Imagine for a moment that a mad scientist is able to grow a human brain in the lab.
Once grown, this brain knows nothing about objects in the physical world. It has had no physical experiences. It has had no interactions with anything or anyone else.
The mad scientist then wires this know-nothing brain to a computer. The computer feeds RGB images from the ImageNet ILSVRC2012 dataset, one image at a time, again and again, to the brain's visual system. Upon feeding each image, the mad scientist measures which of 1000 previously chosen neurons in the brain's cortex fire, each neuron corresponding to exactly one label in the dataset. When the right neuron fires, the scientist activates the brain's pleasure centers; if the wrong neurons fire, the scientist activates the brains's pain centers.
After seeing each image multiple times, one can imagine, the know-nothing brain will learn to activate the correct neuron associated with each image label well enough to get high accuracy on a validation dataset consisting of previously unseen images.
But that poor brain will still know nothing about objects in the physical world.
Substitute "human brain" for "convolutional neural network," and you obtain a remarkably accurate description of how neural nets learn to recognize images.
we often discuss AI in terms of “smartness” which is thought to be inherent to the architecture. But maybe we should talk in terms of “experience”. When will AI surpass human beings in quantity/quality of lived experiences?
That is a very succinct way of expressing the same idea. Many AI researchers understand this. Note that, unlike human beings, the experience that AIs are accumulating via interactions with the real world will be digitally saved and gradually accumulated, and built upon, bit by bit, such that over time, AIs will be able to draw on the accumulated experience of many prior AIs. We can already see this happening today, for example, with the wide availability of pretrained vision and language models that incorporate the experience of AIs developed and trained by others in the past.
Completely irrelevant, but HN is one of the few places online in which you can use the verb to cons, and most people will understand what you mean, even though this verb isn't part of any spoken language.
Once grown, this brain knows nothing about objects in the physical world. It has had no physical experiences. It has had no interactions with anything or anyone else.
The mad scientist then wires this know-nothing brain to a computer. The computer feeds RGB images from the ImageNet ILSVRC2012 dataset, one image at a time, again and again, to the brain's visual system. Upon feeding each image, the mad scientist measures which of 1000 previously chosen neurons in the brain's cortex fire, each neuron corresponding to exactly one label in the dataset. When the right neuron fires, the scientist activates the brain's pleasure centers; if the wrong neurons fire, the scientist activates the brains's pain centers.
After seeing each image multiple times, one can imagine, the know-nothing brain will learn to activate the correct neuron associated with each image label well enough to get high accuracy on a validation dataset consisting of previously unseen images.
But that poor brain will still know nothing about objects in the physical world.
Substitute "human brain" for "convolutional neural network," and you obtain a remarkably accurate description of how neural nets learn to recognize images.