Joho the BlogOctober 2017 - Joho the Blog

October 19, 2017

[liveblog] AI and Education session

Jenn Halen, Sandra Cortesi, Alexa Hasse, and Andres Lombana Bermudez of the Berkman Klein Youth and Media team are leading about a discussion about AI and Education at MIT Media Lab as part of the Ethics and Governance of AI program jointly at the Harvard’s Berkman Klein Center for Internet & Society and the MIT Media Lab.

NOTE: Live-blogging. Getting things wrong. Missing points. Omitting key information. Introducing artificial choppiness. Over-emphasizing small matters. Paraphrasing badly. Not running a spellpchecker. Mangling other people’s ideas and words. You are warned, people.

Sandra gives an introduction the BKC Youth and Media project. She points out that their projects are co-designed with the groups that they are researching. From the AI folks they’d love ideas and better understanding of AI, for they are just starting to consider the importance of AI to education and youth. They are creating a Digital Media Literacy Platform (which Sandra says they hope to rename).

They show an intro to AI designed to be useful for a teacher introducing the topic to students. It defines, at a high level, AI, machine learning, and neural networks. They also show “learning experiences” (= “XP”) that Berkman Klein summer interns came up with, including AI and well-being, AI and news, autonomous vehicles, and AI and art. They are committed to working on how to educate youth about AI not only in terms of particular areas, but also privacy, safety, etc., always with an eye towards inclusiveness.

They open it up for discussion by posing some questions. 1. How to promote inclusion? How to open it up to the most diverse learning communities? 2. Did we spot any errors in their materials? 3. How to reduce the complexity of this topic? 4. Should some of the examples become their own independent XPs? 5. How to increase engagement? How to make it exciting to people who don’t come into it already interested in the topic?

[And then it got too conversational for me to blog…]

Be the first to comment »

October 16, 2017

How to screw up a succah. In a good way.

I know you’re all wondering how I was able to build such a magnificent succah, and how I managed to combine inexpensiveness with convenience. But most of all, you’re wondering what the hell is a succah?

A succah is essentially a temporary Jew shack that you eat in during the holiday of Succos (AKA Sukkot). It has to meet certain requirements that make it somewhat sturdier than a pillow fort: It has to be temporary, covered incompletely on top, closed on at least three sides, etc. If you’re an observant Jew, as elements of my family are, you eat all your meals out there during the 8-day holiday. Some Jews even sleep in them. Far more commonly, the custom is to have guests as often as possible so that meals are extended and highly social. In some Jewish communities, succah-hopping is a thing. A good thing.

For the past 20+ yrs, I’ve been constructing it out of the same set of PVC pipes. I have a rubber mallet, which is comical enough that I should probably have bought it from Acme Hardware, which I use to bang poles into T-fittings. (For the middle uprights, they’re T-s with a third sleeve in the third dimension, which sounds way more complex than it actually is.)

This is fine except for my constant anxiety about wind overcoming the friction that holds the slippery tubes into their slippery connectors. So, every year after I’ve pounded the poles together — and, if you try to visualize the process you’ll see that pounding a tube into one sleeve unpounds it from the sleeve at the other end — I’ve drilled a hole through the sleeve and tube and inserted a weenie nail, just to add some charming shrapnel to the explosion when the wind suddenly tosses it apart like a child knocking down a house made of drinking straws.

So, I did some research and this year built a succah using a remarkable breakthrough in applied physics: threaded connectors. Here’s how.

Our succah is 10′ x 10′. Each side wall consists of two corner uprights, one upright in the middle, and four horizontal poles. The uprights have have fittings with a threaded nut. The horizontals have fittings that screw into the nuts. The fittings are glued on to the poles using PVC glue. You simply screw all the pieces together.

It’s a little more complicated than that, though, because everything is. The threaded fittings are sleeves. But because they’re all designed to connect to lengths of pipe the way you might want to connect one garden hose to another, you can’t use them to connect pipes perpendicularly. But every joint in this construction connects a horizontal to an upright, which means you need 90-degree turns.

