Three types of opto-electronic hybrid neural networks are introduced. The errors induced by diffraction in these networks are emphasized and analyzed with computer simulation using the experimental parameters and input data， where diffraction, optical information processing and neural network theories are applied. It shows near-field diffraction induces large magnitudes of relative errors. When near-field diffraction and far-field diffraction are employed together, the errors are different according to the display modes. When the display mode is a small image, the magnitudes of relative errors are very large. However, when the display mode is a complicated image, the magnitudes of relative errors are small. The feasibility of using linear regression to calibrate the output data is discussed. It is found that linear regression can reduce the errors for about one magnitude. According to the analyses, the errors induced by diffraction can be minimized to a low level in the experiments, and thereby, the recognition rates can be maintained at a high level (larger than 97.7%).