Category: AI

AIAlgorithms

Gödel, Incompleteness and Privacy

Avi Wigderson has a nice talk on how Kurt Gödel’s incompleteness theorems bound what can and cannot be computed (or proved) by simple programs.

In this recent post I talked about how Gödel’s theorem was used to show that for certain kinds of learning algorithms, we can’t know definitively whether the algorithm learns the right thing or not. This is kind of equivalent to saying that we can’t definitively know whether there will be a gap in the program’s learning.

The flip side of this, as Wigderson points out, is that it is probably a good thing that there are certain things that are too hard for a program to figure out. This hardness is the key to privacy — the harder it is to decipher an encrypted message, the more you can have confidence in keeping the message content private. This principle is at the core of what allows e-commerce.

Perhaps there is a way to structure our online communications or transactions so that learning our behavior — in pernicious ways — becomes impossibly hard. This might diffuse a lot of the emerging fears surrounding AI.

Wigderson makes his point — what he calls the “Unreasonable usefulness of hard problems” — about 30 minutes into the talk.

Check out the “Unreasonable usefulness of hard problems” 30 minutes in.
AISocial Justice

Which cities use facial recognition?

San Francisco famously banned the use of facial recognition by police and other municipal authorities on May 14th of this year. Citizens in Detroit angered by the use of facial recognition in Project Green Light forced a moratorium on its use. Although Orlando has halted for an immediate deployment, a trial is being conducted involving police officers only. According to the Natalie Bednarz, the Digital Communications Supervisor in the Orlando office of Communications and Neighborhood Relations

if the City of Orlando Police Department decides to ultimately implement official use of the technology, City staff would explore procurement and develop a policy governing the technology

Email communication from the Orlando office of Communications and Neighborhood Relations

This report by Georgetown Law School reports that Chicago uses facial recognition in policing and throughout its mass transit systems.

Beyond surveillance cameras, several cities have been forced by ICE to turn over drivers license photos to ICE’s facial recognition software to identify persons who are not U.S. citizens. Not only is facial recognition software notoriously bad at identifying faces of African Americans, but systems score poorly in identifying people who identify ethnically as Latinx.

The Georgetown Law School in 2016 put together a list of city and state governments across the U.S. that use facial recognition.

Should facial recognition be banned altogether in policing?

AIData Scienceinclusion

Black In AI workshop call for papers

If you are a student, researcher, or professor at a Historically Black College or University and work actively in data science, machine learning, or artificial intelligence, please consider submitting a paper to the 2019 Black in AI workshop. The deadline is July 30 — I’d encourage submission even (especially!!) if your research and ideas are still coming together. There are also travel grants available and I’ll post that application soon.

The workshop occurs during the 2019 neurlps conference (this is probably the most attending conference on deep learning and other AI architectures). The specific goal of the workshop is to encourage involvement of people from Africa and the African diaspora in the AI field, and to promote research that benefits (and does no harm to) the global Black community.

I’ll include below the call for papers below.

And if you want to know more about data science with Black people in mind, I’m giving away two books on the subject! You can click this link or this one to claim one of them if you like.

Here’s the annoucement

Paper submission deadline: Tue July 30, 2019 11:00 PM UTC

Submit at: https://cmt3.research.microsoft.com/BLACKINAI2019

The site will start accepting submissions on July 7th.

No extensions will be offered for submissions.

We invite submissions for the Third Black in AI Workshop (co-located with NeurIPS). We welcome research work in artificial intelligence, computational neuroscience, and its applications. These include, but are not limited to, deep learning,  knowledge reasoning, machine learning, multi-agent systems, statistical reasoning, theory, computer vision, natural language processing, robotics, as well as applications of AI to other domains such as health and education, and submissions concerning fairness, ethics, and transparency in AI. 

Papers may introduce new theory, methodology, applications or product demonstrations. 

We also welcome position papers that synthesize existing work, identify future directions, or inform on neglected/abandoned areas where AI could be impactful. Examples are work on AI & Arts, AI & Policy, etc.

Submission will fall into one of these 4 tracks:

  1. Machine learning Algorithms
  2. Applications of AI 
  3. Position papers
  4. Product demonstrations

Work may be previously published, completed, or ongoing. The workshop will not publish proceedings. We encourage all Black researchers in areas related to AI to submit their work. They need not to be first author of the work.

Formatting instructions

All submissions must be in PDF format. Submissions are limited to two content pages, including all figures and tables. An additional page containing only references is allowed. Submissions should be in a single column, typeset using 11-point or larger fonts and have at least 1-inch margin all around. Submissions that do not follow these guidelines risk being rejected without consideration of their merits. 

