1The decoding of a sensory or motor variable from neural activity benefits from a known ground truth against 2 which decoding performance can be compared. In contrast, the decoding of covert, cognitive neural activity, 3 such as occurs in memory recall or planning, typically cannot be compared to a known ground truth. As a 4 result, it is unclear how decoders of such internally generated activity should be configured in practice. We 5 suggest that if the true code for covert activity is unknown, decoders should be optimized for generalization 6 performance using cross-validation. Using ensemble recording data from hippocampal place cells, we show 7 that this cross-validation approach results in different decoding error, different optimal decoding parameters, 8 and different distributions of error across the decoded variable space. In addition, we show that a minor 9 modification to the commonly used Bayesian decoding procedure, which enables the use of spike density 10 functions, results in substantially lower decoding errors. These results have implications for the interpreta-11 tion of covert neural activity, and suggest easy-to-implement changes to commonly used procedures across 12 domains, with applications to hippocampal place cells in particular. 13 65 (Zhang et al., 1998; Johnson and Redish, 2007; Pfeiffer and Foster, 2013).
66
Materials and Methods
67Overview 68 Our aim is to describe how the output of decoding hippocampal ensemble activity depends on the configu-69 ration of the decoder. In particular, we examine two components: (1) the split between training and testing 70 data, and (2) the parameters associated with the estimation of firing rates and tuning curves (the encoding 71 model). Both are described in the Analysis section. All analyses are performed on multiple single unit data 72 recorded from rats performing a T-maze task, described in the Behavior section. Data acquisition, annotation, 73 and pre-processing steps are described in the Neural data section.
74All preprocessing and analysis code is publicly available on our GitHub repository, https://github. 75 com/vandermeerlab/papers. Data files are available from our lab server on request by e-mail to the 76 corresponding author. 77 Neural data 78 Subjects and overall timeline. Four male Long-Evans rats (Charles River and Harlan Laboratories), weigh-79ing 439-501 g at the start of the experiment, were first introduced to the behavioral apparatus (described 80 below; 3-11 days) before being implanted with an electrode array targeting the CA1 area of the dorsal hip-81 5 pocampus (details below). Following recovery (4-9 days) rats were reintroduced to the maze until they ran 82 proficiently (0-3 days), at which point daily recording sessions began. On alternate days, rats were water-83 or food-restricted. In parallel with the maze task, some rats (R042, R044, R050) were trained on a simple 84 Pavlovian conditioning task in a separate room (data not analyzed).
85Behavioral task. The apparatus was an elevated T-maze, constructed from wood, paint...