Dynamic Self-Organizing Neurons

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence Hardware
• Solid-State Devices

Background:

• Current deep neural network (DNN) accelerators are often application-specific and lack adaptability to dynamic environments.
• Existing architectures and algorithms in DNN accelerators are rigid, limiting their flexibility.
• Supervised learning has been the primary focus for many DNN accelerators.

Purpose of the Study:

• To propose a novel neuromorphic architecture for self-organizing feature maps (SOFMs).
• To utilize ferroelectric field-effect transistors (FeFETs) for in-memory computation within the neuromorphic architecture.
• To create an adaptable and efficient accelerator for diverse AI applications.

Main Methods:

• Implementation of a self-organizing feature map (SOFM) using ferroelectric field-effect transistors (FeFETs).
• Design of a neuromorphic architecture inspired by biological networks, allowing for neuron growth and adaptive topography.
• In-memory computation for error correction and processing.

Main Results:

• Demonstrated the neuromorphic architecture's ability to adapt to various datasets.
• Showcased lifelong learning and self-repair capabilities of the network.
• Validated the architecture's efficiency in terms of power and speed, alongside robustness to device variability.

Conclusions:

• The proposed FeFET-based SOFM neuromorphic architecture offers a flexible and efficient solution for AI acceleration.
• The architecture's adaptive nature, including neuron growth and topographic modulation, enables lifelong learning and self-repair.
• This approach overcomes the limitations of rigid, application-specific DNN accelerators.