Quetelet

Representation learning and probabilistic methods are huge sub-areas of modern machine learning, just take a look at the proceedings of ICLR2019.

These adversarial examples are generated in a very artificial way that will not be present in natural images (and if you’re thinking of security issues, the attacker needs access to your model...)

They’re still an interesting topic to explore but hardly evidence that neural nets don’t generalize.

You might be confusing the historical use of the sigmoid activation function with probabilistic modeling, neural networks in the 80s were used similarly to how they are today, albeit at a much smaller scale due to hardware limitations at the time.

The development of neural networks is a major contribution of the machine learning community, so even if you’d like to split hairs about whether the “computer is learning” (“learning” has a a precise technical definition by the way), NNs are not “just statistics.”

Actually most modern neural networks are not probabilistic, they are deterministic function approximators.

Also your point 3) isn’t quite correct either, often a “standard” architecture and training procedure (e.g. ResNet50 with Adam) will work on a new task with sufficient training data and minimal modification of the model.