Deep Learning for Training with Noise in Expert Systems
(Room S320C)
29 Nov 17
8:30 AM
-
9:00 AM
Deep learning systems have achieved a remarkable human-machine parity for some key training tasks, including machines that can speak, listen and see at an expert human level. For image recognition, the state-of-the-art features deep convolutional neural networks (CNN). This paper benchmarks their performance on key recognition objectives: 1) automating large-scale image classification, and 2) identifying types of noisy training data that can improve testing outcomes. This benchmarking relies on two classically hard image problems: 1) distinguishing similar objects, and 2) classifying tiny low-resolution icons. The teacher-based recognition embodies supervised learning since image class labels are known. When deployed using Google’s Tensorflow framework, the CNN learns both similar classes and tiny icons with 95.3% and 89.5% accuracy respectively. This paper further explores two competing hypotheses of training noise. The first role for noise may improve outcomes if the noise reduces overfitting. The competing role, however, may diminish learning if noise masks some key object features. We inject quantifiable impulse (or spike) noise to disrupt local object patterns (convolution) and to benchmark learning changes. This choice of localized interference attacks a fundamental assumption underpinning CNN performance, namely that neighboring pixels dominate remote ones. One surprising outcome is that by adding impulse noise to training images, overall classification improves compared to training on unmodified test images (95.4%). The basic principle can be understood as noise benefiting image-based training when it augments data size and diversity or when it obscures background relative to foreground targets. When we apply convolutional neural nets to large-scale image classification, the accuracy compares favorably to the state-of-the-art in published literature and public global data competitions. At least in the case of some multi-class image problems, CNN accuracy e