Neuron Core With Time-Embedded Floating Point for Low-Power Synaptic Dynamics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital neuromorphic systems face challenges in implementing floating point precision neural and synaptic dynamics due to high energy costs and noise variability, while analog systems struggle with controlling and calibrating synaptic operations, leading to over-optimization and inefficient power consumption.
Innovation Solution
A time-embedded floating point unit is used to modulate synaptic weights into membrane potential values at a reference time, integrating time information into floating point numbers to reduce memory access and optimize power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If floating point arithmetic operations are used in digital neuromorphic systems to achieve precise neural and synaptic dynamics, then manufacturing precision and measurement precision are improved, but use of energy increases significantly
Solution Approach 1:
The patent pre-calculates and stores exponential decay factors in lookup tables during system initialization or idle periods. During runtime, neuron cores perform simple table lookups and multiplications instead of complex floating point arithmetic, significantly reducing energy consumption while maintaining precision
Solution Approach 2:
The patent divides the floating point arithmetic operation into separate functional components: exponential decay calculation, lookup table storage, and multiplication operations. This segmentation allows each component to be optimized independently, with the lookup table handling the computationally intensive exponential calculation offline
2Manufacturing precision
If memory access operations are performed frequently to update state variables and calculate temporal dynamics, then manufacturing precision is maintained, but use of energy increases due to redundant access
Solution Approach 1:
The patent pre-computes exponential decay factors and stores them in lookup tables before runtime. This preliminary action eliminates the need for frequent memory access during temporal dynamics calculation, as the decay factors are readily available in the lookup table
Solution Approach 2:
The patent combines the exponential decay calculation with the state variable update operation by using the pre-computed lookup table values directly in the update formula, eliminating separate memory access operations for reading decay factors
3Manufacturing precision
If analog neuromorphic systems are used to naturally support floating point precision synaptic dynamics, then manufacturing precision is improved, but reliability decreases due to noise and variability
Solution Approach 1:
The patent introduces lookup tables as an intermediary between the digital hardware and the continuous synaptic dynamics model. The lookup tables store pre-computed exponential decay factors that mediate between discrete digital operations and continuous temporal dynamics, enabling precise analog-like behavior in a digital system
Solution Approach 2:
The patent replaces the physical analog mechanisms that naturally provide floating point precision with a digital system using lookup tables and multiplication. This substitution eliminates noise and variability inherent in analog devices while maintaining the mathematical equivalence of continuous synaptic dynamics
Data Source
AI summary
Provided is a method of operating a neuron in a neuromorphic system. The method includes evaluating a membrane potential value at a corresponding time when receiving an input spike, time-modulating a synaptic weight of the membrane potential value and converting the time-modulated synaptic weight into a membrane potential value at a reference time, and generating an output spike when the membrane potential value at the reference time exceeds a certain threshold value. The membrane potential value at the reference time is represented by a floating point number including a predetermined bit of exponent and mantissa, and the floating point number includes time information. The method further includes accessing a memory and scanning a neural state variable when a timer is updated to “0” to update the neural state variable to an updated value at a reference time.