An AI is an AI is an AI…or not.

Every morning, of late, when I read the news, there is a slew of headlines of what AI has done for us lately.

Just this morning, I read:

Robert Wickham of Salesforce is the source of the last statement, that AI will be the new electricity, once we are done oohing and ahhing. Or being afraid that we will all lose our jobs.

AI, however, is not like electricity. It is not so straight forward. While it may, eventually, be ubiquitous and unconsidered, so far we cannot provide a single and clear definition for what it is, and thus these reductive metaphors create greater confusion than clarity.

In each article ‘AI’ describes something different. Deep learning, neural networks, robotics, hardware, a combination, etc. Even within deep learning or neural networks, the meanings can be different, as can the nuts and bolts. Most media and humans use ‘AI’ as shorthand for whatever suits their context. AI, without an agreed upon definition, but the lack of clarity, differentiation, and understanding does make it very difficult to discuss in a nuanced manner.

There is code, there is data, there is an interface–for inputs and outputs, and all of these are (likely) different in each instantiation.  Most of the guts are proprietary, in the combination of code and data and training. So we don’t necessarily know what makes up the artificial intelligence.

Even code, as shorthand to a layperson, as the stuff that makes computers do what they do, is a broad and differentiated category. In this case, like language, it is used for a particular purpose, so this reduction is perhaps not as dangerous. We’ve never argued that code is going to take over the world, or that rogue code is creating disasters.  As compared to algorithms, a few years ago, and AI, now.

So much of this lumping is a problem? We lump things, such as humans or cats, into categories based on like attributes, but we do have some ways to differentiate them. These may not be useful categories, nationality, breed, color, behavior, gender.  (Even these are pretty fraught of late, so perhaps our categorization schemes for mammals needs some readdressing.)  On the other side, we could consider cancer, an incredibly reductive title for very a broad range of…well of what? Tumor types? Mechanisms? Genetic predispositions? There are discussions, given recent research, as to whether cancer should be a noun, perhaps it is better served as a verb. Our bodies cancer, I am cancering, to show the current activity of internal cellular misbehavior.

What if we consider of this  on the intelligence side, how do we speak of intelligence, artificial or otherwise? For intelligence, like consciousness, in humans, we do not have clear explanations for what it is or how it works. So not perhaps the simplest domain to borrow language from, and apply it to machines.  Neural networks is one aspect, modeled on human brains, but it is limited to the structural pathways, a descriptor of how information travels and is stored.

The choice to use AI to represent such a broad range of concepts, behaviors, and functions concerns me. Even in the set of headlines above, it is difficult to know what is being talked about, from a continuum of visible outputs to robots who speak.  If we cannot be more clear about what we are discussing it is incredibly complicated to make clear decisions, about functions, about inputs, about outputs, data, biases, ethics, and all the things which have broad impacts on society.

I haven’t seen clear work on how we should use this language, and though I worked with IBM Watson for a while on exactly this concern, I can’t say I have a strong recommendation for how we categorize not just what we have now, but, as importantly, what is being built and what will exist in the future. The near future.

I’ll work on this later, as in soon, ways in which to talk about these parts in a public context that are clearer, and allow for growth of creations into a systems model.  Check back!

 

Grammar checking software, language change, and precursors to the future

Almost 20 years ago I wrote my master’s thesis on ‘Grammar Checking Software as a Tool of Language Change’ or some such title. I’ve since lost track of the paper, it was, in that day, on paper.

I was studing at Georgetown University, and my work focused on language and power, from a sociolinguistic and cultural perspective. I had been using assorted early 90s log files I had collected, from IRC and a forum used at UCSC when I was an undergrad, and assessing markers of power in language in the online environment, watching the evolution of language change, and seeing the ways in which one positioned oneself, via language. One of the areas that particularly interested me, which I delved further into, is how non-native speakers of English marked authority online in an anonymous environment and using a language that was rapidly evolving, often different, in each community.

