Research
My research bridges neuromorphic hardware, bio-inspired algorithms, and efficient AI systems to enable adaptive perception and learning in robotics and edge computing applications.
Research Areas
Efficient AI Algorithms & Architectures
Developing neuromorphic optical flow methods achieving 400% speedup over conventional approaches, with real-time processing capabilities (~1-2ms) for motion prediction, object segmentation, and tracking in dynamic environments like autonomous driving and robotics.
Bio-inspired Neuromorphic Systems
Implementing biologically-realistic perception mechanisms including high-order associative learning, differential sensory processing, and adaptive nociception for safe robotic grasping and navigation in unstructured environments with human-like capabilities.
Simulation & Modeling Resources for Emerging Memory
Creating general memristive transistor models (GEM) with 300% improvement in switching accuracy, real-time state modulation circuits, and comprehensive SPICE implementations for neuromorphic hardware design and validation across different device types and scales.
Featured Projects
Neuromorphic Spatiotemporal Optical Flow
A 2D floating-gate synaptic transistor array encodes temporal motion cues directly in hardware, enabling identification of motion regions in 1-2ms. Combined with spatial gradient information, this method achieves 400% faster processing while maintaining 94% accuracy in autonomous driving scenarios across vehicle operation, UAV navigation, and sports activities.
Key Innovation: Spatiotemporal consistency processing allows selective filtering of visual input, accelerating velocity calculations and task execution to surpass human-level speeds (~150ms).
Differential Neuromorphic Computing for Adaptive Perception
Memristor-based differential processing mimics biological sensory systems for perception in unstructured environments. A single SDC memristor achieves >720% amplification of hazardous stimuli and <50% attenuation of mild stimuli, enabling safe grasping of sharp and slippery objects with sub-millisecond response times.
Key Innovation: Feature extraction drives adaptive memristive modulation schemes, replicating nociceptor, fast-adapting, and slow-adapting receptor behaviors for robot manipulation tasks in dynamic environments.
High-Order Memristive Associative Learning
Bio-inspired associative learning framework using memristor state information to bridge different orders of learning. Achieves 230% improvement in learning efficiency over previous works in Pavlov's conditioning experiments, with memristor power consumption <11μW. Scales to 20×20 arrays for image recognition with 100% accuracy.
Key Innovation: State-involved differential equation and voltage-controlled moving window functions enable transient high-order associations, mimicking neural mechanisms in Drosophila.
GEM: General Memristive Transistor Model
A behavioral-level model incorporating time-dependent differential equations, voltage-controlled moving windows, and nonlinear output functions. Achieves 300% improvement in switching behavior modeling (RMSE <0.013) across carbon nanotube, WS₂/PZT FeFET, and HfS₂/h-BN/FG-graphene devices, while capturing physical limits and output characteristics.
Key Innovation: First general model to incorporate both state-involved dynamics and amplitude-dependent state ranges, enabling accurate simulation of diverse memristive transistor types with SPICE implementation available on GitHub.
Real-Time State Modulation and Acquisition Circuits
Dual-function circuit enabling concurrent memristor state modulation and acquisition for neuromorphic systems. Achieves mean absolute acquisition error <1Ω in nociceptor applications and 0.076kΩ in 4×4 crossbar arrays, with modulation voltage deviation at mV level. Eliminates traditional two-step read-write processes, enhancing real-time performance.
Key Innovation: Feedback op-amp circuit with voltage divider and sample-and-hold leverages memristor threshold characteristics for stable, continuous state observation during modulation without additional read pulses.
Research Impact
Performance
- ✓ 400% speedup in optical flow processing
- ✓ 1-2ms motion region detection
- ✓ 94% accuracy in driving scenarios
- ✓ 230% learning efficiency improvement
Hardware Metrics
- ✓ <11μW memristor power consumption
- ✓ ~100μs synaptic response time
- ✓ >10⁴s retention stability
- ✓ >8,000 programming cycles
Modeling Accuracy
- ✓ 300% improvement in switching models
- ✓ <0.013 RMSE across device types
- ✓ <1Ω state acquisition error
- ✓ mV-level modulation precision