Double-blinded reviews

Submissions will be peer-reviewed by at least 2 reviewers, in addition to an area chair. The reviewing process will be double-blinded at the level of the reviewers. As an author, you are responsible for anonymizing your submission. In particular, you should not include author names, author affiliations, or acknowledgements in your submission and you should avoid providing any other identifying information.

Travel grants

Use this link to apply for travel grants to the conference. They are available for eligible attendees, and should be submitted by  Wed July 31, 2019 11:00 PM UTC at the latest (Note that this is one day after the paper submission deadline).

Content guidelines

Submissions must state the research problem, motivation, and contribution. Submissions must be self-contained and include all figures, tables, and references. 

Here are a set of good sample papers from 2017: sample papers 

Questions? Contact us at bai2019@blackinai.org.

AIAlgorithmsMachine Learning

Gödel, Incompletness, and AI

Kurt Gödel was one of the great logicians of the 20th century. Although he passed away in 1978, his work is now impacting what we can know about today’s latest A.I. algorithms.

Gödel’s most significant contribution was probably his two Incompleteness Theorems. In essence they state that the standard machinery of mathematical reasoning are incapable of proving all of the true mathematical statements that could be formulated. A mathematician would say that that the consistency (or ability to determine which of any two contradictory statements is true) of standard set theory (a collection of axioms know as Zermelo–Fraenkel set theory) is independent of ZFC. That is, there some true things which you just can’t prove with math.

In a sense, this is like the recent U.S. Supreme Court decision on political gerrymandering. The court ruled “that partisan gerrymandering claims present political questions beyond the reach of the federal courts”. Yeah, the court stuck their heads in the sand, but ZFC just has no way to tell truth from falsity in certain cases. Gödel gives mathematical formal systems a pass.

It now looks like Gödel has rendered his ruling on machine learning.

A lot of the deep learning algorithms that enable Google translate and self driving cars work amazingly well, but there’s not a lot of theory that explains why they work so well — a lot of the advances over the past ten years amount to neural network hacking. Computer scientists are actively looking at ways of figuring out what machines can learn, and whether there are efficient algorithms for doing so. There was a recent ICML workshop devoted to the theory of deep learning and the Simons Institute is running an institute on the theoretical foundations of deep learning this summer.

However, in a recent paper entitled Learnability can be undecidable Shai Ben-David, Amir Yehudayoff, Shay Moran and colleagues showed that there is at least one generalized learning formulation which is undecidable. That is, although the particular algorithm might learn to predict effectively, you can’t prove that it will.

They looked at a particular kind of learning that in which the algorithm tries to learn a function that maximizes the expected value of some metric. The authors chose as a motivating example the task picking the ads to run on a website, given that the audience can be segmented into a finite set
of user types. Using what amounts to server logs, the learning function has to output a scoring function that says which ad to show given some information on the user. The scoring function learned has to maximize the number of ad views by looking at the results of previous views. This kind of problem obviously comes up a lot in the real world — so much so that there is a whole class of algorithms Expectation Maximization that have been developed around this framework.

One of the successes of theoretical machine learning is realizing that you can speak about a learning function in terms of a single number called the VC dimension which is roughly equivalent to the number of classes the items that you wish to classify can be broken into. They also cleverly use the fact that machine learning is equivalent to compression.

Think of it this way. If you magically could store all of the possible entries in the server log, you could just look up what previous users had done and base your decision (which ad to show) based on what the previous user had done. But chances are that since many of the users who are cyclists liked bicycle ads, you don’t need to store all of the responses for users who are cyclist to guess accurately which ad to show someone who is a cyclist. Compression amounts to successively reducing information you store (training data or features) as long as your algorithm performs acceptably.

The authors defined a compression scheme (the equivalent of a learning function) and were then able to link the compression scheme to incompleteness. They were able to show that the scheme works if and only if a particular kind of undecidable hypothesis called the continuum hypothesis is true. Since Gödel proved (well, actually developed the machinery to prove) that we can’t decide whether the continuum hypothesis is true or false, we can’t really say whether things can be learned using this method. That is, we may be able to learn an ad placer in practice, but we can’t use this particular machinery to prove that it will always find the best answer. Machine learning and A.I. are by definition intractable problems, where we mostly rely on simple algorithms to give results that are good enough — but having certainty is always good.

