Teaching students to stay away from Physiognomic AI

I read Luke Stark and Jevan Hutson‘s Physiognomic AI paper last night and it’s sparked some thinking about additional reading I could add to my graduate course on statistical theory for engineering next semester (Detection and Estimation Theory).

“The inferential statistical methods on which machine learning is based, while useful in many contexts, fail when applied to extrapolating subjective human characteristics from physical features and even patterns of behavior, just as phrenology and physiognomy did.”

From the (mathematical) communication theory context in which I teach these methods, they are indeed useful. But I should probably teach more about the (non-mathematical) limitations of those methods. Ultimately, even if I tell myself that I am teaching theory, that theory has a domain of application which is both mathematically and normatively constrained. We get trained in the former but not in the latter. Teaching a methodology without a discussion of its limitations is a bit like teaching someone how to shoot a gun without any discussion of safety [*]

The paper describes the parallels between the development of physiognomy and some AI-based computer vision applications to illustrate how claims about the utility or social good arguments made now are nearly identical. They quote Lorenzo Niles Fowler, a phrenologist: “All teachers would be more successful if, by the aid of Phrenology, they trained their pupils with reference to their mental capacities.” Compare this to the push for using ML to generate individual learning plans.

The problem is not (necessarily) that giving students individualized instruction is bad, but that ML’s “internally consistent, but largely self-referential epistemological framework” cherry picks what it wants from the application domain to find a nail for the ML hammer. As they write: “[s]uch justifications also often point to extant scientific literature from other fields, often without delving its details and effacing controversies and disagreements within the original discipline.”

Getting back to pedagogy, I think it’s important to address this “everything looks like a nail” phenomenon. One start is to think carefully even about the cartoon examples we use as examples. But perhaps I should add a supplemental reading list to go along with each topic. We fancy ourselves as theorists, but I think it’s a dodge. Students are taking the class because they want to learn ML because they are excited about doing machine learning. When they go off into industry, they should be able to think critically about whether the tool is right for the job: not just “is logistic loss the right loss function” but “is this even the right question to be asking or trying to answer?”

[*] That is, very American?

A story about Canvas

One upon a time, there was a University whose administration was enthralled by a religion called Canvas. The central tenets of Canvas were held in the highest esteem and those who followed Canvasitic doctrine were expected to prepare their course materials through prescribed rituals and incantations.

When preparing the course, problems for homeworks and quizzes (assessments) could be organized into Question Banks stored in the professor’s office.

  • To create an assessment, a special place was prepared on the wall of the classroom.
  • The professor would photocopy the Questions from the Question Bank, one for each student, and would post them in sheaves on the wall for the students to take.
  • The students would collect the assessment. When they completed it, the scores would be tallied and magically stored in the Professor’s ledger.
  • The go-getting-est among them would realize some error or typo on a Question and inform the professor.
    Even though the professor could alter the Question in the Question Bank, the photocopies had been made and he could not alter the Questions already taken by the students.
  • However, he could amend the photocopies on the wall for those laggards who had not collected their assessment from the wall. For those students, the Question was corrected.
  • The amended Question was only corrected on those assessments left on the wall. Harried for time, they (the professor) may or may not have corrected the Question in the Question Bank. This will be important later.
  • The laggards would have correct tallies but the go-getters’ tallies had to be manually amended, one by one.

At the start of the next year, the professor, exhausted, asked “can I not reuse the materials from last year?” The high priests of Canvas exclaimed “but of course!” in a tone that suggested the professor was interested in mustard of Francophone origin. The professor chanted the strange rituals and lo and behold, a copy of the class was prepared.

For each section of wall (assessment), sheaves of amended Problems were already there, rustling in the autumnal breeze.

  • The amended Problems did not appear in the Question Banks. The magic of Canvas was unable to “scan” a problem from the wall and copy it into a Question Bank. It was as if the Question on the assessment was an entirely different object from the Question in the Question Bank.
  • The professor, full of great plans for improved assessments, began to amend the Question Banks. However, since photocopies had been made, so the professor tore down the sheaves of Problems on the wall and prepared to make a fresh batch of photocopies.
  • However, the magic of Canvas had also created phantom duplicates of the Question Banks in the professor’s office, so the amended Questions and unamended Questions were hard to distinguish. Moreover, problems with editable equations were copied as uneditable photographs (PNGs).

The professor, distraught, turned to their local priest. The priest, unable to help, conveyed the question to higher authorities, where it vanished.

Meanwhile, the professor’s assistants waited for some instruction on how to prepare the course for the hundreds of students already waiting outside the classroom…

[NB: as afar as I can tell from searching forums and documentation, this is actually how Canvas behaves, but I am willing to be shown otherwise.]

