A team of researchers has introduced a revolutionary heuristic compression technique for convolutional neural network models, incorporating three traditional reduction methods in a specific order: integer quantization, network slimming, and deep compression. This innovative approach autonomously determines the optimal network model size through iterative margin calculations. The study, published in IEEE Access (2024), demonstrates the potential of this method to enhance AI technologies.
Artificial intelligence (AI) systems used for image recognition are inspired by human vision and neural processes. To reduce computational requirements, researchers have identified three key methods: network slimming, deep compression, and integer quantization. Previously, the application ratio of these methods was determined through trial and error.
The University of Tsukuba researchers have developed an algorithm that automatically identifies the ideal proportion of each reduction method, aiming to reduce power consumption in AI technologies and facilitate semiconductor miniaturization. Convolutional neural networks (CNNs) play an essential role in applications like facial recognition and object detection.
By minimizing data bits in computations, CNNs can maintain accuracy while significantly reducing computational demands. The new algorithm establishes the optimal sequence of reduction methods as IQ, NS, and DC, with an autonomous determination of application ratios. This advancement allows CNNs to be compressed up to 28 times smaller and 76 times faster than previous models.
This research breakthrough is expected to revolutionize AI image recognition technology by reducing computational complexity, power consumption, and the size of AI devices. The implications of this study can enhance the feasibility of deploying advanced AI systems in various industries.