This work led to me to the early grammar and spelling checkers, and my often curiousity as to why they were so not grammatical. I decided to analyse the grammaticality of the current crop of tools against Fowler’s Modern English Usage. For those non-grammar nerds, Fowler’s is an early 20th century grammar text which is/was the go-to for proper English. As a comparator it had its issues, but those I worked around and wrote about. As I worked through analyzing different software programs, I eventually opted to use on Microsoft Word, due to the enormity of my task.

The outcome of this was that even at its most stringent, the grammar checker was no where near Fowler’s. And at its more casual levels — at the time it had three — the English it was recommending was so odd and so lax that the usage of the tool, in my estimation, would ‘teach’ a user a very different English than one would learn at school.

There was no way this was not purposeful, which led me to many questions, most of which remained unanswered, as MS did not wish to speak to me about them.  At the most basic, I was curious who wrote the program because I felt that no linguist would have built a grammatically incorrect system.  This led me to the hypothesis that the linguists built the systems and the marketers freaked out at all the wavy green and red lines and insisted it have fewer. I would love to have heard more of the inside of this, and if you happened to work on this, I’d love to hear from you.

This has really interesting ramifications for both Americans and non-native English speakers. In general, and in different ways, each can use a boost in grammar, and if you have an authoritative tool telling you that no, you cannot use ‘which’ refering to a cat, it must be ‘that’ then this is a possible shift.  It refuses the use of the passive voice, and other constructs it deemed overly complex. One cannot use a run on sentence nor a fragment.  It flattened, significantly, the ability to be creative in language.

We could say that one could turn this off, or that it could be ignored, but in the course of doing the research, I did eventually become bothered by all the wavy lines and want to change words to adhere to what I called Microsoft English.  Because it was, in fact, a significantly different and evolved English.  One that was being rapidly disseminated by what was becoming the most widely used word processor.

Whether or not Microsoft was attempting to create a new English, or was aware of the possible cultural ramifications, and power structures, that they were creating and/or re-inforcing, changes were happening.

And this was in one language. Years later I went back and re-ran some of my assessments in French, out of curiousity for the formality/informality and what it would recommend, and even in a language which has an Academie to control the language, there were shifts away from the rulings. So perhaps in most languages we see changes due to the judgments of software, and then these likely flow into society, because we do learn from software.

I doubt most people have considered this, that the language promoted by their grammar checking software is ‘wrong’ or at the very least, not a standard, until they created a standard.  I am not writing this, nor did I at the time, to be a stickler for Fowler’s or old grammar rules, but to surface the awareness that changes in the systems’ use of language flow outward into spoken language, and that they often have significant ramifications in how we can think about things. If your software systematically attempts to remove the use of animate pronouns for animals, plants and objects, it becomes of judgment of the humanity of any thing other than a human.   These are the things invisible to most people, yet significant in the ways in which we exist in our worlds.

In the same vein, I still look at the ways in which software is an output into our language, not just how language changes within power and control structures, but also the adoption of words, and the modification of our grammar, to make it easier to interact with the machines. More on this last bit later.

 

 

Temporal translation

I collect old dictionaries, in many languages, translations and otherwise. They are full of rich cultural information, new words, pathways, changed meanings, and I enjoy reading them for the glimpses of other worlds.

Often they contain words that I have to look up in other dictionaries, such as my copy of the first Hebrew-English translation dictionary released in Israel. It has so many words about the desert, about the plants, water, formations, growing, that I had to look a signficant number of them up in English, as I had never heard them.  A more modern Hebrew-French translation dictionary I have does not include nearly as many words of this sort.

I can build these models in my mind, in bits and pieces. But what would it be like to build them in the machine, to provide a rich view into different time periods by pouring in time-specific language data?

What if the machine can translate me to 1700s English? What if the machine translated from time periods, different Englishes, or Frenches? What about dialects?

I don’t know where phonology data would come from. What if I want to translate to Beowulf? How does the machine learn to pronounce the language properly?

I can imagine an amazing visualization, a time line, that I can drag into the past, to hear the sounds. Except it would need regional variation as well.