Distinguished Lecturers should not be vetoed by the US

I attended the IEEE Information Theory Society (ITSOC) Board of Governors meeting at ISIT in Paris this week and found something gnawing at me afterwards from the presentation about the Distinguished Lecturer (DL) program. The presentation said that “IEEE denied the selection of a DL based in Iran due to U.S. sanction.” The name of the particular DL nominee does not appear in the public record.

Why can IEEE deny the selection of a DL? In part, there are requirements for DLs now:

DL should visit IT Society local chapters. DL program pays for airfare and travel. Local chapter pays for local expenses (hotel). If traveling to a different continent, visits to two locations are required. DL lectures should be freely accessible to the public (i.e. no registration fees).

A DL from Iran cannot be reimbursed by IEEE because the IEEE is based in the US and has to abide by US law. By the new rules then, scholars from Iran are automatically disqualified from the DL program.

Being a DL is an important recognition: it is arguably an award. It certainly bestows a certain level of prestige. Acceding to this intervention by IEEE sends the message that “if you are from Iran, you can’t get an award.” Once we go down this road we might as well ban conference submissions, membership, and participation in the academic community for scholars from Iran. Why not go whole hog and become a tool of the US State Department? It’s ludicrous.

ITSOC should not sit by and passively accept this “veto” from IEEE: it’s an assault on academic freedom that devalues scholarship on purely political grounds. To not even name the nominee erases the honor to which they are entitled. In fact, they should be given the honor/award with the stipulation that they are exempt from the reimbursement. It is possible to take a stand without violating the law: recognize this scholar and take a public stand against the encroachment of American foreign policy onto an international academic community.

That review is so… meta

Reviewing has started for NeurIPS 2019 and this time around I am an area chair (AC). We’ve been given a lot of instructions and some tasks: bidding on papers, bidding on reviewers, adjusting reviewers, identifying what we think are likely rejects in the batch of papers we are handling, and so on. It’s a little more involved than being an AC for ICML, but that’s to be expected since the whole reviewing game has been evolving rapidly to adapt to the massive increase in submissions.

Since there is yet another tier of TPC above the ACs (the Senior ACs), how should one approach the meta-review? One view is that the meta-review is AC’s decision/opinion informed by the reviews, the response, the discussion, and their own reading of the paper. This makes the AC a bit like an associate editor at a journal. This also gives the AC quite a bit of flexibility: if the discussion is limited or not particularly useful, the AC can fill in the gap by adding their own voice. The downside is that ACs might bring more of their own preferences (or biases) to the process.

A different approach is to make the meta-review akin to a panel summary as part of an NSF review. In the panels I have been on, there are N people who write reviews of each proposal, one of whom leads the discussion. There is also a scribe for the discussion who has not written a review: a dispassionate observer. The whole panel (even those who didn’t read the proposal) participates in the discussion. The scribe is supposed to draft a summary/synthesis of the discussion and runs it past the panel for edits until they reach a consensus. The N reviews are still there though, with their diversity of opinion.

I think I might prefer the second model. The setup is a bit different, since authors get to respond to the reviews. The meta-review is supposed to augment the existing reviews by incorporating the discussion and author response. The AC is supposed to guide the discussion, which is a role shared by the lead discussant and program officer in the NSF model. The only problem is that the amount of discussion on each paper is highly variable. It’s sometimes like pulling teeth to get reviewers to respond/interact. Reviewers, for their part, might be participating in 5 different discussions, so context switching to each paper can be tough. But for papers with some reasonable discussion, the meta-review as panel summary might be a good way to go.

One complaint about panel summaries is that they often feel anodyne. However, I think this might be desirable in a meta-review, since it could lead to fewer angry authors. One aspect of the NSF model which I think could be adopted, regardless of how the AC views their job, is running the meta-review past the reviewers. I did this for ICML and got some edits and feedback from the reviewers that improved the final review.

CFP: Theory and Practice of Differential Privacy (TPDP) 2019

November 11
London, UK
Colocated with CCS 2019

Differential privacy is a promising approach to privacy-preserving data analysis.  Differential privacy provides strong worst-case guarantees about the harm that a user could suffer from participating in a differentially private data analysis, but is also flexible enough to allow for a wide variety of data analyses to be performed with a high degree of utility.  Having already been the subject of a decade of intense scientific study, it has also now been deployed in products at government agencies such as the U.S. Census Bureau and companies like Apple and Google.

Researchers in differential privacy span many distinct research communities, including algorithms, computer security, cryptography, databases, data mining, machine learning, statistics, programming languages, social sciences, and law.  This workshop will bring researchers from these communities together to discuss recent developments in both the theory and practice of differential privacy.

Specific topics of interest for the workshop include (but are not limited to):

  • theory of differential privacy,
  • differential privacy and security,
  • privacy preserving machine learning,
  • differential privacy and statistics,
  • differential privacy and data analysis,
  • trade-offs between privacy protection and analytic utility,
  • differential privacy and surveys,
  • programming languages for differential privacy,
  • relaxations of the differential privacy definition,
  • differential privacy vs other privacy notions and methods,
  • experimental studies using differential privacy,
  • differential privacy implementations,
  • differential privacy and policy making,
  • applications of differential privacy.

