A Proof of Concept Neural Network Watchdog using a Hybrid Generative Classifier For Optimized Outlier Detection
With the continuous development of tools such as TensorFlow and PyTorch, Neural Networks are becoming easier to develop and train. With the expansion of these tools, however, neural networks have also become more black boxed. A neural network trained to classify fruit may classify a picture of a giraffe as a banana. A neural network watchdog may be implemented to identify such out-of-distribution inputs, allowing a classifier to disregard such data. By building a hybrid generator/classifier network, we can easily implement a watchdog while improving training and evaluation efficiency.
READ FULL TEXT