Vinton G. Cerf's Cerf's Up column "The Internet in the 21st Century" (Sept. 2018) highlighted many challenges facing today's Internet, including risks to privacy, security, and society that did not exist when the network was originally being built in the late 1960s. His concern is warranted and will require us to strike a balance between protecting the democratic and egalitarian values that made the Internet great to begin with while ensuring those values are used for good. The fundamental issue, then, in creating a 21st-century Internet becomes what changes are warranted and who will be responsible for defining and administering them.
On the technology dimension, computer scientists and engineers must develop smarter systems for detecting, addressing, and preventing malicious content on the Web. Cerf's argument on behalf of user training is helpful but will not ultimately solve the problem of an untrustworthy, ungovernable, potentially malicious network. I myself recently fell for a phishing attack, which only proves that today's attacks can fool even savvy, experienced users. Meanwhile, bad actors worldwide exploit the same infrastructure that is used by billions of well-intentioned people every day, including notoriously scammers using Google AdWords in 2017 to impersonate Amazon.1 There is clearly a need for systems deployed across all layers of the Internet stack to prevent, detect, and address such abuses.
Technology by itself is not the solution. Billions of ordinary people thus have a role to play in combatting such potentially destructive forces. Critical thinking and digital literacy must be taught in grade school, enabling even children to question the overload of information, including misinformation churned out by countless bad actors. Society must also reward, rather than stigmatize, people and organizations for doing the right thing, as when an organization admits it has suffered an attack, even if embarrassing to admit.
A free and democratic Internet available to all requires a governance structure that reflects the network's global scope. A major reason it continues to succeed is that no single government agency has total control. Instead, bits of it are maintained by various international non-governmental agencies, including the Internet Corporation for Assigned Names and Numbers. However, we are beginning to see national and regional governments more directly regulate the Internet their citizens access through filters or data privacy laws (such as the European Union's General Data Protection Regulation). Such walled gardens produce a fragmented experience for many users. Needed are international non-governmental organizations that address these concerns, with the legal authority to perform digital forensics to identify cyber attacks and criminals, perhaps even by issuing Interpol-type warrants for alleged criminals. They could also formally pass and enforce rules and publish guidance covering, say, data privacy as well.
James Simpson, York, U.K.
Vinton G. Cerf views himself as a layman on the subject of neural networks, and so do I. However, I think I understand heuristic (lookahead) search, especially for computer chess. I thus limit myself here to Cerf's assertion in his Cerf's Up column "On Neural Networks" (July 2018) that "AlphaGo Zero learned to play chess well enough to beat most (maybe all?) computer-based players in 24 hours." Since he did not mention "search" per se, such a statement suggests that learned neural networks would be likely to perform well in chess, as was reported in Chess News (https://en.chessbase.com/post/thefuture-is-here-alphazero-learns-chess) and mentioned by Cerf in the column.
"Lookahead search" means exploring the state-space rooted in a given (chess) position. Due to combinatorics, exhaustive exploration is normally impossible in computer chess, yielding two successful approaches:
Minimaxing. Exploring to bounded (though usually variable) depth, as Scheucher and I explained,2 why minimaxing heuristic values works at all, despite its theoretical problem identified by ACM Turing Award laureate Judea Pearl and others; and
Monte Carlo tree search. Sampling through selective explorations to "leaf" positions (such as checkmate).
For my point here, the particular search approach needed to achieve excellent performance does not matter, only whether search is involved in the move decision. I conjecture that AlphaZero2 would achieve superhuman performance in chess through minimaxing as well.
In describing AlphaZero (which I am convinced is what Cerf meant to say in the column, rather than "AlphaGo Zero"), Silver et al.3 did not mention Monte Carlo tree search until somewhere in the middle of their paper, reflecting its status as not particularly important. In fact, AlphaZero did not learn but was simply programmed to perform search. Note that Silver et al. thoroughly analyzed the role of search in the predecessor AlphaGo in Nature in 2016.
Although neither AlphaZero nor its learned neural network appear to be publicly available, I have managed to explore a reimplementation called LCZero at http://lczero.org/. I recently matched up its neural network without search against my (and colleagues Helmut Horacek's and Marcus Wagner's) old Merlin chess program, which tied for sixth place out of 24 machines in the 1989 World Computer Chess Championship. I emulated the computer power it had at the time on a five-year-old laptop (with a single core) by setting its performance time at two seconds per move rather than three minutes on a mainframe at the time. While I consider the moves played by LCZero's neural network generally plausible, even without search, Merlin beat it decisively in these games, as a neural network without search often blunders in chess positions where non-trivial tactics really do matter.
