We’re deepening the credibility crisis in data science
1. Background 2. History of AI 3. Limits of ML 4. Next Steps 5. Killing AI 6. Closing Thoughts
TL;DR: Data scientists have a responsibility to shepherd the term artificial intelligence out of the world. We need to show maturity in walking back the near unattainable promise embedded in this misused phrase.
Background
“That’s a really strange pronunciation of machine learning, friend.”
This is my response to a fellow data scientist’s use of the words artificial intelligence during a mid-week phone call about how we might grow our business.
He pauses, and, cognizant that snarky interruptions aren’t an effective teaching tool, I take a deep breath and explain: “Let’s say machine learning because it more accurately describes the process of teaching mathematical models to generate insights by providing them with training examples.”
As I’ve deepened my understanding of data science, I’ve increasingly found that the phrase artificial intelligence sparkles with false promise:
- The term misleads decision-makers into investing in advanced analytics before their organization has reached sufficient data maturity. After all, intelligence should be able to cope with a little data complexity, right?
- It confuses the heck out of young students, who might’ve applied a little extra attention in their statistics class if they understood that was the path to a promising career in data science.
- It distracts from tried and true methodologies within data science.
I’m ready to kill artificial intelligence.
History of Artificial Intelligence
The academic discipline got its start in 1955 with the goal of creating machines capable of mimicking human cognitive function. Learning, problem solving, applying common sense, operating under conditions of ambiguity — taken together, these traits form the basis for general intelligence, the long-standing goal of artificial intelligence.
Since inception, AI research has experienced boom and bust cycles, fueled by an abundance of optimism followed by a collapse of funding. These setbacks have been so dramatic and so endemic to the field that they received their own neologism: AI winter. The two most dramatic winters occurred in the mid-to-late ’70s and the mid-80s to mid-90s. Failure to appropriately manage hype is the commonly cited cause of this unfortunate antipattern.
In the words of the chief data scientist of the Strategic Artificial Intelligence Lab, T. Scott Clendaniel:
“I’m really concerned that we’re going to enter a third AI winter … because I think the field has been so overhyped that there’s no way it can possibly live up to expectations.”
Lately, the machine learning community has been fielding a lot of journalistic inquiries about AI. In May 2020, OpenAI released their GPT-3 model for natural language processing. The cost to initially train the model was $4.6 million, requiring 355 GPU years of computational power. As of now, OpenAI has released the model through a controlled API access point rather than making the code freely available to researchers.
Setting aside concerns of impracticality and inaccessibility for a moment — GPT-3 has produced some impressive feats. And yet, importantly, this overhyped development does not move us closer to artificial intelligence.
If advancing research into AGI is analogous to sending a spacecraft to explore Mars, then the development of GPT-3 is more or less analogous to investing $4.6 million into a rocket that creates a beautiful fire cloud of exhaust without ever leaving the launchpad.
Limits of Machine Learning
The general consensus of the research community is that AGI won’t be attained by deepening machine learning techniques.
Machine learning capabilities are narrow. An ML algorithm may be able to achieve better-than-human performance but only on exceedingly specific tasks and only after immensely expensive training.
Human capabilities are comparatively broad. We seem to be exceptionally good at one-shot learning — making inferences and assigning categories based on a very few examples. Young children quickly master tasks of matching, sorting, comparing, and ordering. Infants have an innate desire to explore novelty and draw conclusions about the wider world.
These capacities are as yet unmatched by machine intelligence. There’s somewhat of a paradox in the fact that machine learning can defeat a grand master at Go, but a robot can’t beat a toddler at sorting blocks.
Following the path of improved machine learning techniques seems unlikely to result in the attainment of human levels of common-sense reasoning or versatile problem solving. To quote machine learning pioneer Stuart Russell:
“I don’t think deep learning evolves into AGI. Artificial General Intelligence is not going to be reached by just having bigger deep learning networks and more data… Deep learning systems don’t know anything, they can’t reason, and they can’t accumulate knowledge, they can’t apply what they learned in one context to solve problems in another context etc. And these are just elementary things that humans do all the time.”
In other words, statistics-based solutions are fairly good at interpolation — i.e., drawing conclusions about novel examples that fall within the bounds of data they’ve already seen. They aren’t very good at extrapolation — i.e., using what they’ve learned to make conclusions about the broader world.
It’s possible that “the greatest trick AI ever pulled was convincing the world it exists.”
Next Steps
I think it’s extremely important for machine learning researchers to ask themselves if they’re solving the right challenges.
Here’s a YouTube video about StarGAN v2, a machine learning model that can take a cat photo and use it to create a bunch of similar images inspired by photos of dogs. Meanwhile, data quality issues cost U.S. organizations $3.1 trillion a year according to analysis from IBM.
Perhaps removing AI from our lexicon will help debunk the notion that an artificially intelligent tool can address substantive data management failures. Low-quality data is unfortunately ubiquitous — and it’ll impair business function and impede the implementation of even the simplest of advanced-analytics tools.
Killing Artificial Intelligence
If you’re a data scientist or machine learning engineer, I hope your takeaway from this article is a sense of responsibility to quit feeding into the hype around AI. Unless you’re on the absolute cutting edge of AI research, your use of the term should be reserved for discussion of the superintelligence control problem and other future-oriented considerations.
If you’re a nontechnical person, you can safely replace just about every use of artificial intelligence with very, very advanced statistics.This is certainly true for any tool available on the marketplace today. There’s still room for philosophical discussion around questions of artificial general intelligence, but that technology is still a long ways off.
And if you happen to be a business leader, particularly one with AI in your title, here’s what Anil Chaudhry, the director of AI implementations at the U.S. GSA, has to say about his role:
“I describe AI to people as augmented intelligence, not artificial intelligence.”
Even leaders with AI in their title are cringing away from AI.
In summary, the future vision of artificial intelligence won’t be achieved through contemporary methods. The hype around massive, impractical models such as GPT-3 reveals a lack of understanding about the current state of machine intelligence — or lack thereof.
The overuse of artificial intelligence isn’t just a whimsical exaggeration — it’s damaging to the data science community and risks tipping the field into a crisis of confidence.
Closing Thoughts
Here are three trends I see for data science in the next three to five years.
Stewardship of the language used in our field
Clearly, I feel strongly that there’s an imperative to remove AI from the lexicon.
Increased reliance on human-centered design to identify risks associated with machine learning malpractice
Not all errors should be treated as equally bad.
To quote Stuart Russell again:
“Some kinds of errors are relatively cheap. Whereas classifying a human as a gorilla, as Google found out, is really expensive, like in the billions of dollars of trashing your goodwill and global reputation. I’m sure it was sort of an innocent error, coming from just using a uniform loss function.”
Identifying the gravity of this potential mistake from the outset of the design process could have saved the ML engineering team at Google a lot of heartache. Incorporating human-centered design into the model-creation process doesn’t just help with selection of the optimal loss function, but it also helps identify potential sources of bias (e.g., racially unbalanced training data) and other risks.
A return to the fundamentals, including a renewed focus on understanding the end-to-end process of data generation
I’d love to see data scientists develop a breadth of skills spanning the data generation pipeline.
Quoting again, this time from Harvard professor Gary King:
“You have to know the whole data generation process from which the data emerge… We always focus on that whole chain of evidence… After all, we’re studying the world — we’re not studying some bunch of data.”
This is why I think it’s crucial for data scientists to develop familiarity with the principles of end-to-end data strategy.