How Intelligent is Artificial Intelligence?

By Oana Marin

Rebooting AI: Building Artificial Intelligence We Can Trust. By Gary Marcus and Ernest Davis. Pantheon Books, New York, NY, September 2019. 288 pages, $28.95.

“To reboot, or not to reboot?” is an almost Shakespearean question that all users face when their computers fail to respond to simple commands. This universal query gains even more depth when humans relinquish their intelligence—the competitive advantage of their species—to a computer, referring to this machine capability as artificial intelligence (AI). Rebooting AI by Gary Marcus and Ernest Davis is both a “food for thought” text as well as a scientific assessment of AI’s status in a world that is overwhelmed with information.

“To boot” is a contronym, as it has two opposite meanings. Intended or not, the pun in the book’s title is not lost on the reader. Depending on the context, “to boot” can either mean to start or be kicked out. The authors do not position themselves on either side of the word, but caution readers as to where relevant pitfalls may lie. The first chapter of Rebooting AI, entitled “Mind the Gap,” makes this point clear. The text is packed with lexical musings, such as “deep learning just ain’t that deep.” This is a jab at the word “depth,” which in the context of deep learning refers solely to the depth of a neural network. So perhaps we may conclude that Marcus and Davis are indeed toying with the vocabulary at a level of subtlety that is only accessible to humans, possibly to strengthen the point made in the title of chapter four: “If Computers Are So Smart, How Come They Can’t Read?”

Historical accounts take the reader from incipient promises made as early as the 1960s, which predicted that computers would flawlessly mimic human intelligence within 20 years, to the present day, where the mirror of our own biases leaves us confounded. Although impartially written and safely grounded in references, the book allows several questions to transpire. Where can we expect AI to succeed? How can AI best serve humankind? In what ways should AI improve? The wealth of facts in the text is not devoid of emotion; the authors express concern when appropriate, and the tone is almost tender and tolerant of the human condition’s fragility.

Marcus and Davis’ exposition follows the basic steps of scientific procedure: problem, hypothesis, and solution. In this spirit, the authors reveal AI’s foundational insufficiencies, hypothesize on their nature, and quickly reassure readers with an array of possible solutions.

The essential distinction between AI and machine learning is often muddled in casual language — particularly in the physics-dominated branches of applied mathematics, where AI robotics applications are of lesser relevance. Rebooting AI clarifies that machine learning is merely a subset of AI that is focused on learning from data, while AI also includes topics like reasoning and knowledge representation.

As the most promising subfield of machine learning, deep learning receives its deconstructive treatment in chapter three (entitled “Deep Learning, and Beyond”) and is not entrenched in the overly-standard supervised/unsupervised/reinforcement learning classification. The authors’ path is instead rather pragmatic, as they explore the extent to which a neural network can perform “end-to-end” tasks. Their discussion about the role of graphical processing units in propelling the AI revolution and consolidating the belief that machines could outperform humans in terms of processing speed (only to hit the wall of big data soon after) would likely be of interest to applied mathematicians.

Proceeding on the assumption that an “end-to-end” machine might be possible, the next chapters bring the reader into the fields of neuroscience, linguistics, and cognitive psychology. Marcus and Davis posit that an “inventory of knowledge” and a “representation of knowledge” are the prerequisites of any reliable AI machine that is free of human confirmation bias. To support this stance, the book establishes a set of 11 clues from the cognitive sciences to guide AI’s development. My personal favorite is “Concepts are embedded in theories,” with its punchline that “no fact is an island.”

From a technical viewpoint, the major challenge of AI pertains to the integration of elusive concepts like common sense. As such, the authors browse various key components in an effort to determine how to render common sense programmable. The text pivots from linguistic theory, following Noam Chomsky’s assertion that “the essence of language is...infinite use of finite means,” to the computer science limitations of finite-bit representations. Similarly, semantic networks—like ConceptNet, which is intended to facilitate reasoning tasks—are swiftly reinterpreted via formal logic.

It is refreshing that Rebooting AI does not shy away from AI’s delicate failures, such as algorithms that displayed discriminatory tendencies (as with image-labeling technology for Google photos in 2015) or anti-Semitic behavior (as with Microsoft’s experimental chatbot, Tay, in 2016). Although the algorithms themselves are not inherently prejudiced, they reinforce human bias through their very nature. Does this fact call for a change in the machine learning paradigm, or is it entirely due to the training data sets?

The closing chapter, “Trust,” launches with an epigraph that is attributed to Zora Neale Hurston: “Gods always behave like the people who make them.” Readers are left to wonder about the identity of the gods. Are they the computers who may overtake humanity in doomsday scenarios, or the humans who—as AI creators—will always hold the advantage of untranslatable infinity of depth (much as the universe has for them)? This last chapter dives deeper into robust engineering design, safeguards, and general scientific practices to answer the question of trust, which is understood as reliability. Readers are also made aware of one of the central puzzles of scientific progress; in interdisciplinary fields like AI, the individual scientific branches that comprise the whole unit do not advance at the same speeds. If this detail already causes difficulties in the physical sciences, having to co-opt the humanities and establish a bridge between these fields is an even more significant challenge.

Though it abounds with warnings about AI, the book ends on an optimistic note and assures readers of AI’s transformative power. Marcus and Davis even voice their collective belief that a machine that can teach itself to be an expert is within reach. When discussing overall impact on humanity, the authors ultimately compare the AI revolution with the Industrial Revolution. The Epilogue idealistically suggests that while automation might lead to higher unemployment rates, it may eventually make way for Oscar Wilde’s vision of life as reduced to “enjoying cultivated leisure.”

In summary, Rebooting AI offers something for everyone: the scientist learning about the historical evolution of AI, the sociologist concerned with AI’s impact on society, the neuroscientist interested in constructing a replica of the human brain, the anthropologist intrigued by the naissance of human intelligence, the science fiction writer searching for a dystopian angle, etc. It is invaluable for a popular science book to seamlessly unite so many types of readers and incite joint efforts to answer major questions that pertain to civilization. Given the authors’ backgrounds, it is not accidental that Rebooting AI inspires en masse, and we are thankful that Marcus and Davis continue to reach popular audiences with their work.

Oana Marin is an Assistant Applied Mathematics Specialist at Argonne National Laboratory. She is active in machine learning applications to mathematical physics.