|
rlm@333
|
1 *Machine Learning and Pattern Recognition with Multiple Modalities
|
|
rlm@333
|
2 Hyungil Ahn and Rosalind W. Picard
|
|
rlm@333
|
3
|
|
rlm@333
|
4 This project develops new theory and algorithms to enable computers to
|
|
rlm@333
|
5 make rapid and accurate inferences from multiple modes of data, such
|
|
rlm@333
|
6 as determining a person's affective state from multiple sensors—video,
|
|
rlm@333
|
7 mouse behavior, chair pressure patterns, typed selections, or
|
|
rlm@333
|
8 physiology. Recent efforts focus on understanding the level of a
|
|
rlm@333
|
9 person's attention, useful for things such as determining when to
|
|
rlm@333
|
10 interrupt. Our approach is Bayesian: formulating probabilistic models
|
|
rlm@333
|
11 on the basis of domain knowledge and training data, and then
|
|
rlm@333
|
12 performing inference according to the rules of probability
|
|
rlm@333
|
13 theory. This type of sensor fusion work is especially challenging due
|
|
rlm@333
|
14 to problems of sensor channel drop-out, different kinds of noise in
|
|
rlm@333
|
15 different channels, dependence between channels, scarce and sometimes
|
|
rlm@333
|
16 inaccurate labels, and patterns to detect that are inherently
|
|
rlm@333
|
17 time-varying. We have constructed a variety of new algorithms for
|
|
rlm@333
|
18 solving these problems and demonstrated their performance gains over
|
|
rlm@333
|
19 other state-of-the-art methods.
|
|
rlm@333
|
20
|
|
rlm@333
|
21 http://affect.media.mit.edu/projectpages/multimodal/ |
