norse.dataset.memory module

class norse.dataset.memory.MemoryStoreRecallDataset(samples, seq_length=100, seq_periods=12, seq_repetitions=4, population_size=5, poisson_rate=100, dt=0.001, seed=None)

Bases: Generic[torch.utils.data.dataset.T_co]

A memory dataset that generates random patterns of 4-bit data, and a 2-bit command pattern (store and recall).

Note that you can control the randomness by setting a manual seed in PyTorch.

Inspired by Bellec et al.: Biologically inspired alternatives to backpropagation through time for learning in recurrent neural nets.

Parameters

• samples (int) – Number of samples in the dataset.

• seq_length (int) – Number of timesteps to simulate per command. Defaults to 100.

• seq_periods (int) – Number of commands in one sample. Defaults to 12.

• seq_repetitions (int) – Number of times one store/recall pair occurs in a single sample. Defaults to 4.

• population_size (int) – Number of neurons encoding each command. Defaults to 5.

• poisson_rate (int) – Poisson rate for each command in Hz. Defaults to 250.

• dt (float) – Timestep for the dataset. Defaults to 0.001 (1000Hz).

• seed (Optional[int]) – Optional seed for the random generator