Temporal Convolution Neural Networks With Quantized Filters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Recurrent neural networks (RNNs) are not heavily quantizable, which limits their performance on mobile processing units, and they consume more energy compared to neural networks with temporal convolutions, despite processing temporal information effectively.
Innovation Solution
Implementing neural networks with temporal convolutions and quantizing filter coefficients and activations, allowing for reduced bit representation, which reduces memory and energy consumption, enabling efficient processing of time series data without the need for recurrent links.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Recurrent Neural Networks (RNNs) are used to process temporal information, then the ability to process temporal dimension data is improved, but energy consumption increases and heavy quantization is not feasible
Solution Approach 1:
The patent changes the fundamental parameter of the neural network architecture from recurrent connections to temporal convolutions with quantized parameters. By transforming the network structure and applying quantization to filter coefficients and activations, the system achieves efficient temporal processing with reduced energy consumption suitable for mobile devices.
Solution Approach 2:
The patent replaces the mechanical recurrent connection mechanism (which requires storing and reusing hidden states across time steps) with a convolutional mechanism that processes temporal sequences through filtered operations. This substitution eliminates the need for complex recurrent link storage and computation, reducing energy requirements.
2Adaptability or versatility
If Recurrent Neural Networks (RNNs) are used to process temporal information, then the ability to process temporal dimension data is improved, but quantization capability deteriorates
Solution Approach 1:
The patent transforms the network architecture to use temporal convolutions instead of recurrent connections, enabling the parameters (filter coefficients and activations) to be effectively quantized. This parameter transformation allows the system to achieve heavy quantization (reducing precision to save memory) while maintaining temporal processing capabilities.
3Loss of energy
If neural networks with temporal convolutions and quantized parameters are used, then energy consumption and memory requirements are reduced, but computational intensity of convolution increases
Solution Approach 1:
The patent applies quantization to the convolution parameters (filter coefficients and activations), reducing their precision requirements. This allows the computationally intensive convolution operations to be performed with lower precision arithmetic, reducing the actual power consumption despite the high computational intensity of the convolution itself.
Data Source
AI summary
A machine learning system, including at least one temporal filter. An input variable, encompassing a chronological sequence of images, is processed with the aid of the machine learning system, using the filter. The machine learning system is configured to use the filter on a sequence of pixels, which are all situated at identical coordinates of the images, or at identical coordinates of intermediate results. Filter coefficients of the filter are quantized. A method, a computer program, and a device for creating the machine learning system are also described.