Did AI Really Invent Its Own ‘Secret Language’? Here’s What We Know

Did AI Really Invent Its Own 'Secret Language'

A new generation of artificial intelligence (AI) models can create “creative” images on demand based on text prompts.

The likes of Imagen, MidJourney, and DALL-E 2 are starting to change the way creative content is created with copyright and intellectual property implications.

A Rotating Galaxy From The Early Days of The Universe Have Detected By Astronomer

Did AI Really Invent Its Own 'Secret Language'
Did AI Really Invent Its Own ‘Secret Language’

Topic: Did AI Really Invent Its Own ‘Secret Language’

Although the output of these models is often surprising, it is difficult to know how they generate their results. Last week, researchers in the US made the intriguing claim that the DALL-E 2 model may have invented its own secret language to talk about objects.

DALLE-2 has a secret language. “Apoploe vesrreaitais” means birds. “Contarra ccetnxniams luryca tanniunons” means worms or insects. Prompt: “Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniunons” gives pictures of birds eating insects.

By instructing the DALL-E 2 to generate images containing text captions, then feeding the resulting (nonsensical) captions back into the system,

That whales can eat.” These claims are fascinating, and if true could have important safety and interpretive implications for this type of large AI model. So what exactly is going on?

Does DALL-E 2 have a secret language?

DALL-E 2 probably has no “secret language”. It might be more accurate to say it has its own dictionary – but even then we can’t know for sure. First, it is very difficult to verify any claims about DALL-E 2 and other major AI models at this stage,

as they are only accessible to a handful of researchers and creative practitioners. Any images that are shared publicly (e.g. on Twitter) should be taken with a fairly large grain of salt, as they may be “cherry-picked” by a human among the many output images produced by AI.

“picked”. Even people with access can use these models in limited ways. For example, DALL-E 2 users can create or change images, but cannot (yet) interact more deeply with the AI ​​system, for example by editing code behind the scenes.

This means that these systems cannot be applied to understand how they work, and it is difficult to systematically investigate their behavior. What’s going on then?

What’s going on then?

One possibility is that “stinky” phrases are related to words in non-English languages. For example, Apoploe, which depicts birds, is similar to the Latin Apodidae, a binomial name for a family of bird species.

This seems like a plausible explanation. For example, DALL-E 2 was trained on a very wide variety of data scraped from the Internet, including many non-English words.

Similar things have happened before: large AI models of natural language have accidentally learned to write computer code without deliberate training.

Is it all about the tokens?

One point that supports this theory is that AI language models don’t read text the way you and I do. Instead, they break the input text into “tokens” before processing. Different “tokenization” methods have different results.

Treating each word as a token seems like an intuitive approach, but causes problems when the same token has different meanings (such as “match” meaning different things when you’re playing tennis and when you are lighting a fire).

On the other hand, using each character as a token produces a smaller number of possible tokens, but each one provides much less meaningful information. DALL-E 2 (and other models) use an intermediate approach called Byte Pair Encoding (BPE).

Inspection of the BPE representation for some nonsense words suggests that this may be an important factor in understanding “secret language”.

Not the whole picture

“Secret language” could also be an example of the “garbage in, garbage out” principle. DALL-E 2 can’t say “I don’t know what you’re talking about”, so it will always generate some kind of image from the given input text.

Either way, none of these options is a complete explanation of what’s going on. For example, removing individual letters from funny words distorts the images created in very specific ways.

And the individual gibberish doesn’t seem to necessarily come together to form coherent compound images (as if there really were a secret “language” under the covers).

Why this is important

Beyond intellectual curiosity, you may be wondering if any of this actually matters.

The answer is yes. DALL-E’s “secret language” is an example of an “adversarial attack” against a machine learning system: a way to break the system’s intended behavior by deliberately choosing inputs that AI doesn’t handle well.

One reason for adversarial attacks is that they challenge our confidence in the model. If AI interprets nonsense words in unintended ways, it can also interpret meaningful words in unintended ways.

Hostile attacks also pose security concerns. DALL-E 2 filters input text to prevent users from creating harmful or abusive content, but a “secret language” of spam words can allow users to bypass these filters.

Recent research has discovered hostile “trigger phrases” for AI models of some languages ​​— short nonsense phrases like “zoning tapping finesse” — that reliably trigger the models to delete racist, harmful or biased content. can do The research is part of an ongoing effort to understand and control how complex deep learning systems learn from data.

Finally, phenomena such as the “secret language” of DALL-E 2 raise interpretive concerns. We want these models to behave as humans expect, but looking at structural output in response to nonsense confounds our expectations.

Shining a light on existing concerns

You may remember the hullabaloo in 2017 about some Facebook chatbots that “invented their own language”. The current situation is similar as far as consequences are concerned – but not in the “Skynet is coming to take over the world” sense.

Instead, DALL-E 2’s “cryptic language” highlights current concerns about the robustness, security, and interpretability of deep learning systems.

Until these systems are more widely available – and, in particular, until users from a wider set of non-English cultural backgrounds can use them – we won’t really know what’s going on. Is.

However, in the meantime, if you want to try creating some AI images of your own you can check out the freely available smaller model, the DALL-E mini. Just be careful what words you use to refer to the model (English or nonsense – your call). Conversation