摘要 |
A device for recognising structure in data such as handwritten digits within a scanned image. The device is implemented on a computer (10) and comprises an array of ten sub-devices which can be used to recognise, for example, US postal codes. Each sub-device generates a score in respect of an input digit image which it has been trained to recognise. Each sub-device is trained on a training input image by adjusting weights which determine the score generated by the sub-device in response to an input image according to a five stage process in which a reconstruction of a training input signal is formed and used to adjust the weights to try to give a higher score to the training input signal than to the reconstruction. |