Information is carried in the brain by the joint electrical spiking patterns of large
groups of neurons, also known as the "neural code". The response of neurons to
repeated stimuli is surprisingly unreliable. The noisy nature of neurons limits the
capacity of spiking patterns to convey information, and how it can be read by the
brain. To accurately decode the joint activity of neurons, the brain must overcome
this noise and identify which patterns are semantically similar. We show that, in the
vertebrate retina, accurate models of network noise allow us to build a neural
thesaurus, measuring the similarity between population responses to visual stimuli
based on the content they carry. This thesaurus reveals that the neural code is
organized in clusters of synonym-like patterns that are similar in meaning, but may
appear different syntactically. This structure is highly reminiscent of the codebook
organization of engineered codes. We suggest that the brain may use this structure,
and show how it allows the accurate decoding of novel stimuli from novel spiking
patterns.