Thread by @emilymbender

Some thoughts from @GaryMarcus on overclaims around deep learning (and how things could be different). A few reactions from me:

nautil.us/deep-learning-…

1/

@GaryMarcus First, I had a good giggle at the thought of someone trying to implement something like the copy/paste functionality of an OS via deep learning. How frustrating would it be for some non-trivial % of the time for the info being pasted to be randomly different?

2/

@GaryMarcus Second, I disagree that photo labeling (i.e. assigning words to images) is necessarily "low stakes". Real harm can come from such systems, when the (photo, label) pair reproduces racism & other -isms. Key work here by @Abebab and colleagues:

arxiv.org/abs/2110.01963

3/

@GaryMarcus @Abebab See also this paper by @Abebab
and @vinayprabhu
ieeexplore.ieee.org/abstract/docum…

4/

@GaryMarcus @Abebab @vinayprabhu Finally, I think a really key point is that symbols are *signs*: pairings of signifier and signified. I believe what @GaryMarcus means by "algebra" in this context is rules for manipulating signifiers so that the resulting signifieds are also valid.

5/

The lack of such rules in large language models paired with the fact that they are manipulating signifiers that we (humans) know how to pair with signifieds is part of what makes them dangerous. (See esp Sec 6 of Stochastic Parrots: dl.acm.org/doi/pdf/10.114… )

6/

But of course, when we talk about manipulating the signifiers of natural language (rather than e.g. arithmetic) the constraints on validity include not just the syntax/semantics of the NL, but also whatever domain is being talked about.

7/

That is an enormous task, and quite possibly an inherently infeasible one. Though I agree that if it is approachable at all, something that combines the strengths of symbolic methods with data-driven search for generalization is a likelier candidate than either alone.

8/

That said, I personally have no interest in working towards so-called "AGI". I think we'd be better off focusing on building effective tools for human use rather than striving towards autonomy.

9/

With the big data regime we are way out over our skis in terms of working at scales where we aren't positioned to mitigate (or ideally prevent) harms from reproducing & amplifying biases.

10/

I believe @ZeerakTalat is working right in this space --- looking at how the choices we're making now about where machine learning fits in are key & we have real agency to choose better futures.

11/

@ZeerakTalat In conclusion: it is important both for flourishing research and for the resulting futures we can build to deflate the deep learning bubble and make room for other things.

https://twitter.com/emilymbender/status/1430944351358648324?s=20&t=4mIY0Lsaoc3UPqpteV8-HQ