News for hackers. Stuff that matters.
Re:Where we see shapes, AI sees textures by w_t_payne on Monday July 08, 2019 @01:33PM (#20384805)
Depends on the training data.

These networks are correlation-finding machines, and will simply latch onto the simplest correlation possible that produces the correct results. Only by explicitly controlling the correlations in the training data can you force the network to focus on the attributes that you wish.

This is why domain-randomization (such as the approach that NVIDIA is taking) helps support generalization -- it's a very direct and efficient form of regularization, removing correlations that do not generalize, forcing the network to work harder to find correlations that do generalize.

This also has the added benefit of being something that we can tie back to requirements and physical properties, an important ingredient in making these systems understood and safe.

There are no comments on this comment yet.