Gary Marcus

I think it is quite refreshing that the community is working hard and honestly to quantify the limitations of the current SOTA. I’m not a die-hard advocate of DNNs. Let’s turn this from negative to positive. Gary, I’d love to hear more about what you advocate as the way forward.

It's a bit like those reinforcement learning algorithms that stop working if you simply rescale the reward function by a constant. It's outrageous that these algorithms don't satisfy basic invariance requirements

I'm surprised this is news. Another way to confuse a DNN is to reclassify a screen grab of an image. The .jpg compression often caused the original and the screengrab to be classified differently when I tried this (though I can't remember which net I used)