The Bonsai.ML.PointProcessDecoder package provides a Bonsai interface to the PointProcessDecoder package used for decoding neural activity (point processes), and relies on the Bonsai.ML.Torch package for tensor operations.
The package can be installed by going to the bonsai package manager and installing the Bonsai.ML.PointProcessDecoder. Additional installation steps are required for installing the CPU or GPU version of the Bonsai.ML.Torch package. See the Torch installation guide for more information.
The PointProcessDecoder package is a C# implementation of a Bayesian state space point process decoder inspired by the replay_trajectory_classification repository from the Eden-Kramer Lab. It can decode latent state observations from spike-train data or clusterless mark data based on point processes using Bayesian state space models.
For more detailed information and documentation about the model, please see the PointProcessDecoder repo.
The following workflow showcases the core functionality of the Bonsai.ML.PointProcessDecoder package.
:::workflow
:::
The CreatePointProcessModel node is used to define a model and configure it's parameters. For details on model configuration, please see the PointProcessDecoder documentation. Crucially, the user must specify the Name property in the Model Parameters section, as this is what allows you to reference the specific model in the Encode and Decode nodes, the two main methods that the model will use.
During encoding, the user passes in a tuple of Observations and SpikeCounts. Observations are variables that the user measures (for example, the animal's position), represented as a (M, N) tensor, where M is the number of samples in a batch and N is the dimensionality of the observations. Spike counts are the data you will use for decoding. For instance, spike counts might be sorted spike data or clusterless marks. If the data are sorted spiking units, then the SpikeCounts tensor will be an (M, U) tensor, where U is the number of sorted units. If the data are clusterles marks, then the SpikeCounts tensor will be an (M, F, C) tensor, where F is the number of features computed for each mark (for instance, the maximum spike amplitude across electrodes), and C is the number of independant recording channels (for example, individual tetrodes). Internally, the model fits the data as a point process, which will be used during decoding.
Decoding is the process of taking just the SpikeCounts and inferring what is the latent Observation. To do this, the model uses a bayesian state space model to predict a posterior distribution over the latent observation space using the information contained in the spike counts data. The output of the Decode node will provide you with an (M x D*) tensor, with D* representing a discrete latent state space determined by the parameters defined in the model.