This is a self-contained memory module for the Dynamic Kanerva Machine, as reported in the NIPS 2018 paper: Learning Attractor Dynamics for Generative Memory.
Disclaimer: This is not an official Google product.
A memory used in the Omniglot experiment in the paper can be set-up as follows:
from memory import Kanervamemory
# Initialisation
memory = KanervaMemory(code_size=100, memory_size=32)
prior_memory = memory.get_prior_state(batch_size)
# Update memory posterior
posterior_memory, _, _, _ = memory.update_state(z_episode, prior_memory)
# Read from the memory using cues z_q
read_z, dkl_w = memory.read_with_z(z_q, posterior_memory)
# The KL-divergence between posterior and prior memory, used in training
dkl_M = memory.get_dkl_total(posterior_memory)z_episode is a [episode_size, batch_size, code_size] tensor as the
embedding of an episode. dkl_w is the KL-divergence between the w used reading.