Where to find information about models, data and labels used by Image/AudioIdentify?

Question

I am thinking about using Mathematica 12's ImageIdentify and AudioIdentify functions for an interdisciplinary research project. The social science part of it requires to describe a little bit the learning phase (architecture of the neural nets, datasets used, etc) and, more importantly, the list of the labels used so we can clearly identify the premises of the models in terms of what is considered by the function, and what is not.

I can't find such information in Wolfram's documentation. Is there a place where it can be found ? If not, how researchers usually justify the use of high-level function of Mathematica in their papers ? (I have in mind the way scikit-learn or networkx Python libraries tend to list academic references of the choices they made to implement a specific function).

Thanks in advance for you help !

Answer 1

I think you are looking for the Neural Network Repository

Here are the specific links for AudioIdentify and ImageIdentify.

These are the raw networks used underneath ImageIdentify and AudioIdentify. I am not aware of any papers on the arXiv or elsewhere that explain the underlying networks. However, you can download the "construction notebook" for the network, which specifies how the network is built, from the bottom of the page. These pages also have examples of how to train, transfer learn, and in some cases use and extend the networks, as well as the specific labels used by the trained networks (or rather, the datasets the networks were trained on). I believe "Reference" at the bottom is the preferred reference URL and style.