Symbiotic Hybrid Neural Network Watchdog For Outlier Detection
This addresses the issue of neural network reliability for users in safety-critical applications, but it is incremental as it builds on existing watchdog methods.
The paper tackled the problem of neural networks misclassifying out-of-distribution inputs by comparing separate versus symbiotic watchdog architectures for outlier detection, finding that the symbiotic approach performs better empirically.
Neural networks are largely black boxes. A neural network trained to classify fruit may classify a picture of a giraffe as a banana. A neural network watchdog's job is to identify such inputs, allowing a classifier to disregard such data. We investigate whether the watchdog should be separate from the neural network or symbiotically attached. We present empirical evidence that the symbiotic watchdog performs better than when the neural networks are disjoint.