Local Target Signal Augmentation for Low-Power Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional CPUs are inadequate for processing-intensive tasks like machine learning due to power consumption and processing capability limitations, and there is a need for robust training data to enhance neural network performance, especially in voice-activated devices, while addressing user privacy concerns.
Innovation Solution
Local generation of extended training data through audio manipulation tools on neural network chips, using augmentation techniques to create new signal samples without transmitting user data externally, enhancing neural network robustness and reducing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If traditional CPUs are used for machine learning processing, then processing capability is maintained, but power consumption increases and processing capability becomes insufficient
Solution Approach 1:
The patent replaces traditional CPU-based machine learning processing with a specialized neural network chip that uses dedicated hardware circuits (including analog multipliers and accumulators) to perform neural network operations. This substitution of general-purpose mechanical computing with specialized neural hardware achieves superior processing capability for machine learning tasks while significantly reducing power consumption, as the neural chip is specifically optimized for these operations.
2Reliability
If more training data is collected externally, then neural network robustness improves, but user privacy concerns increase and data transmission requirements increase
Solution Approach 1:
The patent implements a self-service mechanism where the neural network chip performs self-training by generating synthetic training data through audio manipulation tools (pitch shifting, time stretching, noise addition) applied to locally stored target signals. This eliminates the need to externally collect and transmit user data, as the system autonomously generates diverse training samples from its own local data, thereby improving robustness while preserving user privacy.
Solution Approach 2:
The patent creates synthetic copies of target signals by applying various audio manipulations (pitch shifting, time stretching, noise addition) to original target signals stored locally. These copied and modified signals serve as additional training data, enabling the neural network to be trained on diverse samples without requiring external data collection or user privacy compromise.
3Power
If neural network chip contains more transistors for increased processing capability, then processing power increases, but chip complexity and manufacturing difficulty increase
Solution Approach 1:
The patent segments the neural network chip into distinct functional blocks including input interfaces, analog multiplier arrays, accumulators, weight memory, and output interfaces. Each block performs a specific function in the neural network processing pipeline, allowing the complex chip to be designed and manufactured through modular assembly of specialized components rather than as a monolithic complex circuit.
Solution Approach 2:
The patent implements universal computational units (analog multipliers and accumulators) that can be reused across multiple neurons and layers of the neural network. These multi-functional building blocks perform the same mathematical operations (multiplication and accumulation) for different neural network computations, reducing overall chip complexity compared to having dedicated circuits for each operation.
Data Source
AI summary
Provided herein is an integrated circuit for generating augmented training data, including a host processor configured to receive a signal stream. The integrated circuit also has a co-processor commutatively coupled to the host-processor includes a neural network with at least a first set of weights configured to identify one or more target signals from the signal stream received from the host processor. A plurality of augmentation tools are also accessible to the integrated circuit. Finally, the integrated circuit, coupled computing device or other suitable digital signal processor stores a plurality of the one or more identified target signals, and upon reaching a predetermined threshold of identified target signals, utilizes the plurality of augmentation tools to generate an extended set of target signals, and generates a second set of weights for the neural network based on the extended set of target signals generated by the augmentation tools.