14.02.2013 Views

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Chapter 20. Signal Process<strong>in</strong>g 86(3):603-623, 2006 275<br />

(a) A JADE<br />

0.6<br />

0.4<br />

0.2<br />

0<br />

0.6<br />

0.4<br />

0.2<br />

1 2 3 4 5 6 7 8<br />

(d) A sNMF<br />

0<br />

1<br />

2<br />

3<br />

4<br />

5<br />

6<br />

7<br />

8<br />

1<br />

2<br />

3<br />

1<br />

2<br />

3<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

channel<br />

0<br />

1<br />

2<br />

3<br />

4<br />

5<br />

6<br />

7<br />

8<br />

0.6<br />

0.4<br />

0.2<br />

(b) A NMF<br />

signal<br />

0<br />

1<br />

2<br />

3<br />

4<br />

5<br />

6<br />

7<br />

8<br />

1<br />

2<br />

3<br />

(e) A sNMF*<br />

1<br />

2<br />

3<br />

0.6<br />

0.4<br />

0.2<br />

0<br />

1<br />

2<br />

3<br />

4<br />

5<br />

6<br />

7<br />

8<br />

0.6<br />

0.4<br />

0.2<br />

(c) A NMF*<br />

(f) A SCA<br />

0<br />

1<br />

2<br />

3<br />

4<br />

5<br />

6<br />

7<br />

8<br />

1<br />

2<br />

3<br />

1<br />

2<br />

3<br />

Fig. 4. Mix<strong>in</strong>g matrices recovered for the synthetic s-EMG us<strong>in</strong>g different methods;<br />

(a) ICA us<strong>in</strong>g jo<strong>in</strong>t approximate diagonalization of eigenmatrices (JADE), (b, c)<br />

nonnegative matrix factorization with different preprocess<strong>in</strong>g (NMF, NMF∗), (d, e)<br />

Sparse NMF (sNMF, sNMF∗) with the same two preprocess<strong>in</strong>g methods as NMF,<br />

and (f) sparse component analysis (SCA).<br />

3.1 Artificial signals<br />

In the first example, we compare performance <strong>in</strong> the well-known sett<strong>in</strong>g of<br />

artificially created toy-signals.<br />

3.1.1 S<strong>in</strong>gle s-EMG<br />

For visualization, we will first analyze a s<strong>in</strong>gle artificial s-EMG record<strong>in</strong>g, and<br />

only later present batch-runs over multiple separate realizations to test for<br />

statistical robustness. As data set, we use toy signals as <strong>in</strong> section 3.1 but<br />

with only three source components for easier visualization. The ICA-result<br />

is produced us<strong>in</strong>g JADE after PCA to 3 components. Please note that here<br />

and <strong>in</strong> the follow<strong>in</strong>g we perform dimension reduction because <strong>in</strong> the small<br />

sensor volumes <strong>in</strong> question not many MUs are present. This is confirmed by<br />

consider<strong>in</strong>g the eigenvalue structure of the covariance matrix: tak<strong>in</strong>g the mean<br />

over 10 real s-EMG data sets — further discussed <strong>in</strong> section 3.2 — the ratio<br />

of third to first largest eigenvalue is only 0.11, and the ratio of the fourth to<br />

the first only 0.04. Tak<strong>in</strong>g sums, <strong>in</strong> the mean we loose only 5.7% of raw data<br />

14

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!