Neural activity in a window can be represented as a coordinate in high-dimensional space. When mammal brains enact object recognition, encoded memories guide cortical neurons to “replay” previously seen neural states, which happen to be states that correspond to familiar scenes.
The spike2vec algorithm will enable us to track the trajectory of the network between familiar and unfamiliar states using a high-dimensional coordinate scheme. A network’s ability to revisit an encoded coordinate is testable, and so a spike2vector test of object recognition could be construed as a formal hypothesis test.
Some preliminary code that performs the spike2vec analysis is here. In addition to previously stated benefits the modern language makes large-scale model visualisation and analysis more computationally tractable because of its support for compressed data formats, and support for reduced precision types.