A bio-inspired bistable recurrent cell allows for long-lasting memory.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Recurrent neural networks (RNNs) excel at tasks requiring memory, often using gated cells like GRU and LSTM.
• Standard RNNs store information at the network level, with long-term memory influenced by connection weights.
• Biological neurons exhibit bistability, enabling cellular-level long-term information storage.

Purpose of the Study:

• To develop a novel recurrent cell inspired by biological neuron bistability.
• To improve RNN performance on time-series data demanding very long memory.
• To explore biologically plausible mechanisms for RNNs, including neuromodulation.

Main Methods:

• Introduction of a new bistable, biologically-inspired recurrent cell.
• Utilizing only cellular connections (neuron-to-itself) for memory storage.
• Equipping the cell with recurrent neuromodulation to link with standard GRU cells.

Main Results:

• The new bistable cell significantly enhances RNN performance on long-memory time-series tasks.
• The cell achieves this improvement using only self-connections, independent of other neurons.
• Neuromodulation provides a link to GRU cells, advancing biological plausibility.

Conclusions:

• Biologically inspired bistability offers a powerful mechanism for long-term memory in RNNs.
• The novel cell architecture improves RNNs for challenging time-series data.
• This work opens avenues for more biologically plausible RNNs using neuromodulation.