Selective Deterministic Computations for Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional neural networks, particularly convolutional neural networks, are computationally expensive due to the large number of operations required, leading to significant resource consumption in data centers, necessitating more efficient models.
Innovation Solution
The implementation of selective deterministic computations based on statistical models generated from historical data, which indicate the probability of value selection by layers such as ReLU and pooling layers, allowing for the omission of unnecessary convolution operations and potential removal of selection layers, thereby reducing computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional neural networks perform all deterministic computations, then computational accuracy is maintained, but computing resource consumption increases significantly
Solution Approach 1:
The patent applies partial action by selectively performing only those deterministic computations that are likely to be selected by subsequent selection layers (ReLU or pooling layers). The system uses statistical models to determine which computations are necessary, avoiding unnecessary operations while maintaining accuracy for critical computations.
Solution Approach 2:
The patent implements preliminary action by pre-computing and storing statistical models that indicate the probability of selection for each deterministic computation. These models are generated in advance using historical data, allowing the system to make informed decisions about which computations to perform during actual inference without requiring real-time analysis of all possible operations.
2Reliability
If all convolution operations are performed, then neural network accuracy is maintained, but computational time increases
Solution Approach 1:
The system performs only the subset of convolution operations that are statistically likely to produce selected outputs. By using statistical models to predict selection probabilities, the system identifies and executes only the necessary computations, thereby reducing total computational time while maintaining accuracy for critical features.
Solution Approach 2:
Statistical models are pre-computed using historical data to determine which convolution operations are likely to be selected by subsequent layers. This preliminary analysis allows the system to optimize computation during actual inference by skipping operations with low selection probabilities, thus improving analysis speed without sacrificing accuracy.
3Productivity
If statistical models are generated from historical data, then selective computation performance is improved, but data processing requirements increase
Solution Approach 1:
The system changes the parameter approach by transforming raw historical data into statistical models that capture selection probabilities. Instead of processing all historical data individually during inference, the system uses these pre-computed statistical representations, which summarize the essential patterns from historical data in a more compact and efficient format.
Data Source
AI summary
Selective performance of deterministic computations for neural networks is disclosed, including: obtaining a statistical model for a selection layer of the neural network, the statistical model indicating probabilities that corresponding values are selected by the selection layer, the statistical model being generated using historical data; selectively performing a subset of a plurality of deterministic computations on new input data to the neural network, the plurality of deterministic computations being associated with the deterministic computation layer, the selective performance of the deterministic computations being based at least in part on the statistical model and generating a computation result; and outputting the computation result to another layer in the neural network.