Abstract: Neuroscience is slowly transitioning into a data rich discipline and large data sets allow new approaches. Brain decoders use neural recordings to infer what someone is thinking, viewing, or their intended movement. The problem has always been phrased as a supervised learning problem.  Here we introduce a new method for brain decoding that does not require supervised data, i.e. the knowledge of the intended movement while the neural activity is recorded. Our approach is inspired by code breaking techniques used in cryptography where it is asked which mapping from from encrypted to decrypted text leads to text that most resembles the known structure of language. Analogously, we find a transformation of neural data (decoder) that aligns the distribution of the decoder output with the distribution of the user's intended movement. On a standard primate center-out reaching task, we demonstrate that we can obtain similar performance with that of a decoder with access to supervised data. However, current datasets are still too small to ask many relevant questions about neural computation and I am collaborating with neuroengineers to change that.