In the tradtion of vac, in which sound matters, the sound and the meaning intricately entwined, what histories can we learn by having the ability to translate to other places in time, not just other languages?

Do men dream of electric families?

In two recent articles I’ve read, men have created robots have modeled after female family members.

Martine Rothblatt, who created BINA48 in the image of his wife, Bina Aspen Rothblatt, and Hiroshi Ishiguro, who created his in the image of his (then) five-year old daughter.

Why?

Sources:

  • https://futurism.com/the-most-life-life-robots-ever-created/
  • https://www.wired.com/2017/10/hiroshi-ishiguro-when-robots-act-just-like-humans/

 

Neural Machine Translation architecture

Almost all such systems are built for a single language pair — so far there has not been a sufficiently simple and efficient way to handle multiple language pairs using a single model without making significant changes to the basic NMT architecture.

Google’s engineers working on NMT released a paper last year detailing a solution to multilingual NMT systems that avoided making significant changes to the architecture, which was based on a single language pair translation.

This makes me wonder what the NMT architectural structure would be and how it would differ from what is currently in place, to be optimized for multilinguality.  And what the differences would be, in how it behaves, if any.

I wonder if the system were architected in a language other than English, if it would be different. What do you get if you cross Sapir-Whorf with systems architecture?

I wonder how the machine would translate ‘soy milk’ to Romanian. Would it assume English is the source language because of ‘milk’?

 

No humans involved! (Except where they were.)

The headlines declaiming no humans were involved in AlphaGo Zero’s mastery continue to amuse me. Now if a machine had taught AGZ the principles and set it off on its path, that would be really something.

The press’ continued erasure of the humans who built the machines and provided the first principles from which to learn the game are indicative of the larger issue in that we don’t see the humans behind these decisions which are made, and thus have no insight into bias etc.

The linked headline above at least specifies that humans were not involved in the mastery, rather than the creation. AGZ played significantly more games than AGL, though this is also not often mentioned either. If you harken back to Gladwell’s supposition that it takes 10,000 hours of practice to become an expert, and we are looking at machines playing from 100,000 games to 4MM games, to ‘learn’ to excel, it is not surprising that they are outplaying humans.  We simply do not have the capacity (or longevity, and likely desire) to play so many games.

The description below, of allowing AGZ t to have access to all past experiences, which AGL did not have except when playing humans is very interesting and I’d love to know more about this decision and why it was taken.

AlphaGo Zero’s creators at Google DeepMind designed the computer program to use a tactic during practice games that AlphaGo Lee didn’t have access to. For each turn, AlphaGo Zero drew on its past experience to predict the most likely ways the rest of the game could play out, judge which player would win in each scenario and choose its move accordingly.

 

AlphaGo Lee used this kind of forethought in matches against other players, but not during practice games. AlphaGo Zero’s ability to imagine and assess possible futures during training “allowed it to train faster, but also become a better player in the end,” explains Singh, whose commentary on the study appears in the same issue of Nature.

 

Stronger attention to language, please.

AI Now Institute is an independent research institution looking at the ‘social implications of artificial intelligence,’ something very needed as we continue to have such rapid and significant change in AI, driven by a very limited set of creators.

The Institute itself has four domains which it uses to bucket to the work they do on “the social implications of artificial intelligence”.

  • Rights & Liberties
  • Labor & Automation
  • Bias & Inclusion
  • Safety & Critical Infrastructure

Reading their 2017 Report, I believe it should have more emphasis on language in its recommendations, about which I would like to say more. Linguists, specifically those outside the computational linguistics field, need to be more integrated in the creation of AI as technology and interface.

Language is an underlying consideration in Bias & Inclusion, with this description:

However, training data, algorithms, and other design choices that shape AI systems may reflect and amplify existing cultural prejudices and inequalities.

Which does not have a strong enough inclusion or consideration of the language(s) used to train these systems. While bias work is written about language in AIs, it is more likely to fall on the corpus, that is to say, which words are used in descriptions, and the like. When you feed a corpus into a machine, it brings all its biases with it, and since these data sets are extant, they come with all the societal issues that seem to be more visible now than at earlier times.

