Neuromorphic Synapse Memory Using Split-Precision Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neuromorphic processors require large memory usage for storing synaptic weights due to the vast number of synapses, leading to increased chip size.
Innovation Solution
Decompose synaptic weights into presynaptic and postsynaptic weights, where presynaptic weights are stored with a higher precision and postsynaptic weights with lower precision, reducing memory requirements while maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If synaptic weights are stored with high precision to maintain data processing accuracy, then measurement precision is improved, but memory usage and chip area increase
Solution Approach 1:
The patent segments the synaptic weight into two components: a presynaptic weight matrix (M×1) containing presynaptic weights, and a postsynaptic weight matrix (M×N) containing postsynaptic weights. This segmentation allows different precision levels to be applied to different parts of the weight representation, reducing overall memory requirements while maintaining computational accuracy.
Solution Approach 2:
The patent applies local quality by storing presynaptic weights with higher precision (first number of bits) and postsynaptic weights with lower precision (second number of bits, where second number < first number). This differential precision allocation optimizes memory usage by recognizing that different components of the weight decomposition have different importance for accurate computation.
2Quantity of substance
If the number of bits for storing synaptic weights is reduced to decrease memory usage, then memory usage and chip area are reduced, but data processing accuracy deteriorates
Solution Approach 1:
By segmenting the weight representation into presynaptic and postsynaptic components stored in separate matrices, the system can allocate different bit depths to each component. This segmentation enables reduced precision for postsynaptic weights while maintaining sufficient accuracy through the presynaptic weight component.
Solution Approach 2:
The patent changes the precision parameter differentially across the weight matrices. The presynaptic weight matrix uses a first number of bits while the postsynaptic weight matrix uses a second, smaller number of bits. This parameter change optimizes the balance between memory efficiency and computational accuracy.
3Quantity of substance
If synaptic weights are decomposed into presynaptic and postsynaptic weights to reduce memory usage, then memory usage is reduced, but device complexity increases
Solution Approach 1:
The decomposition of synaptic weights into presynaptic and postsynaptic weights creates a segmented matrix structure. While this increases structural complexity, it enables more efficient memory usage and can be implemented through standard matrix operations in the neuromorphic processor architecture.
Data Source
AI summary
A method for processing data based on a neural network including a first layer including axons and a second layer including neurons, includes receiving synaptic weights between the first layer and the second layer; generating presynaptic weights, a number of which is identical to a number of the axons, and postsynaptic weights, a number of which is identical to a number of the synaptic weights, from the synaptic weights; and storing the presynaptic weights and the postsynaptic weights in a synapse memory.