Submissions

The goal of TPDP is to stimulate the discussion on the relevance of differentially private data analyses in practice. For this reason, we seek contributions from different research areas of computer science and statistics.  Authors are invited to submit a short abstract (4 pages maximum) of their work. Submissions will undergo a lightweight review process and will be judged on originality, relevance, interest and clarity. Submission should describe novel work or work that has already appeared elsewhere but that can stimulate the discussion between different communities at the workshop. Accepted abstracts will be presented at the workshop either as a talk or a poster.  The workshop will not have formal proceedings and is not intended to preclude later publication at another venue. Selected papers from the workshop will be invited to submit a full version of their work for publication in a special issue of the Journal of Privacy and Confidentiality.

Submission website: https://easychair.org/conferences/?conf=tpdp2019

Important Dates

Submission: June 21 (anywhere on earth)

Notification: August 9

Workshop: 11/11

Program Committee

  • Michael Hay (co-chair), Colgate University
  • Aleksandar Nikolov (co-chair), University of Toronto
  • Aws Albarghouthi, University of Wisconsin–Madison
  • Borja Balle, Amazon
  • Mark Bun, Boston University
  • Graham Cormode, University of Warwick
  • Rachel Cummings, Georgia Tech University
  • Xi He, University of Waterloo
  • Gautam Kamath, University of Waterloo
  • Ilya Mironov, Google Research – Brain
  • Uri Stemmer, Ben-Gurion University
  • Danfeng Zhang, Penn State University

For more information, visit the workshop website at https://tpdp.cse.buffalo.edu/2019/.

Signal boost: travel grants for SPAWC 2019

Passing a message along for my colleague Waheed Bajwa:

As the US Liaison Chair of IEEE SPAWC 2019, I have received NSF funds to support travel of undergraduate and/or graduate students to Cannes, France for IEEE SPAWC 2019. Having a paper at the workshop is not a prerequisite for these grants and a number of grants are reserved for underrepresented minority students whose careers might benefit from these travel grants. Please share this with any interested students and, if you know one, please encourage her/him to consider applying for these grants.

ICML 2019 encouraged code submission. That is great!

ICML 2019 had an optional code submission for papers. As an area chair, I handled a mix of papers, some more theoretical than others, but almost all of them had some empirical validation. Not all of them submitted code. For a paper with a theorem, the experiments can range from sanity checks to a detailed exploration of the effects of some parameters for problem sizes of interest. For more applied/empirical papers, the experiments are doing the heavy lifting of making a case. A survey just went out to Area Chairs asking to what degree code submission was taken as a factor in our recommendations to the senior program committee.

Absent a compelling reason not to submit code, I think that ensuring some form of reproducibility is important for both transparency and the open communication of ideas. Reviewers already approach reading a paper with some skepticism — the burden of proof is on the authors to make a compelling argument in their paper. But if the argument is largely empirical (e.g. “this heuristic works very well for problem A”) then the burden of proof consists of making a case that the experiments, as described in the paper, were in fact carried out and not mere fabrications. How better to do that than to provide the implementation of the method?

Providing implementations is not always possible: examples abound in multiple fields, including electrical engineering. In antenna design the schematic might be provided in the paper, but the actual fabricated antenna and anechoic chamber are not available to the reviewers. Nobody seems to think this is a problem: reviewers somehow trust that the authors are not making things up. Shouldn’t we trust ML authors as well?

One factor that makes a difference is that conferences are just not as competitive outside of computer science. Conferences have a short review period in which to evaluate a large volume of papers. The prestige conferred by getting a paper accepted to a top CS conference is often compared to getting a paper accepted to a top journal. Authors benefit a lot from the research community accepting their paper. It is only appropriate that they also share a lot.

Let’s take an example. Suppose you are working in academia and developed a new method for solving Problem X. You are going to launch a startup based on this method. How much more appealing would it be to funders if you had one (or more!) ICML papers about how you’ve totally nailed Problem X, showing that you are a total rockstar in the ML/AI community? But your competitive advantage might be at risk if reviewers (and then later the community) has access to your code. So then you write a paper where you discuss the main ideas behind your approach and give the experimental results but no implementation with the 5 other things you had to do to make the thing actually work. In this case you’re getting the stamp of approval while not sharing with the rest of the research community.