Language and power, language and bias, minority languages, all of these have long been topics in the field of linguistics. They are also touched on in behavioral economics, with priming and other ways in which we set into the minds of humans particular considerations. You can also see this in the racial bias work from Harvard that was very prevalent on the web a few years back.

Language is a core human communication tool that entails so much history and meaning that without a greater attention to the social and cultural implications of the language we choose, from how we discuss these topics, to how language is embedded in the interaces of our AI systems we are glossing over something of great meaning with far too little attention.

I don’t think that language belongs only in the realm of bias & inclusion, in the long run. It may create outsiders at this time, but language is such a core consideration, it seems larger than any of these four domains.  Though to note, as well, none of these domains explicitly attend to the interfaces and the ways in which we interact with these systems, so language would belong there as well, as an input and an output, with differing needs and attentions on each side.

 

 

 

 

“AI Learns Sexism Just by Studying Photographs”

Two articles, one from Wired and one from the MIT Technology Review on bias in software. the quotes below, on gender bias.

As sophisticated machine-learning programs proliferate, such distortions matter. In the researchers’ tests, people pictured in kitchens, for example, became even more likely to be labeled “woman” than reflected the training data. The researchers’ paper includes a photo of a man at a stove labeled “woman.”

 

Machine-learning software trained on the datasets didn’t just mirror those biases, it amplified them. If a photo set generally associated women with cooking, software trained by studying those photos and their labels created an even stronger association.

These are interesting to me for several reasons.

First, it assumes that there is a bias against the kitchen, or women in the kitchen. But when one considers that most top chefs (Michelen 5 star) are men, it isn’t just about women in the kitchen, where the bias exists, but on a different level, and perhaps a nuance the machines don’t yet grasp.

Second, they are labeled as the AI learning sexism. I would be more inclined to suggest that the AI learned American categorization and labeling systems. The humans added the value judgement. I wonder why the machine was looking at images and labeling gender. How does a computer understand man/woman? By what criteria is it taught to differentiate?  Which inevitably brings us to the discussion of who creates the algorithms and programs, chooses the data sets, and teaches the labels to the machines.

It feels like solving an issue of ‘distortion’ in the way a machine learns, if that machine is reflecting both the programmers and the society, isn’t a solve, if it’s machine-level only. This is, perhaps, not the entire conversation, or even the wrong conversation.

It makes me think we need a deeper discussion on what the AI sees, how it applies labels, and how humans both interpret them and understand them. It reminds me of Lombroso and Lacassagne. Are we on our way to recreate the past, with different parameters?

 

 

Linguistic anthropology and AI, part 2

I posted the original set of questions so I could shoot them over to a few people, to get their thoughts on my thoughts. Delivered even more than expected.  In the emails and conversations I’ve had since then, there are ever more questions, that I am going to keep documenting here.

  • If it were possible to allow the AIs to interrupt each other, to cut in before one finished what it was saying, what would happen?
  • What happens if you have three AIs in conversation or negotiation?
  • Are the AIs identical in the beginning? If, so, who modifies language first, and do they do it differently? In concert? In reaction?
  • Does an AI who changes language get considered a new incarnation of the AI? Does it modify itself, as it modifies its language?
  • If you have two AIs with different programming, two different incarnations, of a sort, what modifications do they make, vs two instantiations of the same thing?
  • Does language come about as a means of addressing desires and needs? [Misha wrote this and I find I don’t agree, which is really a deeply fascinating place to go with this.]
  • Can machines have desires and needs? How would we know the answer to this?
  • Is the assumption that machines modify language for reasons of efficiency overly deterministic?
  • What is the role of embodiment in the creation of language? Is it required for something to be meaningful? Does it change the way language works? Would it ‘count’ for cyborgs?