Although the authors caution that it is a restricted case and other formulations might lead to better results, there are some two other significant consequences I can see. First, the compression scheme they develop is precisely the same structure that are used in Generative Adversarial Networks (GANs). The GAN neural network is commonly used to generate fake faces and used in photo apps like Pikazo http://www.pikazoapp.com/. The implication of this research is that we don’t have a good way to prove that a GAN will eventually learn something useful. The second implication is that there may be no provable way from guaranteeing that popular algorithms like Expectation Maximization will avoid optimization traps. The work continues

It may be no coincidence that the Gödel Institute is in the same complex of buildings as the Vienna University AI institute.

Next door to the Gödel Institute is the Vienna AI institute

Avi Wigderson has a nice talk about the connection between Gödel’s theorems and computation. If we can’t event prove that a program will be bug free, then we shouldn’t be too surprised that we can’t prove that a program learns the right thing.

A nice talk by Avi Wigderson. Sometimes hacking is all you got.
AIAlgorithmsAtlanta

The city of Atlanta doesn’t use facial recognition — so why does Delta Airlines?

I recently made an inquiry with the City of Atlanta’s Mayor’s office as to the use of facial recognition software. I received the following reply on the Mayor’s behalf from the Atlanta Police Department

The Atlanta Police Department does not currently use nor the capability to perform facial recognition. As we do not have the capability nor sought the use of, we not have specific legislation design for or around facial recognition technology.

Delta Airlines, a company based in Atlanta, continues to promote the use of facial recognition software, and according to this wired article makes it difficult for citizens to opt out of its use.

There are several concerns with use of facial recognition technology, succinctly laid out by the Electronic Frontier Foundation:

Face recognition is a method of identifying or verifying the identity of an individual using their face. Face recognition systems can be used to identify people in photos, video, or in real-time. Law enforcement may also use mobile devices to identify people during police stops. 

But face recognition data can be prone to error, which can implicate people for crimes they haven’t committed. Facial recognition software is particularly bad at recognizing African Americans and other ethnic minorities, women, and young people, often misidentifying or failing to identify them, disparately impacting certain groups.

Additionally, face recognition has been used to target people engaging in protected speech

Electronic Frontier Foundation at https://www.eff.org/pages/face-recognition

So in other words, the technology has the potential for free assembly and privacy abuses and because the algorithms used are typically less accurate for people of color (POC), the potential abuses are multiplied.

There are on going dialogs (here is the U.S. House discussion on the impact on Civil Liberties) on when/how/if to deploy this technology.

Do me a favor? If you happen to fly Delta, or are a member of their frequent flyer programs, could you kindly ask for non-facial recognition check in? Then asking for more transparency on the use and audit of the software used would be an important step forward.

AIAlgorithmsSocial Justice

San Francisco passes facial recognition ordinance

San Francisco recently passed an ordinance controlling the use of facial recognition in the city. The ordinance was in large part thanks to the pioneering research of Joy Buolamwini.

The argument against the technology is twofold: first, the technology is highly invasive in public spaces and may constitute a direct threat to basic (US) constitutional rights of freedom of assembly; secondly the feature extraction and training set construction methodologies (for newer deep learning based models) have been shown to have racial and gender biases “baked in”. For example, the systems analyzed in Buolamwini’s work are less accurate for Black people and women — either because the data sets used for training include mostly white male faces, or the image processing algorithms focus on image components and make assumptions more common to European faces.

Consider uses in policing, where an inaccurate system mis-identifies a Black or LatinX person as a felon. Especially when there is no transparency into the use or internals of such systems, the chances for abuse and injustice are in incredible. Despite these concerns, Amazon shareholders think it is ok to release the technology on the public.

Do you know if such a system is deployed in your city? If so, are there measures to control its use, or make audits available to your community? If not, have you considered contacting your elected representatives to support or discuss appropriate safeguards?

AIHistorically Black Collegesinclusion

AI at HBCUs Fall 2018

If you teach/study computer science at a Historically Black College or University, or know of someone who does, please check out and pass along the site https://charlescearl.github.io/ai-hbcu/.

I put it up a year or so ago to document the work being done at those institutions to increase the impact and participation of the African Diaspora in shaping the way that AI technologies are developed and used. As this recent report by the ACLU points out, yes algorithmic racism is still a thing.

If you know of important initiatives, interesting classes, or discussions going on at HBCUs around this issue, please feel empowered to check out the repository https://github.com/charlescearl/ai-hbcu and send a pull request. Or drop a comment below.

I was encouraged to see that the Neural Information Processing conference (one of the most attended AI conferences) is taking steps towards inclusion. They keep promising to change their name.