When additional data is much more numerous than target-distribution data, limited computational resources can make it expensive to model both sources well. Down-weighting the auxiliary source can reduce the need for a very large neural network while still using the additional data.

Training Data Matching Matters More When Model Capacity Is Limited

Adding internet images can force the neural network to expend some capacity learning properties specific to internet images, such as higher resolution or different framing. If those properties differ greatly from mobile-app images, less capacity remains for recognizing mobile-app data, and this could hurt performance.

Google

If a sufficiently large neural network is trained long enough on both mobile-app images and internet images, there is no theoretical harm in trying to make the algorithm do well on both sources.

Using Additional Internet Images with a Large Enough Neural Network

Machine Learning Yearning, a free ebook from Andrew Ng, teaches you how to structure Machine Learning projects.
https://www.deeplearning.ai/machine-learning-yearning/

Note: The content of the book is aligned with the Coursera Deeplearning.ai specialization.  https://www.deeplearning.ai/deep-learning-specialization/ 

Machine Learning Yearning (Deeplearning.ai)

Additional internet images give the neural network more examples of what cats do or do not look like. Because internet images and user-uploaded mobile-app images share some similarities, the neural network can apply some knowledge acquired from internet images to mobile-app images.

Learn Before

Related

Learn After