|
rlm@334
|
1 *Machine Learning and Pattern Recognition with Multiple
|
|
rlm@334
|
2 Modalities Hyungil Ahn and Rosalind W. Picard
|
|
rlm@333
|
3
|
|
rlm@334
|
4 This project develops new theory and algorithms to enable
|
|
rlm@334
|
5 computers to make rapid and accurate inferences from
|
|
rlm@334
|
6 multiple modes of data, such as determining a person's
|
|
rlm@334
|
7 affective state from multiple sensors--video, mouse behavior,
|
|
rlm@334
|
8 chair pressure patterns, typed selections, or
|
|
rlm@334
|
9 physiology. Recent efforts focus on understanding the level
|
|
rlm@334
|
10 of a person's attention, useful for things such as
|
|
rlm@334
|
11 determining when to interrupt. Our approach is Bayesian:
|
|
rlm@334
|
12 formulating probabilistic models on the basis of domain
|
|
rlm@334
|
13 knowledge and training data, and then performing inference
|
|
rlm@334
|
14 according to the rules of probability theory. This type of
|
|
rlm@334
|
15 sensor fusion work is especially challenging due to problems
|
|
rlm@334
|
16 of sensor channel drop-out, different kinds of noise in
|
|
rlm@334
|
17 different channels, dependence between channels, scarce and
|
|
rlm@334
|
18 sometimes inaccurate labels, and patterns to detect that are
|
|
rlm@334
|
19 inherently time-varying. We have constructed a variety of
|
|
rlm@334
|
20 new algorithms for solving these problems and demonstrated
|
|
rlm@334
|
21 their performance gains over other state-of-the-art methods.
|
|
rlm@333
|
22
|
|
rlm@334
|
23 http://affect.media.mit.edu/projectpages/multimodal/
|
|
rlm@334
|
24
|
