Quanta has a fair round up of recent advances in deep learning. Most interesting is the recent performance on natural language understanding tests that are close to or exceed mean human performance. Inevitably, John Searle’s Chinese Room argument is brought up, though the author of the Quanta article suggests that inferring the Chinese translational rule book from the data itself is slightly different from the original thought experiment. In the Chinese Room there is a person who knows no Chinese but has a collection of translational reference books. She receives texts through a slot and dutifully looks up the translation of the text and passes out the result. “Is this intelligence?” is the question and it serves as a challenge to the Strong AI hypothesis. With statistical machine translation methods (and their alternative mechanistic implementation, deep learning), the rule books have been inferred by looking at translated texts (“parallel” texts as we say in the field). By looking at a large enough corpus of parallel texts, greater coverage of translated variants is achieved as well as some inference of pragmatic issues in translation and corner cases.
As a practical matter, it should be noted that modern, professional translators often use translation memory systems that contain idiomatic—or just challenging—phrases that they can reference when translating new texts. The understanding resides in the original translator’s head, we suppose, and in the correct application of the rule to the new text by checking for applicability according to, well, some other criteria that the translator brings to bear on the task.
In the General Language Understand Evaluation (GLUE) tests described in the Quanta article, the systems are inferring how to answer Wh-style queries (who, what, where, when, and how) as well as identify similar texts. The collection of questions is designed to be linguistically challenging, from anaphora resolution to morphological negation, from monotonicity to world knowledge. And the machines keep getting better and better, partially through pretraining of the networks on large corpora to just build a baseline system, but also through clever architectural advances like attention moving mechanisms that allow for the scanning of the text region for clues in a bidirectional fashion, with the attention behavior itself trained into the network along with the recognition of the surface features (words, primarily) of the text.
Yet we can easily scoff at the intelligence present in the systems. They are certainly useful. I asked my watch for the dates of World War I this morning because of the sudden appearance of American flags along the central drag of the town. Siri gave me a fine answer. And my car can self-drive in a limited fashion. But other than basic procedural knowledge isolated to a given domain, all of the hallmarks of human intelligence, including behavioral plasticity in the face of variable challenges and the situational knowledge that derives from growing up in the world and a culture, are largely missing.
So we have an uncanny valley situation, in a way, where the deep learning networks are approaching human capability on a narrow task but it’s uncertain whether they can be anything more than useful tools. Perhaps that’s enough and saves us from a future where AIs are so canny we can’t distinguish them from real people.
But what might it take to move out of that valley? Let’s assume that Strong AI is true. That is that a computer programmed to function like a human mind and with normal inputs is, in fact, indistinguishable from a human mind. The central problem with the implementation must be contingency. That is, human minds are biological and contingent on their natural history and that a given mind is contingent on its cultural and personal history of maturation. Any given, practical, attack on an aspect of human cognition is not built in the same way. It can be useful but is destined to be merely an uncanny approximation. The real thing is possible but must be so narrowly attached to a biological architecture and all the complexities of growing in a world that the possibility is very low.
This perhaps corresponds to Searle’s claim that human intelligence must be associated with brains, but with the caveat that simulated brains with sufficient accuracy and complex histories might pull it off or get close enough that they are like Dan Dennett’s zombie humans. In the most likely outcome, however, we will be forever in the uncanny valley.