Inception-v4 is one of Google’s flagship neural networks, with a depth of 299 layers and a total of 41,264,932 trainable weights. It was developed for the ImageNet Large-Scale Visual Recognition Challenge in 2014. In the medical domain, pre-trained Inception-v4 networks have been successfully retrained to classify skin cancer, amongst other tasks.
Using our framework for the validation of machine learning libraries, we have recently validated the Keras / TensorFlow implementation of this huge machine learning network as a software of unknown provenance (SOUP). This validation is a mandatory pre-requisite for its use within a medical device. The solution consists of:
- a light-weight generic library that contains the specified tensor operations for each different layer type;
- an architecture description that assembles the required layers to a test oracle;
- a test data generator;
- a pytest test driver that compares the predictions of the system (model) under test with the test oracle’s prediction for the generated test data;
- and a textual SOUP validation plan describing the validation approach.
The Inception-v4 validation is a major break-through and proof-of-concept of our validation approach in two respects: It proves that the approach is applicable to deep neural networks of virtually any depth and size, and it exemplifies how easily and straightforward it can be applied to new model architectures.
Interested? Want to know more? Please contact us, we’ll be happy to give you more information and talk about your specific validation requirements