I thus consider Cerf's claim about computer chess and neural networks misleading, but is such a claim indeed worth pointing out here, in a letter to the editor? What worries me is the effect of the kind of "telephone game" begun by Silver et al.,3 a message mis-characterized in Chess News and now again in Communications by an author of Cerf's stature. We can only imagine how such a telephone game would continue to play out in the regular nontechnical press if we did not cut the thread now.
Hermann Kaindl, Vienna, Austria
Kaindl is correct. I meant to say AlphaZero and mistakenly wrote AlphaGo Zero. I also I agree that search is important for the AlphaZero system and should have drawn attention to it but honestly had not done enough due diligence with the DeepMind team, a deadline effect...
I appreciate the additional color and clarity.
Vinton G. Cerf, Mountain View, CA
I am not one given to writing complementary letters to publications but must say Communications (Sept. 2018) was brilliant in exploring several extremely relevant intellectual challenges with its readers, specifically Vinton G. Cerf's discussion of the "Treaty of Westphalia" in his Cerf's Up column "The Peace of Westphalia" and its relevance to ongoing international interference in elections around the world; Moshe Y. Vardi's discussion of "disruptive technology" in his Vardi's Insights column "Move Fast and Break Things," noting that while computer scientists should "celebrate" their achievements, they also need to, as Vardi put it, "drive very carefully"; the passion of the letters to the editor concerning a prior Vardi column "How the Hippies Destroyed the Internet" (July 2018) on the past (and future) of the Internet; and my favorite, Adam Barker's Viewpoint "An Academic's Observations from a Sabbatical at Google" on the importance and relevance of software practice to software academics worldwide.
I also noted the then-recently announced "China Region Special Section" (Nov. 2018). As a participant in a 1987 People to People international travel visit to China with other international computer scientists, I recall being amazed to find that, in what was at the time a fairly primitive country technologically, Chinese computer scientists were almost all theoreticians, rather than pragmatists. It took considerable thought for me to realize, or perhaps rationalize, that this was happening because labor was notably inexpensive in China at a time that doing things manually was significantly less costly than doing them with computer support.
Robert L. Glass, Toowong, QLD, Australia
James Simpson's letter to the editor "Side with ACM Ethical Values" (Aug. 2018) concerning Moshe Y. Vardi's Vardi's Insights column "Computer Professionals for Social Responsibility" (Jan. 2018) rightly suggested that ACM policies are inextricably linked to ethical and moral values and hence to political considerations. Simpson even urged establishment of a new ACM special interest group dedicated to ethics and public policy.
But Simpson's implicit encouragement of ACM to lobby for immigration of tech workers into the U.S. fell flat on two accounts: First, it would avoid responsibility for home-growing the talent American industry needs. The news story "Broadening the Path for Women in STEM" by Esther Stein (also in Aug. 2018) included data on the declining participation of women in CS majors in the U.S. Meanwhile, separately, members of the U.S. team in the International Mathematical Olympiad (http://www.ams.org/news?_id=4446) are almost always entirely Asian-Americans. And second, and perhaps more important from an ethical point of view, U.S. hiring of foreign tech workers can amount to a form of "colonial brain drain" from other countries that need their own IT professionals at home.
A broad call to U.S. students to pursue STEM or CS careers is not the answer. American companies do not need workers who are poorly prepared, do not love what they do, or do not care about the moral dimensions of their work. Tech companies, as well as computer science and engineering generally, need people called to IT as a profession, not just a career. ACM should thus consider these issues before deciding to lobby on behalf of bringing in tens of thousands more foreign tech workers.
Paul J. Campbell, Beloit, WI, USA
1. Vaas, L. Scammers slip fake Amazon ad under Google's nose. Naked Security (Feb. 10, 2017); https://nakedsecurity.sophos.com/2017/02/10/scammers-slip-fake-amazon-ad-under-googles-nose/
3. Silver, D. et al. Mastering chess and Shogi by self-play with a general reinforcement learning algorithm. arXiv, 2017; https://arxiv.org/abs/1712.01815
Communications welcomes your opinion. To submit a Letter to the Editor, please limit yourself to 500 words or less, and send to [email protected].
©2018 ACM 0001-0782/18/12
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and full citation on the first page. Copyright for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or fee. Request permission to publish from [email protected] or fax (212) 869-0481.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2018 ACM, Inc.
"American companies do not need workers who are poorly prepared, do not love what they do, or do not care about the moral dimensions of their work. Tech companies, as well as computer science and engineering generally, need people called to IT as a profession, not just a career."
Although I agree with later part of the statement, does the author of this letter imply that immigrant tech workers from China, India and other countries "are poorly prepared, do not love what they do, or do not care about the moral dimensions of their work"?
Displaying 1 comment