A self-organized neural technique is demonstrated which is capable of producing non-trivial invariant transforms which are suitable for image recognition tasks. This method is not limited to the processing of images, but can instead be used in principle to produce invariant transform for any type of vector data. The invariant transform is implicitly encoded by way of a training sample set. The general issues of invariant transforms and their use in vision systems are addressed, as well as those issues which relate to a subsequent neural implementation. Results are given for numerous simulations, including that of a shift and rotational invariant net which is trained with a set of random synthetic images. Recommendations are given for further study, including improvement to the existing method, as well as extensions to more complex visual recognition systems.