From AI to ML to AI: On Swirling Nomenclature & Slurried Thought

Artificial intelligence is transforming the way we work (Venture Beat), turning all of us into hyper-productive business centaurs (The Next Web).  Artificial intelligence will merge with human brains to transform the way we think (The Verge). Artificial intelligence is the new electricity (Andrew Ng).  Within five years, artificial intelligence will be behind your every decision (Ginni Rometty of IBM via Computer World ).

Before committing all future posts to the coming revolution, or abandoning the blog altogether to beseech good favor from our AI overlords at the AI church, perhaps we should ask, why are today’s headlines, startups and even academic institutions suddenly all embracing the term artificial intelligence (AI)?

In this blog post, I hope to prod all stakeholders (researchers, entrepreneurs, venture capitalists, journalists, think-fluencers, and casual observers alike) to ask the following questions:

1. What substantive transformation does this switch in the nomenclature from machine learning (ML) to  artificial intelligence (AI) signal?
2. If the research hasn’t categorically changed, then why are we rebranding it?
3. What are the dangers, to both scholarship and society, of mindlessly shifting the way we talk about research to maximize buzz?

In the Beginning, There was “AI”

Well, perhaps not in “the beginning”, but certainly when I entered this world, AI was the preferred term of art to label the broad academic efforts at investigating machine intelligence.

Since the sudden embrace of the term AI reverses a previous rebranding, we ought to ask, what warranted the switch from AI → ML in the first place?

For us to be more than sales-{men,women}, we ought to have reasons for choosing our words beyond those same reasons that motivate marketers. It’s worth asking: what real change justified the initial rebranding of the field from AI to ML? And does this reversal in nomenclature correspond to a 180 in research direction along that same axis?

A brief incomplete history of AI

It’s hard to define AI, but CMU’s Dean of the School of Computer Science Andrew Moore gave my favorite definition to date in this interview with Forbes:

Artificial intelligence is the science and engineering of making computers behave in ways that, until recently, we thought required human intelligence.

It’s a nice definition because it captures (1) the balance of theoretical inquiry and engineering, (2) the breadth of the term, and (3) the sense in which the term is aspirational, a moving target based on those capabilities that humans possess but which machines do not.

Earlier AI work encompassed a broad set of methodological approaches, among which ML was only one. For the uninitiated reader, ML is simply the study of algorithms whose performance on tasks improves as a result of  interactions with data, either via by mining patterns from some compiled dataset (as in supervised learning) or by interacting with a dynamic environment (as in reinforcement learning).

For example, given a large dataset of MRI images and corresponding diagnoses, an ML algorithm might generate a predictive model. Then given subsequent MRI images, the model could predict the probability that an MRI depicts cancer (based on the patterns gleaned from the training data).

Back in the 1980s, many researchers pursued other popular avenues of AI research that had no reliance on pattern recognition. At times, these other approaches attracted wider attention. For example, the IBM supercomputer Deep Blue relied entirely on tree search algorithms and heuristics. Other prominent researchers were focused on rule-based expert systems, attempts to codify the knowledge of domain experts as collections of if-then rules that could be executed automatically by computers. While the hype over expert systems never quite rivaled today’s fawning over machine learning, it did capture the attention of business executives and received coverage in popular venues like Harvard Business Review.

However, (apologies to readers for whom the narrative’s become hackneyed), many of these research sub-communities within artificial intelligence overpromised and underdelivered. In a thought-piece published in AAAI in 1997, UCLA Professor of Computer Science Richard Korf noted that the word artificial intelligence had acquired a “taint” due to “failing to deliver on its early promises”. In the same article, he reveals that while IBM often described Deep Blue as a “supercomputer”, they took pains to explicitly claim “that Deep Blue did not use artificial intelligence”.

Per Korf’s analysis, the name artificial intelligence “itself, … seems much more grandiose and boastful than other early alternatives”. Moreover, he wrote at the time that:

 “work in AI is often thought to be lacking in rigor and without scientific basis. To some extent this a product of the nature of the problem being studied, but it also is the fault of those of us in the field who have not been sufficiently diligent in applying rigorous scientific standards to our work.”

In this climate, the machine learning community, which was burgeoning with activity and (at that time) stood on increasingly sound theoretical footing, sought to differentiate itself by identifying primarily under the name machine learning and publishing primarily in venues more strictly associated with data-driven approaches, e.g. the NIPS, ICML, and COLT conferences, and the Machine Learning Journal, subsequently superseded by the Journal of Machine Learning Research.

The name gave machine learners a fresh slate in the eyes of funding agencies, entrepreneurs, and venture capitalists, and it more closely aligned ML with mathematics and statistics.

My purpose here is not to evangelize the ML approach or to impugn the alternatives. I simply argue that the switch to ML as the preferred identifier corresponded to an actual shift in research direction and that the term itself is reasonable.

Why Switch Again?

If the first name-change came to salvage a discipline amid crisis, then this recent adoption of AI most certainly did not.

