Phase-Summing Injection Circuit for PVT-Robust Neural Readout
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solutions for artificial neural networks face challenges in reducing power consumption and footprint while being sensitive to PVT variations and requiring complex auxiliary circuits for frequency-domain read-out.
Innovation Solution
A phase-domain summing circuit using a locked oscillator per injection interface circuit, where synchronization signals are phase-shifted to encode input and output signals, and synaptic weights are adjusted via injection currents or local impedances, allowing for efficient power usage and reduced footprint.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If phase-locked loop circuits are used for oscillatory neural networks, then robustness to PVT variations is improved, but footprint increases due to passive loop filters
Solution Approach 1:
The patent extracts and removes the passive loop filter component from the phase-locked loop circuit, retaining only the essential phase-comparison and feedback functionality. This extraction eliminates the large footprint requirement while preserving the robustness to PVT variations through phase-differential measurement.
Solution Approach 2:
The patent uses a simplified model that copies only the essential functionality of phase-locked loops (phase comparison and feedback) without replicating the full complex circuitry including passive filters. This allows achieving similar robustness with reduced footprint.
2Use of energy by stationary object
If charge redistribution digital-to-analogue converters are used for weighted sum computation, then consumption and footprint are improved, but sensitivity to PVT variations increases
Solution Approach 1:
The patent replaces the charge redistribution mechanism (electrical/mechanical system) with a phase-domain computational approach using oscillators and phase-differential measurements. This substitution maintains low consumption characteristics while improving robustness to PVT variations through frequency-based operation.
Solution Approach 2:
The patent changes the operational parameter domain from charge/voltage (analog domain sensitive to PVT) to phase/frequency domain (more robust to PVT). By encoding information in phase differences rather than voltage levels, the system achieves both low consumption and PVT robustness.
3Quantity of substance
If nano-magnetic oscillators are used for oscillatory neural networks, then integration density and energy efficiency are improved, but read-out complexity increases due to frequency-domain output
Solution Approach 1:
The patent introduces phase-differential measurement as an intermediary mechanism that translates the frequency-domain output of nano-magnetic oscillators into a more easily exploitable form. By measuring phase differences between oscillator outputs, the system simplifies the read-out process while maintaining high integration density and energy efficiency.
Solution Approach 2:
The patent changes the read-out parameter from direct frequency measurement (complex) to phase-differential measurement (simpler). This parameter transformation maintains the benefits of high-frequency oscillator operation while reducing read-out complexity through differential phase comparison.
4Power
If remote cloud servers are used for complex AI computations, then computing power is improved, but communication volume, data security, and latency are worsened
Solution Approach 1:
The patent segments the AI computing system into local edge devices capable of performing neural network computations and remote cloud servers. By implementing oscillatory neural networks with efficient phase-domain computation at the edge, the system reduces the volume of data that needs to be communicated with the cloud, addressing both energy loss and security concerns while maintaining computational capability.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The proposed solution significantly decreases power consumption by encoding information in the phase domain, reduces footprint by eliminating the need for phase-locked loops, and enhances robustness against PVT variations.
Implementation Method 1
The circuit according to the invention is based on one locked oscillator per a plurality of injection interface circuits that are controlled by synchronization signals
Data Source
AI summary
A multi-injection phase-summing circuit includes an oscillator having a natural frequency of oscillation F0, a reference injection interface circuit controlled by a reference synchronization signal at a reference frequency and at least one additional injection interface circuit controlled by a secondary synchronization signal at the reference frequency, each injection interface circuit having a variable injection parameter, the secondary synchronization signals each being phase shifted with respect to the reference synchronization signal, the oscillator being configured to generate an output signal at the reference frequency and phase shifted with respect to the reference synchronization signal by a phase shift dependent on a sum of the respective phase shifts of each secondary synchronization signal with respect to the reference synchronization signal, said sum being weighted by the injection parameters.