BrainForest: Neuromorphic Multiplier-Less Bit-Serial Weight-Memory-Optimized 1024-Tree Brain-State Classification Processor | JoVE Visualize

Area of Science:

Neurotechnology and Biomedical Engineering
Neuromorphic Computing
Epilepsy Treatment

Background:

Personalized brain implants offer revolutionary potential for neurological disorder treatment and cognitive augmentation.
Current epilepsy implants require low-power integrated circuits, limiting the integration of advanced machine-learning classifiers.
Existing devices necessitate energy-efficient solutions for continuous, life-long operation and improved therapeutic outcomes.

Purpose of the Study:

To introduce BrainForest, a neuromorphic processor designed for energy-efficient brain-state classification in personalized implants.
To enable machine-learning driven seizure detection for improved therapeutic stimulation in epilepsy treatment.
To overcome the power constraints hindering advanced computational capabilities in implantable devices.

Main Methods:

Developed a multiplier-less, bit-serial, weight-memory-optimized neuromorphic processor architecture.
Utilized two layers of neuron models: resonate-and-fire for EEG biomarker extraction and leaky integrator for multi-timescale classification.
Implemented sparse neural firing activity for clock-gating logic, reducing power consumption by 93%.
Integrated a 1024-tree boosted decision forest for pathological brain state classification and stimulation triggering.

Main Results:

Achieved state-of-the-art energy efficiency with ultra-low power consumption (9.6µW best case, 118µW typical) in 65nm CMOS.
Demonstrated high classification performance with 97.5% seizure sensitivity.
Reported a low false detection rate of 2.08 per hour, crucial for reliable therapeutic intervention.

Conclusions:

BrainForest offers a highly energy-efficient solution for on-chip brain-state classification, enabling advanced machine learning in low-power implants.
The architecture significantly reduces power consumption through sparse neural activity and optimized design.
This processor paves the way for more sophisticated and personalized neuromodulation therapies, particularly for epilepsy management.