Information Technology Reference
In-Depth Information
Table 1. Correct ratio of RBF networks when SNR of test signals decreasing. To converge in
reasonable time at a relatively low SNR, we have to larger the tolerance of the training period.
Because the neural network has good generalization ability, it's not necessary to train the net-
work for every group of data. So the train sets are only made up with samples of -10dB, -30dB,
-35dB.
Test Signals(dB)
Correct Ratio(%)
Tolerance
Train Signals(dB)
-10
98.67
0.5
-10
-15
99.33
0.5
-10
-20
99.0
0.5
-10
-25
96.33
8
-30
-30
86.67
8
-30
-35
75.67
60
-35
-40
73.67
60
-35
Table 2. Comparison of the correct ratio using RBF networks and BP networks. The results show
that RBF networks achieve a better result when SNR comes to -35dB or lower.
Test Signals(dB)
K-means RBFNN(%)
BPNN(%)
-10
98.7
99.3
-15
99.3
98.7
-20
99.0
98.7
-25
96.3
96.0
-30
86.7
90.0
-35
75.7
69.3
-40
73.7
71.3
5 Conclusion
In this paper we propose 2-stage RBF network method for neural spike sorting. The
results show that RBF network performs well under high noise. So RBF network can be
an effective tool in spike sorting. However, there are some problems with this method:
Some parameters such as width of reception field need to be set manually, which is a
time-consuming task. The performance is affected greatly by distribution of samples
because of the application of K-means method in the first stage of RBF network.
Therefore, in our future work we will consider improve the performance using Genetic
Algorithm to make RBF network self-adaptive [6]. We will do further research on
neural decoding in Brain-Computer Interface [7].
Search WWH ::




Custom Search