Of course, one can imagine that submissions from industry authors might rely on proprietary code bases which they cannot (for policy reasons) provide. An academic conference is about the open and free exchange of ideas, knowledge, and techniques. It seems that a trade show would be a more appropriate venue for showing the results without sharing the methods. I’m not trying to suggest that industry researchers are nefarious in some way, but it’s important to think about the incentives and benefits. The rules for submission (in this case code submission) articulate some of the values of the research community. Encouraging (but not requiring) code submission requires authors to signal (and allows reviewers to consider) whether they agree to the social contract.

Readings

I’m on sabbatical now, which ostensibly means having a bit more time to read things (technical and not). I’m not exactly burning through books but 2019 has already had a few good ones.

Convenience Store Woman (Sayaka Murata): The book has a lot of critical acclaim but I can see many readers being put off by it: the story is pretty disturbing in the end. The narrator and protagonist is (I think) neuro-atypical, which comes across in the writing (I’d love to read some notes from the translator). On the other hand, it is also a critique of Japanese work culture, I think, although not the usual office-drone/salaryman/Aggretsuko type, which is refreshing.

Golden Hill (Francis Spufford): This was a really great picaresque with a lot of detail about early New York that made me want to tour around lower Manhattan with a copy of the manuscript to trace out some of the locations. There’s a twist (as always!) but I don’t want to give it away. Spufford, as usual, has a real ear for the language: although it took a little getting used to, I eventually settled in and it was a real page-turner.

The Ministry of Pain (Dubravka Ugrešić): A novel by a Balkan author living in Amsterdam about a Balkan refugee teaching Balkan literature in Amsterdam to (mostly) other refugees. There’s a lot about language and the war and “our language” as she puts it. The story unfolds slowly but I think the atmosphere and ideas were what I appreciated the most about it. The discomfort of addressing while not reenacting trauma is palpable.

Binti: Home and Binti: The Night Masquerade (Nnedi Okorafor) books 2 and 3 in a sci-fi trilogy. The world expands quite a bit beyond the first one and I thought Binti’s character arc was quite dramatic. I wish there had been more to learn about the other characters as well. But these books are novellas so perhaps I should do a bit more work to fill in the gaps with my imagination?

Stranger in a Strange Land: Searching for Gershom Scholem and Jerusalem (George Prochnik): A rather discursive biography of Gershom Scholem, who almost single-handedly (it seems) made started the academic study of Kabbalah which is interleaved with the author’s autobiography of moving to Jerusalem, taking up graduate studies, starting a family, and becoming disenchanted. I thought it was a stretch at times to relate the two, and I had no prior information about Scholem but I found myself almost wanting two books: a straight biography and a straight memoir. Both had their merits but the alternation made it a bit of a slog to read.

gender inclusivity in communication models

I submitted a paper to ISIT in which I tried something different. It’s about communication models with a jammer, so there are three parties: Alice, Bob, and the jammer. Alice wants to send a message to Bob. The jammer wants to prevent it from being reliably received.

We always use Alice for the encoder/transmitter. Knowing nothing more than the name, I would use the pronouns she/her/hers to refer to Alice.

We always use Bob for the encoder/transmitter. Knowing nothing more than the name, I would use the pronouns he/him/his to refer to Bob.

What about the jammer? In previous papers (and in our research discussions) we called the jammer various names: Calvin (to get the C) or James (for the J). We ended up also using he/him/his for the jammer too.

This time I proposed we use Jamie for the jammer. Knowing nothing more than the name, I suggested they/them/their as the most appropriate. In my mind, Jamie may be gender nonconforming, right?

At this point many readers (if there are any) would say I’m being a bit on the nose. Why make James into Jamie and why deliberately change the pronouns? Won’t it just confuse people?

There are so many responses to this.

First, just on pragmatics. This makes pronouns which are uniquely decodable to the parties in the communication model. What can be clearer?

Second, if pronouns create a problem for a mathematically-minded reader, then they are far too obsessed with (gendered) Alice/Bob metaphor. It’s a mathematical engineering paper, not a kid’s story.

But finally, and most importantly, even though all the authors of this paper may be cis-gendered, writing the stories in our papers in a more inclusive way is the right thing to do. Why Alice and Bob? Why not Aarti and Bhaskar, Anting and Bolei, Avital and Binyamin, or Arash and Babak? I’ve heard arguments that we should be more ecumenical in the national origin of our communicating parties. Can we be more inclusive by gender as well?

We can and should!

What signals sent by author lists

I recently had a conversation about ordering of author lists for papers. Of course, each field has its own conventions but as people start publishing in multiple communities’ venues things can get a bit murky. There are pros and cons and different people have different values, etc.

This is all standard and has been hashed to death.

But what happens when you merge two papers with different author lists? Alphabetical makes things very easy, but if you go a different route, then the primary authors have to slug it out to see who gets first author credit. To split the difference, you could put a footnote saying that authors are in alphabetical order. In the conversation, it came up that putting the footnote implies that there was some tension between the two author groups and so this was the compromise solution after a debate. That was new to me: is this the correct inference to make in most cases?