ElefHead / neuromodulated-brc

Neuromodulated bistable recurrent cell from https://arxiv.org/abs/2006.05252

Neuromodulated Bistable Recurrent Cell

Tensorflow and Keras based implementation of a recurrent cell from RNN layer from A bio-inspired bistable recurrent cell allows for long-lasting memory by Nicolas Vecoven, Damien Ernst and Guillaume Drion.

Usage

To use the library, clone the repo and place bistablernn folder inside your project directory.

The Neuromodulated Bistable Recurrent Cell can be imported as:

from bistablernn import NBRCell

and use the cell to create an RNN layer as per keras API as:

tf.keras.layers.RNN(NBRCell(num_units))

Alternatively, a Neuromodulated Bistable RNN layer can be imported and used as:

from bistablernn import NBR

For example:

model = tf.keras.Sequential([
  NBR(units=num_hidden, input_shape=input_shape),
  tf.keras.layers.Dense(num_classes)
])

---

An example of training and evaluation using MNIST data is in the notebooks folder.

Dependencies

• Tensorflow 2 for layers.
• Tensorflow Datasets for MNIST.
• Jupyter for notebook.

Notes

• The Bistable Recurrent Cell modifies the GRU. My code also inherits keras GRUCell and GRU and overloads the functions to reflect the equation changes.
• The implementation is based on my understanding of the equations and modifications. The authors' implementation can be found here.
• The option to add dropout and recurrent dropout remain (assuming they work same as on GRU layers) but are untested.