EO-MTRNN: evolutionary optimization of hyperparameters for a neuro-inspired computational model of spatiotemporal...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spatiotemporal learning with neural networks often relies on manual hyperparameter tuning, a complex process for multiple timescale networks.
• This manual approach is time-consuming and limits the autonomy of artificial agents like robots.
• Existing methods lack efficient automatic hyperparameter optimization for complex spatiotemporal tasks.

Purpose of the Study:

• To develop an automated method for setting hyperparameters in multiple timescale recurrent neural networks (MTRNNs).
• To enhance the learning performance and autonomy of artificial agents using spatiotemporal data.
• To introduce the evolutionary optimized multiple timescale recurrent neural network (EO-MTRNN) inspired by neural plasticity.

Main Methods:

• The proposed evolutionary optimized multiple timescale recurrent neural network (EO-MTRNN) utilizes evolutionary optimization for timescale adjustment and network rewiring (neurons and synapses).
• The network does not require separate pre- and post-processing neural networks for input-output data.
• Validation involved benchmark datasets and robot sensory-motor data, comparing manual vs. automatic hyperparameter estimation.

Main Results:

• Automatically estimated hyperparameters in the EO-MTRNN resulted in approximately 43% better performance compared to manually set ones.
• The EO-MTRNN demonstrated effective learning without overfitting the training data.
• The network exhibited strong generalization abilities, successfully learning data not used during hyperparameter estimation.

Conclusions:

• The EO-MTRNN offers a significant advancement in automated spatiotemporal learning by optimizing hyperparameters effectively.
• This approach enhances learning efficiency and generalization capabilities, reducing reliance on manual expert tuning.
• The EO-MTRNN shows promise for improving the autonomy and performance of artificial agents in complex environments.