Researchers develop novel method for compactly implementing image-recognizing AI
What specific AI technology does the article focus on, and what are its primary applications?
The article focuses on a specific AI technology called convolutional neural networks (CNNs), which are primarily used in applications such as facial recognition at airport immigration and object detection in autonomous vehicles. CNNs are composed of convolutional and fully connected layers, with the former simulating human vision and the latter enabling the brain to deduce the type of image from visual data.
What are the three known methods to reduce data usage in AI image recognition as mentioned in the article?
The three known methods to reduce data usage in AI image recognition, as mentioned in the article, are:
- Network Slimming (NS): This method aims to minimize the visual components in the image recognition process.
- Deep Compression (DC): This method focuses on reducing the neuronal components, which are involved in processing and interpreting visual data.
- Integer Quantization (IQ): This method decreases the number of bits used in computations, thereby reducing the overall data usage.
These three methods, when applied in the optimal sequence (IQ followed by NS and DC), can significantly reduce computational complexity, power consumption, and the size of AI semiconductor devices, enhancing the feasibility of deploying advanced AI systems.
How does the new algorithm developed by researchers at the University of Tsukuba determine the optimal application ratio of these methods?
The new algorithm developed by researchers at the University of Tsukuba determines the optimal application ratio of the three reduction methods (network slimming, deep compression, and integer quantization) by automatically identifying the optimal proportion of each method3. This removes the necessity for trial and error in determining the order of implementation or allocation of these methods. The algorithm enables a convolutional neural network to be compressed to 28 times smaller and 76 times faster than previous models, significantly reducing computational complexity, power consumption, and the size of AI semiconductor devices.