The author of the article Does AI Truly Learn And Why We Need to Stop Overhyping Deep Learning argues that deep learning (DL) is overhyped because DL systems don’t actually think.
For instance, he says this about the way researchers describe their systems:
In contrast, data scientists all too often treat their algorithmic creations as if they were alive, proclaiming that their algorithm “learned” a new task, rather than merely induced a set of statistical patterns from a hand-picked set of training data under the direct supervision of a human programmer who chose which algorithms, parameters and workflows to use to build it.
I certainly see the intent behind this. It’s a way to caution against the attribution of human-like qualities to a series of differential equations and engineering optimizations. This sentence most clearly illustrates his intent:
Why does this matter? It matters because as we increasingly deploy AI systems into mission critical applications directly affecting human life, from driverless cars to medicine, we must understand their very real limitations and brittleness in order to properly understand their risks.
Yes, it’s true. The idea that these systems, as they are deployed in more and more places and affect the lives of millions, need to be scrutinized and double-checked for fairness, accountability, and transparency. Some good research is being done in this space every day. This research area greatly interests me, and I keep up with it as much as I can.
On the other hand, it is also important to consider the language with which we describe the differences between learning machines and sentient humans. To hold up the machines in contrast with humans and declare the machines to be “mundane piles of code,” and by contrast, humans to have some indescribable essence, is fraught territory.
The author of the post seems to be making a similar argument to the Chinese Room. The Chinese Room is a thought experiment by philosopher John Searle in which a person is, sadistically, locked in a room with a big book. The book contains Mandarin grammar rules, but no vocabulary or translations to English. One day, someone slides a piece of paper covered in Mandarin writing under the locked door. The prisoner uses the book of grammar rules writes a reply, even though they don’t have a clue what they’re writing. They slide it out through the bottom of the door. Searle argues that no new understanding has been created inside this room, but the room has spit out a cogent reply. By analogy, this is what a machine does when it outputs something that humans interpret to have meaning. He argues that the machine doesn’t really understand what it’s doing like a human would.
This argument has generated a ton of replies from other philosophers: Replies to The Chinese Room Argument (Stanford Encyclopedia of Philosophy)
I take a number of issues with the Chinese Room argument:
- This thought experiment was conceived in the 80s, and artificial intelligence has come a long way since then. The thrust of Searle’s argument was aimed against symbol manipulation. Symbol manipulation and purely syntactic methods, especially in natural language processing, is at best a lost art and at worst a research dead end.
- The Chinese Room argument is underpinned with a bit of dishonesty that rubs me the wrong way. I feel that it starts with a conclusion and uses an illustrative analogy to make a point without citing anything else. Though it’s a compelling story about a prisoner with a rulebook, it’s not robust scientific thinking.
- The argument fails to adequately define the following terms: mind, understanding, and meaning. Because of this, it’s easy for Searle to move the goalposts of the argument such that any reply could be dismissed with, “But that’s not real understanding,” or “But I’m talking about real meaning.” The lack of definitions are a crutch to avoid really considering counterarguments.
- The argument posits that there’s something special about the substrate of organic brains that set them apart from other materials like rooms, rulebooks, and slips of paper. It’s an organocentric (yeah, I just coined it) view that I don’t believe will stand up to scrutiny the further we move into the future.
Overall, the author of the Forbes post has the right intent. It’s important to remember that machine learning systems do not have morals or intent, and are tools we use to make existing systems more efficient. However, it’s also important not to mischaracterize what they are in comparison with human minds. They might be a key pathway to understanding what understanding is.