One thing I have discovered is that I go at this from a different perspective than many of my conversation partners, which is that I accept that it is possible that everything we think we know is wrong, both about humans, and about machines.  As I wrote, we assume humans are rational in order to make models of human behavior, which are faulty, because we are not. We assume machines are rational, because we programmed them to be, but what if they, too, are not? There seems to be a sense that binary does not allow for irrationality, or anomaly, but..what if it does?

I think I need to wrap into these discussions four things:

  1.  a primer on computational linguistics for those who don’t have it
  2.  a bit of an overview on general linguistics, and where we stand on that
  3.  an overview of creole linguistics, because I think it is a very interesting model to use for the evolution of AI languages, particularly and perhaps except, for the bit where it requires a power dynamic, historically.
  4. some discussion of the genetic evolution of algorithms, deep learning, adversarial networks etc.

Misha’s last really interesting question to me: “Can you evolve language without pain?” is a bit acontextual as I toss it here, but what an interesting question about feedback loops.

 

Nuance, AI, Cancer, and Leadership

I read an article on AI and leadership this morning, where the concern is that if the workers are all AIs, we won’t learn how to lead, as humans.  It is an interesting consideration, and not one I’d have come to without it being raised by someone else. But I don’t want to talk about that, I am more interested in the selection of sources, and the realistic abilities of an AI who replaces a human in cancer diagnosis and treatment being a successful option.

The article, without noting dates etc. uses old sources, a video from 2014 which makes some interesting assumptions (“Replacing human muscle with mechanical muscle frees people to specialize and that leaves everyone better off. This is how economies grow and standards of living rise.”) to bolster the argument that the thinking machines will take our thinking jobs.  The focus is on the robots, so statements such as the above one is just supposed to slide past as being obviously true. I don’t agree carte blanche with that statement, or many others in the video.

The author also cites the Watson PR piece about Watson diagnosing a rare leukemia and recommending a treatment that consequently saves the life of a woman in Japan. This, too, is over a year old, which is not noted in the article. I am not suggesting this invalidates the OOOH factor, but let’s just note that this is a single instance that has been all over the press as a sign of the future of Watson in medicine, and Watson being better than humans at diagnosing cancer.  And this is what I actually want to comment on.

Watson is fed data, millions of records of cancers, tests, treatments, outcomes, any genetic information on the cancer and the patient. Watson is also trained over years to get the right answers, so that Watson can continue to refine. Given that massive amount of data, it is not surprising that Watson can not only pick up a rare instance, but also recommend an unusual treatment which saves a life. I do agree that there are instances like this, where data and processing power will win.  What we never hear of, of course, are the instances where patients die, where we thought Watson could have done better.

And it may be that there will be a greater variety of treatments offered to patients because of the machine’s broad view of all treatments for that cancer to date, in particular populations, and their success rates. BUT the full set of recommended treatments may also be small, it will be the set of treatments which have already worked, because that is what an AI will choose from. It won’t come up with something new and novel, it will apply what already works.  This is fundamentally limiting.

The NCI has already done this, created a set of protocols with fixed recommendations for treatments. I am told it takes years to change these. But at the moment, you can discuss these recommendations with your doctor, specifict to your case. I’d hate to imagine a world where there is no dialogue, but perhaps that is me. I have seen studies which show that the machine is better than humans at diagnosing lung cancer. (This is a different role, not the oncologist who decides treatment path, but the pathologist who diagnoses your cancer, and determines its stage, which of course, feeds heavily in to treatment options.)

I’d like to point out, explicitly, that the end goal of treatments in these instance, in particular in the American medical system, is the continuation of life, and there is _nothin_ about the quality of life, just the extension.  If you look at how hospitals and doctors are rated, this is also obvious. Quality of life post-treatment isn’t on the list.

I am reminded of the dogs who can smell cancer, and other means of diagnosis that are novel but perhaps should be done in concert with humans.

I’d argue that cancer diagnosis and treatment is as much art as science, and if the machines are the science, we shouldn’t drop the art. Or the humanity, which, for the moment, the humans still corner the market on.