present task should be learned, based on the experience with a previous task.
When there are no previous tasks, only the baseline prior is used by setting m=1 =>PNBC.
Sharing tasks to have similar , not exactly the same(advantages over the Dirac delta function used in previous MTL work).
s'
1
1
211 ),;()|(
11),..,|(
m
lmllmmmm Np
mp Iγ
Baseline prior Prior transferred from previous tasks
Balance parameter
13
14
15
Thanks
top related