Core Advantage: RC is highly energy-efficient because it only requires training the readout layer rather than updating all the weights across a traditional neural network.
The Challenge: Standard RC implementation on hardware is an engineering nightmare, comparable to “mapping a random/dense forest onto a map”.
The Solution: Shifting focus to a Simple Cycle Reservoir (SCR) architecture, which is significantly easier to implement on hardware.
Novelty Direction: While prior works successfully implemented an SCR, it relied on an off-chip readout. Developing an on-chip readout provides a distinct, novel research direction.
2. Readout Algorithm & Hardware Architecture
Algorithm Selection:QRD-RLS (an evolution of LMS used in active noise cancellation headphones) is identified as a strong fit for the readout algorithm. Theoretically, it serves as a “gold standard” for AI, despite not being commonly used.
Hardware Mapping: The QRD-RLS algorithm can be easily mapped to hardware using Systolic Arrays (grid-like structures similar to Google’s TPU).
3. Domain Shift: Analog vs. Digital Domain
Analog Constraints: Though an analog SCR architecture is great for ultra-low energy, the Analog-to-Digital Converter (ADC) block presents a major bottleneck. Furthermore, previous analog implementations (like Hokkaido Univ. operating at a subthreshold 1kHz) suffer from the memory-wall problem because the readout sits off-chip, making ultra-low power claims somewhat misleading (similar to issues encountered by IBM Northpole).
Digital Advantages: Implementing the SCR entirely in the digital domain bypasses the ADC hurdle since everything shares the same clock domain and allows for pre-hardware simulation.
4. Current Focus & Next Steps
Non-Linearity Optimization: The primary immediate task is designing and implementing non-linearity to ensure rich ripples .
Data Format Evaluation: Verifying design viability by running comparative tests between Fixed Point (FXP) and Floating Point (FP) formats.
Hyperparameter Tuning via Genetic Algorithm (GA): Exploring GAs to run tests and breed parameters for optimal results. This includes tuning hyperparameter () to make some nodes linear while others remain non-linear, saving energy by avoiding unnecessary non-linear activations.
