Neuromorphic Leakage Messaging Without Per-Neuron Time Storage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neuromorphic processors face challenges in managing leakage operations without requiring additional storage, particularly in asynchronous approaches that need storage for tracking the last update time of neuromorphic elements, leading to large storage requirements due to varying update intervals from microseconds to days.
Innovation Solution
A neuromorphic processor design that incorporates a message-based leakage mechanism, where leakage is computed from the current state value and correlated with the leakage message, eliminating the need for additional storage by using a leakage value monotonically related to the state value, and optionally clipped between minimum and maximum values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If asynchronous approach is used to compute leakage based on time since last update, then leakage computation flexibility is improved, but storage requirements increase due to need to store last update time for each neuromorphic element
Solution Approach 1:
The patent extracts the time-tracking function from individual neuromorphic elements and relocates it to a centralized time management unit. Each element no longer stores its own last update time; instead, the centralized unit maintains a single counter that tracks the current simulation time, eliminating the need for multiple storage locations while preserving asynchronous leakage computation.
Solution Approach 2:
The patent introduces a message passing mechanism as an intermediary between the centralized time management unit and individual neuromorphic elements. The time management unit sends current time information via messages to elements that need it, allowing elements to compute leakage asynchronously without directly accessing or storing time information themselves.
2Quantity of substance
If synchronous approach is used to compute leakage periodically, then storage requirements are reduced, but computation overhead increases due to periodic subtraction operations for all elements
Solution Approach 1:
The patent enables neuromorphic elements to compute their own leakage independently using the current time information received from the centralized time management unit. Each element calculates its leakage based on the difference between current time and its own last update time, eliminating the need for centralized periodic subtraction operations and reducing computation overhead.
3Adaptability or versatility
If time interval range is increased from microseconds to days, then temporal dynamic range is improved, but storage precision requirements increase to accurately represent such a wide range
Solution Approach 1:
The patent handles the wide temporal dynamic range by separating the time representation into two dimensions: a high-order time counter that tracks large time intervals (days) and fine-grained timing information that captures small intervals (microseconds). This dimensional separation allows the system to represent both very long and very short time intervals with adequate precision without requiring excessively large storage for a single time value.
Data Source
AI summary
A neuromorphic processor and a neuromorphic processing method are provided. The neuromorphic processor includes a plurality of neuromorphic elements, a message exchange facility, and a computation unit. The plurality of neuromorphic elements each has a respective state memory entry for storing their state. The message exchange facility enables neuromorphic elements to transmit neural event messages, and to receive transmitted neural event messages. A neural event message includes a message type indicating that the neural event message is an accumulation message or a leakage message. An accumulation message instructs the computation unit to modify a state value of a destination neuromorphic element by addition or subtraction of a value indicated as a message parameter in the accumulation message. A leakage message instructs the computation unit to modify the state value by reduction of the state value with a leakage value that is correlated with the state value.