While it has many benefits, this concept also has a number of limitations. First of all, the limited computing power of devices means that some of the most powerful models cannot be considered.
Also, many on-device deep learning frameworks are not compatible with the most innovative or the most complex layers. For instance, TensorFlow Lite is not compatible with custom LSTM layers, making it hard to port advanced recurrent neural networks on mobile using this framework.
Finally, making models available on devices implies sharing the weights and the architecture with users. While encryption and obfuscation methods exist, it increases the risk of reverse engineering or model theft.
