Simple framework for constructing functional spiking recurrent neural networks.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neural network modeling

Background:

• Cortical microcircuits feature complex recurrent architectures with dynamic properties.
• Neurons communicate via discrete spikes, making it unclear how these generate dynamics for complex computations.
• Continuous rate-coding neural network models can be trained for sophisticated tasks.

Purpose of the Study:

• To develop a framework for constructing biologically realistic spiking recurrent neural networks (RNNs) capable of learning diverse tasks.
• To bridge the gap between continuous rate models and spiking neural network models.
• To enable spiking RNNs to achieve performance comparable to rate-based RNNs.

Main Methods:

• A continuous-variable rate RNN is trained with biophysical constraints.
• Learned dynamics and constraints are transferred to a spiking RNN in a one-to-one manner.
• A single additional parameter is introduced to equate rate and spiking RNN models, with further parameter optimization.

Main Results:

• A simple framework is presented for building spiking RNNs from rate RNNs.
• The framework establishes a clear relationship between rate and spiking RNN models.
• Spiking RNNs constructed via this framework demonstrate performance on par with continuous rate networks.

Conclusions:

• The proposed framework facilitates the creation of effective spiking recurrent neural networks.
• This approach allows spiking RNNs to leverage the computational power of rate-based models.
• Biologically realistic spiking neural networks can be engineered to perform complex computational tasks effectively.