Using a neural network, we have developed a program for fast and precise detection of action potentials (AP) in raw multi-channel optical recording data. The AP detection was performed in two steps: first, peaks were detected in raw optical data, and, second, the peaks were classified by the neural network into APs, noise and undecided peaks. The network was optimized and trained by the backpropagation learning algorithm, employing some thousands of manually classified peaks. The performance of the optimized network was found to be not completely satisfactory, although it was better than the classification by template matching and nearest-neighbor rules. The addition of a signal-to-noise ratio (SNR) of a peak to the network classification improved the classification performance: in comparison with the manual classification results, 96% of manually classified APs were detected. The causes of classification errors were discussed. In spite of the fact that the program required a slight amount of human intervention for undecided peaks, the program could allow mostly automatic AP detection.