Deep learning, which modifies the causation of artificial neural networks (ANNs), has big problems – for example, see the links below. Such ANNs are used in self-driving vehicle AI control systems. In further posts I’d like to discuss possible reasons for these problems.
Mistaken identification:
mistaking a coffee machine for a cobra
mistaking a toy turtle of a rifle
mistaking a rifle for a helicopter
mistaking a building for an ostrich and a panda for a gibbon
mistaking a STOP sign for something that’s not a STOP sign
explaining the mistakes of neural networks… (PDF)
deep learning networks are easily fooled (PDF)
unawareness of deep learning mistakes
deep reinforcement learning doesn’t work yet
White-box testing
thousands of errors in self-driving cars
Black Box
We don’t know how AI deep learning generates its output. We can’t tell how well an ANN will classify its inputs by examining the ANN.
associative-ai.com