Around  6 years ago, when I first thought about applying for a PhD, ML had just begun to enter the public discourse. Some companies, like Google, had long been using ML. Even at Google, the moneymaker was Google Web Search, widely known to rely more on expertly assembled pipelines of traditional information retrieval tools than on ML. The general perception was that two primary paths were open for ML scientists: online advertising and academic research.  Overall, financial prospects appeared worse than riding the web 2.0 wave in SF, a reasonable price to pay for doing more interesting work.

In the last few years, of course, that idea has become quaint. Breakthroughs in deep learning (a subfield of ML) have rendered technologies feasible that power speech recognition and image classification aboard every major smartphone sold today. Each year, more companies acquire data suitable to ML and the algorithms have continued to grow more powerful. Demand for ML talent has never been higher, the tools never better, and the bar for employment never lower. ML technology appears poised to impact ever more businesses.

However, just a couple years after ML entered the public consciousness, despite the absence of any appreciable pivot away from ML approaches, something strange happened: people started calling it AI.

The Sudden Embrace of “AI”

A reasonable person stumbling upon the branding gymnastics might suppose that research must have shifted away from ML and back towards less data-driven approaches.

Puzzlingly, that’s not the case. If anything the dominance, both in academic and industrial spheres, of data-driven methods has only increased. The ML subfield has never consumed a greater share of the AI pie. Instead, perhaps the shift is better explained via the vices that Korf associated with the term AI itself: grandiosity and boastfulness. Perhaps, as Ali Rahimi noted in his test of time talk at NIPS 2017, the shift also signals a renewed lack of rigor.

Below, I outline some specific reasons that I speculate explain why stakeholders, whether consciously or by passively following incentives, are pushing the term “AI” at this moment.

To COnvey A SENSE OF MOMENTUM

As ML research has transitioned from a primarily academic discipline to a matter of public interest more of the discourse takes place in public venues, through popular articles, press releases, and punditry.  In these shallower dialogues, the discussion of technology is at best gestural. What company is doing big data? Doing ML? Doing AI? How much are researchers doing X getting paid?

Because the technology itself is discussed so shallowly, there’s little opportunity to convey any sense of technological progress by describing the precise technical innovations. Instead, the easiest way to indicate novelty in the popular discourse is by changing the name of the field itself!

What was Google working on 6 years ago? Big data. Four years ago? Machine learning. Two years ago? Artificial intelligence. Two years from now? Artificial general intelligence!

Whether or not the technological progress provides any intellectually sensible justification for relabeling a field of research, readers respond to periodic rebranding. Researchers in turn have an incentive to brand their work under the new name in order to tap in to the press coverage.

EXPLOITING MISCONCEPTIONS

The term AI carries 63 years of baggage. While a generation of grizzled scholars might have associated the term with expert systems, the common reader associates AI with an expansive body of fictional references from cinematic and literary science fiction. In most of these works, “AI” isn’t used as the name of a field but rather to refer to specific entities as “artificial intelligences”. They are generally depicted as conscious, anthropomorphic agents.

By using the term AI, journalists and evangelists get their audiences excited at the expense of misleading them. The New York Times, Wired, and other credible sources have full out embraced the term AI as a substitute for ML. As a result, when the charlatans writing for futurism.com announce “The World’s First Album Composed and Produced by an AI“, audiences are liable to take the bait.

INTELLECTUAL WILDCARD

In a recent blog post, Berkeley Professor Michael Jordan took on the hype around AI, and delved a bit into the use of the word itself in the discourse on machine intelligence.  While he goes on to propose addressing an already confused nomenclature by introducing a host of new acronyms (never a good idea), his strongest point to my ears was this:

The current public dialog about these issues too often uses “AI” as an intellectual wildcard, one that makes it difficult to reason about the scope and consequences of emerging technology.

In other words, the lack of specificity allows journalists, entrepreneurs, and marketing departments to say virtually anything they want. The failure to articulate a clean point is weaponized to blunt opposition, similarly to how the president uses the phrase “we’ll see”.

Conclusion

To wrap up, I’d argue that rigor and scholarship don’t just pertain to one’s discipline in making arguments in papers and running proper ablation studies. These are questions of culture. Do we fundamentally value clarity, honesty, and scholarship, or are we willing to do or say anything because it’s presently expedient or trendy?

To put the present argument in context, I wouldn’t rule out the use of the term artificial intelligence. I’m actually quite fond of Andrew Moore’s definition, and as an aspirational term it may be a good word to keep around. For example, in some contexts like academic departments, the term really is used to refer more expansively to a combination of ML and non-ML approaches to building machines that perform tasks that until recently only humans could perform – e.g. the intersection of robotics, planning, and machine learning so fruitfully explored by many of my colleagues at CMU.

Still, this doesn’t absolve us of the responsibility to introspect and to critically consider the abrupt pivot in nomenclature when applied to purely machine-learning-based research. Why are writers who only cover machine learning suddenly called AI journalists? And why are startups that simply apply linear models and/or deep learning in web applications suddenly called AI (vs ML) startups? We ought to examine the underlying motivations, and to consider what this capricious subservience to trendiness says about the culture of scholarship in our field.

The AI Misinformation Epidemic

This post continues a long concentration on the AI misinformation epidemic.