Concept Safety: Producing similar AI-human concept spaces

A frequently-raised worry about AI is that it may reason in ways which are very different from us, and understand the world in a very alien manner. For example, Armstrong, Sandberg & Bostrom (2012) consider the possibility of restricting an AI via “rule-based motivational control” and programming it to follow restrictions like “stay within this lead box here”, but they raise worries about the difficulty of rigorously defining “this lead box here”. To address this, they go on to consider the possibility of making an AI internalize human concepts via feedback, with the AI being told whether or not some behavior is good or bad and then constructing a corresponding world-model based on that. The authors are however worried that this may fail, because

Humans seem quite adept at constructing the correct generalisations – most of us have correctly deduced what we should/​should not be doing in general situations (whether or not we follow those rules). But humans share a common of genetic design, which the OAI would likely not have. Sharing, for instance, derives partially from genetic predisposition to reciprocal altruism: the OAI may not integrate the same concept as a human child would. Though reinforcement learning has a good track record, it is neither a panacea nor a guarantee that the OAIs generalisations agree with ours.

Addressing this, a possibility that I raised in Sotala (2015) was that possibly the concept-learning mechanisms in the human brain are actually relatively simple, and that we could replicate the human concept learning process by replicating those rules. I’ll start this post by discussing a closely related hypothesis: that given a specific learning or reasoning task and a certain kind of data, there is an optimal way to organize the data that will naturally emerge. If this were the case, then AI and human reasoning might naturally tend to learn the same kinds of concepts, even if they were using very different mechanisms. Later on the post, I will discuss how one might try to verify that similar representations had in fact been learned, and how to set up a system to make them even more similar.

Word embedding

A particularly fascinating branch of recent research relates to the learning of word embeddings, which are mappings of words to very high-dimensional vectors. It turns out that if you train a system on one of several kinds of tasks, such as being able to classify sentences as valid or invalid, this builds up a space of word vectors that reflects the relationships between the words. For example, there seems to be a male/​female dimension to words, so that there’s a “female vector” that we can add to the word “man” to get “woman”—or, equivalently, which we can subtract from “woman” to get “man”. And it so happens (Mikolov, Yih & Zweig 2013) that we can also get from the word “king” to the word “queen” by adding the same vector to “king”. In general, we can (roughly) get to the male/​female version of any word vector by adding or subtracting this one difference vector!

Why would this happen? Well, a learner that needs to classify sentences as valid or invalid needs to classify the sentence “the king sat on his throne” as valid while classifying the sentence “the king sat on her throne” as invalid. So including a gender dimension on the built-up representation makes sense.

But gender isn’t the only kind of relationship that gets reflected in the geometry of the word space. Here are a few more:

It turns out (Mikolov et al. 2013) that with the right kind of training mechanism, a lot of relationships that we’re intuitively aware of become automatically learned and represented in the concept geometry. And like Olah (2014) comments:

It’s important to appreciate that all of these properties of W are side effects. We didn’t try to have similar words be close together. We didn’t try to have analogies encoded with difference vectors. All we tried to do was perform a simple task, like predicting whether a sentence was valid. These properties more or less popped out of the optimization process.

This seems to be a great strength of neural networks: they learn better ways to represent data, automatically. Representing data well, in turn, seems to be essential to success at many machine learning problems. Word embeddings are just a particularly striking example of learning a representation.

It gets even more interesting, for we can use these for translation. Since Olah has already written an excellent exposition of this, I’ll just quote him:

We can learn to embed words from two different languages in a single, shared space. In this case, we learn to embed English and Mandarin Chinese words in the same space.

We train two word embeddings, Wen and Wzh in a manner similar to how we did above. However, we know that certain English words and Chinese words have similar meanings. So, we optimize for an additional property: words that we know are close translations should be close together.

Of course, we observe that the words we knew had similar meanings end up close together. Since we optimized for that, it’s not surprising. More interesting is that words we didn’t know were translations end up close together.

In light of our previous experiences with word embeddings, this may not seem too surprising. Word embeddings pull similar words together, so if an English and Chinese word we know to mean similar things are near each other, their synonyms will also end up near each other. We also know that things like gender differences tend to end up being represented with a constant difference vector. It seems like forcing enough points to line up should force these difference vectors to be the same in both the English and Chinese embeddings. A result of this would be that if we know that two male versions of words translate to each other, we should also get the female words to translate to each other.

Intuitively, it feels a bit like the two languages have a similar ‘shape’ and that by forcing them to line up at different points, they overlap and other points get pulled into the right positions.

After this, it gets even more interesting. Suppose you had this space of word vectors, and then you also had a system which translated images into vectors in the same space. If you have images of dogs, you put them near the word vector for dog. If you have images of Clippy you put them near word vector for “paperclip”. And so on.

You do that, and then you take some class of images the image-classifier was never trained on, like images of cats. You ask it to place the cat-image somewhere in the vector space. Where does it end up?

You guessed it: in the rough region of the “cat” words. Olah once more:

This was done by members of the Stanford group with only 8 known classes (and 2 unknown classes). The results are already quite impressive. But with so few known classes, there are very few points to interpolate the relationship between images and semantic space off of.

