Ren Tianling's team from the School of Integrated Circuits proposed an adaptive spatiotemporal information processing paradigm inspired by the attention mechanism.
Ren Tianling's team from the School of Integrated Circuits proposed an adaptive spatiotemporal information processing paradigm inspired by the attention mechanism.
A research team from Tsinghua University’s School of Integrated Circuits has developed an attention-inspired in-memory computing device that mimics the human brain’s ability to efficiently process both temporal and spatial information with low energy consumption.
Inspired by the brain’s attention mechanism, particularly the frontoparietal network’s dynamic filtering of sensory input. The team created a 0D-2D heterodimensional interface using MoS₂ channels and Ag conductive filaments. This structure enables nonvolatile storage of past data and analog computation with adjustable weights between current and stored information, allowing adaptive, real-time attention allocation.
The device performs spatiotemporal processing directly in memory, reducing hardware overhead and power use. A 5×5 array was fabricated and tested to demonstrate its ability to dynamically prioritize moving (point A) versus stationary (point B) objects by adjusting gate voltages—controlling the intensity of temporal and spatial signals.
Simulations show the system can support real-time, context-aware attention allocation in autonomous driving scenarios, enabling roadside and vehicle-side systems to respond adaptively to changing environments with full-range (0–100%) flexibility.
This breakthrough paves the way for energy-efficient, low-power edge intelligence applications—especially in dynamic fields like autonomous vehicles.
Inspired by the brain’s attention mechanism, particularly the frontoparietal network’s dynamic filtering of sensory input. The team created a 0D-2D heterodimensional interface using MoS₂ channels and Ag conductive filaments. This structure enables nonvolatile storage of past data and analog computation with adjustable weights between current and stored information, allowing adaptive, real-time attention allocation.
The device performs spatiotemporal processing directly in memory, reducing hardware overhead and power use. A 5×5 array was fabricated and tested to demonstrate its ability to dynamically prioritize moving (point A) versus stationary (point B) objects by adjusting gate voltages—controlling the intensity of temporal and spatial signals.
Simulations show the system can support real-time, context-aware attention allocation in autonomous driving scenarios, enabling roadside and vehicle-side systems to respond adaptively to changing environments with full-range (0–100%) flexibility.
This breakthrough paves the way for energy-efficient, low-power edge intelligence applications—especially in dynamic fields like autonomous vehicles.
