Hacker News Reader

Comment by Barrin92 | Hacker News Reader

Good article. I think issues like the mistranslation "The bear headed man needed a hat" really show that what we need is not an "AI algorithm" but AI agents. The lack of common sense that we see in machines is very much intended, it's what makes them machines in the first place.

Humans don't make these mistakes because humans are able to create stories and place even a mundane translation in the deep context of a 'world' that must be coherent. Heads simply don't have bears on them, we wouldn't trust that translation even if we had heard it clearly, and we'd rather think we hallucinate before we'd drive on a road that goes nowhere in an impossible direction.

Algorithms that are just glorified feature extraction machines, in my opinion, can never create a coherent story, so I'm very skeptical about the claim that we will have somehow solved intelligence in the next ten years. It seems to me like we are almost nowhere closer to it than we were decades ago.

View on HN

I agree with the general skepticism regarding AGI, but I think it's worth pointing out that there is some progress towards robust representations of abstract conceptual knowledge (http://science.sciencemag.org/content/331/6022/1279) that are far beyond feature extraction.

Isn't it also true that what we define as coherent is based on observed data that could be fed into these algorithms?

In a sense you can always assume that the answer to these problems is just 'make the neural net bigger' but I find this deeply unhelpful for two reasons:

1. This clearly does not seem to be the way humans learn. Humans can learn from very few examples, in entirely unguided environments, and they don't face the same issues that existing algorithms suffer from. (for example humans have no big problem with rotational invariance, whereas ML vision algorithms do).

2. It's essentially surrendering to the fact that we aren't able to understand how cognition works and build higher-level representations as a result. The goal of AI research can not just be to feed data blindly into enormous primitive structures, it must also be to get a grasp on what sort of complex structures are part of intelligent agents and how they interact.

Humans can learn from very few examples, in entirely unguided environments, and they don't face the same issues that existing algorithms suffer from. (for example humans have no big problem with rotational invariance, whereas ML vision algorithms do).

That's because, contrary to the zeitgeist, humans are not a blank slate. Our brains are the result of billions of years of evolution. They are extremely well adapted to modelling the natural environment and the behaviour of other beings around us. This is in stark contrast to computers which we start from nothing and force feed a huge amount of data without context and then expect results. The fact that this approach works at all for some tasks is staggering.

Not an expert, but as long as there is no way of knowing WHAT it is the algorithm is learning it seems to me that it could never work reliably. It might look perfectly reasonable until you hit one of the triggers the algorithm used to segment the data.

Someone somewhere shared a story about using machine learning to spot the difference between US and Russian tanks ; which apparently worked fine until field testing, where it failed miserably. What the algorithm had learned was the difference between great quality photos of US tanks and poor quality photos of Russian. True or not, this is exactly the kind of issues that will keep popping up.

Plenty of people are spending plenty of time figuring out how to mess with facial recognition as we speak by taking advantage of the same fundamental weakness.

Oh, yep! It's super simple to induce systemic errors like that. Take https://github.com/kevin28520/My-TensorFlow-tutorials/tree/m... . Lighten every dog by 20%. Darken every cat by 20%. Train. Take image of cat, lighten 20%, watch as it's transformed into a dog!

For large corpora, it's impossible to know what features got selected. They probably aren't any feature a human would consider.

> Heads simply don't have bears on them

https://encrypted-tbn0.gstatic.com/images?q=tbn%3AANd9GcSzPe...

Perhaps the man proceeding towards the grizzly had a negative economic outlook.

There are literally hundreds of word-sense-disambiguation papers being published per year, and have there have been for about a decade. The state-of-the-art has been advancing relentlessly, and examples like the ones cited here are being zoomed past right now by the latest thing, which is even larger models pre-trained on unsupervised data. There will be a next thing, and for NLP this will be backgammon or chess to 2020s Go.

Embodiment is a better criticism, but this article is still bad, and is written as though Intelligence were some essential soul-like property that agents either possess or don't.

It is really amazing how often articles with this viewpoint repeat one particular trope: they put the word "understanding" in italics. Look out for it.

But if you do an error analysis on even ShotgunWSD (Betnaru 2017), you get errors that humans would not produce - archaic favors, missing recent terms, confusing opposing meanings, ... When the average researcher can easily select the correct sense. And that's a DETERMINISTIC algorithm! Swapping ELMO for pure WordNet gives us slightly better results in that sense but introduces literal random chance errors. And the older, Bayesian models are worse!

While we'll probably solve WSD to a high degree of accuracy (such that engineering bugs cause more failures than the model), our current approach is still fundamentally flawed. We need a better method of imbuing causal relationships than correlation+chance.

> archaic favors, missing recent terms, confusing opposing meanings,

Humans make errors like these all the time (e.g., I don't know what you mean by 'favor' here, and I'm sure many people wouldn't have the type "ELMO"), and the dictionary definition of the word shifts accordingly, except we can't really call them 'errors' when it's a natural phenomenon that we're trying to match. Humans make seemingly random errors sometimes too, confusing one word for a mother, I mean another. When we do it we infer (perhaps correctly) deep psychological causes. Current machines don't have personal identities and desires and goals like we do, but embodied agents with a natural selection process could have those things, there isn't an obvious technological barrier. And if you build in a language model that shares some of the general computation machinery we could reasonably expect it to make those kinds of errors.

> Heads simply don't have bears on them

Also, humans may refer to other humans as toy-bears and get punished for it (https://www.theguardian.com/world/2018/aug/07/china-bans-win...). Human language is hard, metaphors are hard.

Reply on news.ycombinator.com