Communications of the ACM

ity into this consequential question: How effective will function-based approaches be when applied to new and broader applications than those already targeted, particularly those that mandate more stringent measures of success? The question has two parts: The first concerns the class of cognitive tasks whose corresponding functions are simple enough to allow compact representations that can be evaluated efficiently (as in neural networks) and whose estimation is within reach of current thresholds— or thresholds we expect to attain in, say, 10 to 20 years. The second alludes to the fact that these functions are only approximations of cognitive tasks; that is, they do not always get it right. How suitable or acceptable will such approximations be when targeting cognitive tasks that mandate measures of success that are tighter than those required by the currently targeted applications?

The Power of Success

Before I comment on policy considerations, let me highlight a relevant phenomenon that recurs in the history of science, with AI no exception. I call it the “bullied-by-success” phenomenon, in reference to the subduing of a research community into mainly pursing what is currently successful, at the expense of pursuing enough what may be more successful or needed in the future.

Going back to AI history, some of the perspectives promoted during the expert-systems era can be safely characterized today as having been scientifically absurd. Yet, due to the perceived success of expert systems then, these perspectives had a dominating effect on the course of scientific dialogue and direction, leading to a bullied-by-success community.s I saw a similar phenomenon during the transition from logic-based approaches to probability-based approaches for commonsense reasoning in the late 1980s. Popular arguments then, like “People don’t reason probabilistically,” s A colleague could not but joke that the broad machine learning community is being bullied today by the success of its deep learning sub-community, just as the broader AI community has been bullied by the success of its machine learning sub-community.

man speech, and African grey parrots can generate sounds that mimic human speech to impressive levels. Yet none of these animals has the cognitive abilities and intelligence typically attributed to humans.

One of the reactions I received to such remarks was: “I don’t know of any animal that can play Go!” This was in reference to the AlphaGo system, which set a milestone in 2016 by beating the world champion in the game. Indeed, we do not know of animals that can play a game as complex as Go. But first recall the difference between performance and intelligence: A calculator outperforms humans at arithmetic without possessing human or even animal cognitive abilities. Moreover, contrary to what seems to be widely believed, AlphaGo is not a neural network since its architecture is based on a collection of AI techniques that have been in the works for at least 50 years.o This includes the minimax technique for two-player games, stochastic search, learning from self-play, use of evaluation functions to cut off minimax search trees, and reinforcement learning, in addition to two neural networks. While a Go player can be viewed as a function that maps a board configuration (input) to an action (output), the AlphaGo player was not built by learning a single function from input-output pairs; only some of its components were built that way.p The issue here is not only about assigning credit but about whether a competitive Go function can be small enough to be represented and estimated under current data-gathering, storage, and computational thresholds. It would be quite interesting if this was the case, but we do not yet know the answer. I should also note that AlphaGo is a great example of what one can achieve today by integrating model-based and function-based approaches.

Pushing Thresholds

One cannot of course preclude the
possibility of constructing a competi-
tive Go function or similarly complex
o Oren Etzioni of the Allen Institute for Artificial
Intelligence laid out this argument during a
talk at UCLA in March 2016 called Myths and
Facts about the Future of AI.
p AlphaZero, the successor to AlphaGo, used one
neural net work instead of two and data generat-
ed through self-play, setting another milestone.
functions, even though we may not
be there today, given current thresh-
olds. But it begs the question: If it
is a matter of thresholds, and given
current successes, why not focus all
our attention on moving thresholds
further? While there is merit to this
proposal, which seems to have been
adopted by key industries, it does
face challenges that stem from both
academic and policy considerations.
I address academic considerations
next while leaving policy consider-
ations to a later section.

From an academic viewpoint, the history of AI tells us to be quite cautious, as we have seen similar phenomena before. Those of us who have been around long enough can recall the era of expert systems in the 1980s. At that time, we discovered ways to build functions using rules that were devised through “knowledge engineering” sessions, as they were then called. The functions created through this process, called “expert systems” and “knowledge-based systems,” were claimed to achieve performance that surpassed human experts in some cases, particularly in medical diagnosis.q The term “knowledge is power” was used and symbolized a jubilant state of affairs, resembling what “deep learning” has come to symbolize today.r The period following this era came to be known as the “AI Winter,” as we could finally delimit the class of applications that yielded to such systems, and that class fell well short of AI ambitions.

While the current derivative for progress on neural networks has been impressive, it has not been sustained long enough to allow sufficient visibil- q One academic outcome of the expert system era was the introduction of a dedicated master’s degree at Stanford University called the “Master’s in AI” that was separate from the master’s in computer science and had significantly looser course requirements. It was a two-year program, with the second year dedicated to building an expert system. I was a member of the very last class that graduated from the program before it was terminated and recall that one of its justifications was that classical computer science techniques can be harmful to the “ heuristic” thinking needed to effectively build expert systems.