Brain inspired neuronal silencing mechanism to enable reliable sequence identification

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Real-time sequence identification is crucial for applications like event recognition and code verification.
• Current recurrent neural networks (ANNs) for this task face training challenges and complexity.
• Developing feedback-loop-free methods for sequence identification remains an open problem.

Purpose of the Study:

• To present a new neuronal plasticity mechanism for high-precision feedforward sequence identification networks (ID-nets).
• To demonstrate the effectiveness of ID-nets in identifying sequences without relying on feedback loops.
• To explore the generalization and potential applications of this novel mechanism.

Main Methods:

• Introduced a neuronal long-term plasticity mechanism that temporarily silences neurons after spiking.
• Developed feedforward identification networks (ID-nets) utilizing this silencing mechanism.
• Tested ID-nets on handwritten digit sequences and generalized to deep convolutional ANNs for image sequences.

Main Results:

• ID-nets reliably identified 10 handwritten digit sequences.
• The mechanism generalized to deep convolutional ANNs trained on image sequences.
• Achieved high classification performance for sequences, even with limited training data, outperforming individual object recognition.

Conclusions:

• The presented neuronal plasticity mechanism enables effective, feedback-loop-free sequence identification.
• ID-nets offer a promising approach for various sequence recognition tasks, including writer-dependent recognition and encrypted authentication.
• This mechanism opens new avenues for advanced artificial neural network algorithm development.