The Google group did a much larger version – instead of 8 categories, they used 1,000 – around the same time (Frome et al. (2013)) and has followed up with a new variation (Norouzi et al. (2014)). Both are based on a very powerful image classification model (from Krizehvsky et al. (2012)), but embed images into the word embedding space in different ways.

The results are impressive. While they may not get images of unknown classes to the precise vector representing that class, they are able to get to the right neighborhood. So, if you ask it to classify images of unknown classes and the classes are fairly different, it can distinguish between the different classes.

Even though I’ve never seen a Aesculapian snake or an Armadillo before, if you show me a picture of one and a picture of the other, I can tell you which is which because I have a general idea of what sort of animal is associated with each word. These networks can accomplish the same thing.

These algorithms made no attempt of being biologically realistic in any way. They didn’t try classifying data the way the brain does it: they just tried classifying data using whatever worked. And it turned out that this was enough to start constructing a multimodal representation space where a lot of the relationships between entities were similar to the way humans understand the world.

How useful is this?

“Well, that’s cool”, you might now say. “But those word spaces were constructed from human linguistic data, for the purpose of predicting human sentences. Of course they’re going to classify the world in the same way as humans do: they’re basically learning the human representation of the world. That doesn’t mean that an autonomously learning AI, with its own learning faculties and systems, is necessarily going to learn a similar internal representation, or to have similar concepts.”

This is a fair criticism. But it is mildly suggestive of the possibility that an AI that was trained to understand the world via feedback from human operators would end up building a similar conceptual space. At least assuming that we chose the right learning algorithms.

When we train a language model to classify sentences by labeling some of them as valid and others as invalid, there’s a hidden structure implicit in our answers: the structure of how we understand the world, and of how we think of the meaning of words. The language model extracts that hidden structure and begins to classify previously unseen things in terms of those implicit reasoning patterns. Similarly, if we gave an AI feedback about what kinds of actions counted as “leaving the box” and which ones didn’t, there would be a certain way of viewing and conceptualizing the world implied by that feedback, one which the AI could learn.

Comparing representations

“Hmm, maaaaaaaaybe”, is your skeptical answer. “But how would you ever know? Like, you can test the AI in your training situation, but how do you know that it’s actually acquired a similar-enough representation and not something wildly off? And it’s one thing to look at those vector spaces and claim that there are human-like relationships among the different items, but that’s still a little hand-wavy. We don’t actually know that the human brain does anything remotely similar to represent concepts.”

Here we turn, for a moment, to neuroscience.

Multivariate Cross-Classification (MVCC) is a clever neuroscience methodology used for figuring out whether different neural representations of the same thing have something in common. For example, we may be interested in whether the visual and tactile representation of a banana have something in common.

We can test this by having several test subjects look at pictures of objects such as apples and bananas while sitting in a brain scanner. We then feed the scans of their brains into a machine learning classifier and teach it to distinguish between the neural activity of looking at an apple, versus the neural activity of looking at a banana. Next we have our test subjects (still sitting in the brain scanners) touch some bananas and apples, and ask our machine learning classifier to guess whether the resulting neural activity is the result of touching a banana or an apple. If the classifier—which has not been trained on the “touch” representations, only on the “sight” representations—manages to achieve a better-than-chance performance on this latter task, then we can conclude that the neural representation for e.g. “the sight of a banana” has something in common with the neural representation for “the touch of a banana”.

A particularly fascinating experiment of this type is that of Shinkareva et al. (2011), who showed their test subjects both the written words for different tools and dwellings, and, separately, line-drawing images of the same tools and dwellings. A machine-learning classifier was both trained on image-evoked activity and made to predict word-evoked activity and vice versa, and achieved a high accuracy on category classification for both tasks. Even more interestingly, the representations seemed to be similar between subjects. Training the classifier on the word representations of all but one participant, and then having it classify the image representation of the left-out participant, also achieved a reliable (p<0.05) category classification for 8 out of 12 participants. This suggests a relatively similar concept space between humans of a similar background.

We can now hypothesize some ways of testing the similarity of the AI’s concept space with that of humans. Possibly the most interesting one might be to develop a translation between a human’s and an AI’s internal representations of concepts. Take a human’s neural activation when they’re thinking of some concept, and then take the AI’s internal activation when it is thinking of the same concept, and plot them in a shared space similar to the English-Mandarin translation. To what extent do the two concept geometries have similar shapes, allowing one to take a human’s neural activation of the word “cat” to find the AI’s internal representation of the word “cat”? To the extent that this is possible, one could probably establish that the two share highly similar concept systems.

One could also try to more explicitly optimize for such a similarity. For instance, one could train the AI to make predictions of different concepts, with the additional constraint that its internal representation must be such that a machine-learning classifier trained on a human’s neural representations will correctly identify concept-clusters within the AI. This might force internal similarities on the representation beyond the ones that would already be formed from similarities in the data.

Next post in series: The problem of alien concepts.