So you get yourself some plain old fittings, like the ones I used in the prior version. You attach them to the uprights. But those fittings are sleeves designed to join two pipes. The threaded fittings are also sleeves. How do you join a sleeve to a sleeve? With a pipe! So, for each join, cut a 2″ piece of pipe. Glue one end into the sleeve on the upright. Glue the other end to the threaded end of the threaded joins. Press them in so that they’re flush. Below is an example where the connector was a little too long, so the joins are not flush, purely for illustrative purposes I assure you:

Now assemble the pipes. Our uprights are 7′. The horizontals are 47″ each, which, with the additional lengths imposed by the fittings, worked out to about 10′. But if you need exactness, you should cut them to fit. Just remember to label them so they’ll go together next year. Also, wear eye protection: the pipes cut easily with a circular saw, but it creates a lot of flying plastic jaggies.

Here’s the invoice for the fittings:


You might want to get ourself a spare or two. I’m still amazed that I got away without needing one.

Note that the outer rings tighten counter-clockwise. You have to get the pieces lined up pretty well to be able to screw them together. I suggest that you assemble it from the ground up so that you won’t have to expand magic suspending pieces in mid-air.

The succah worked out well. It seemed pretty robust for a plastic structure made out of pipes not intended for that purpose. It disassembled quite easily. My only concern is how many years we’ll get out of the threaded pieces; they seem rugged but so did I once. (Actually, I didn’t.)

1 Comment »

October 10, 2017

[liveblog][bkc] Algorithmic fairness

I’m at a special Berkman Klein Center Tuesday lunch, a panel on “Programming the Future of AI: Ethics, Governance, and Justice” with Cynthia Dwork, Christopher L. Griffin, Margo I. Seltzer, and Jonathan L. Zittrain, in a discussion moderated by Chris Bavitz.

NOTE: Live-blogging. Getting things wrong. Missing points. Omitting key information. Introducing artificial choppiness. Over-emphasizing small matters. Paraphrasing badly. Not running a spellpchecker. Mangling other people’s ideas and words. You are warned, people.

They begin with brief intros of their interests:

Chris Griffin: One of the big questions for use of algorithms in the justice system is what: is the alternative? Human decision making has its own issues.

Margo Seltzer: She’s been working on transparent models. She would always prefer to be able to get an index card’s worth of explanation of how a machine learning system has come up with its output.

Cynthia Dwork: What is our definition of fairness, and how might we evaluate the fairness of our machine systems? She says she’s not that big a fan of insisting on explanations.

Jonathan Zittrain: What elements of this ought to be contracted out? Can we avoid the voting machine problem of relying on a vendor we don’t necessarily trust? Also, it may be that expalantions don’t help us that much. Also, we have to be very wary of biases built into the data. Finally, AI might be able to shed light on interventions before problems arise, e.g., city designs that might lower crime rates.

Chris Bavitz: Margo, say more about transparency…

Seltzer: Systems ought to be designed so that if you ask why it came up with that conclusion, it can tell you in a way that you can understand. Not just a data dump.

Bavitz: The legal system generally expects that, but is that hard to do?

Seltzer: It seems that in some cases you can achieve higher accuracy with models that are not explicable. But not always.

Dwork: Yes.

Zittrain: People like Cynthia Rudin have been re-applying techniques from the 1980s but are explainable. But I’ve been thinking about David Weinberger’s recent work [yes, me] that reality may depend on factors that are deeply complex and that don’t reduce down to understandable equations.

Dwork: Yes. But back to Margo. Rule lists have antecedents and probabilities. E.g., you’re trying to classify mushrooms as poisonous or not. There are features you notice: shape of the head, odor, texture, etc. You can generate rules lists that are fairly simple: if the stalk is like this and the smell is like, then it’s likely poisonous. But you can also have “if/else” conditions. The conclusions can be based on very complex dependencies among these factors. So, the question of why something was classified some way can be much more complicated than meets the eye.

Seltzer: I agree. Let’s say you were turned down for the loan. You might not be able to understand the complex of factors, but you might be able to find a factor you can address.

Dwork: Yes, but the question “Is there a cheap and easy path that would lead to a different outcome?” is a very different quesiton than “Why I was classified some particular way?””

Griffin: There’s a multi-level approach to assessing transparency. We can’t expect the public to understand the research by which a model is generated. But how is that translated into scoring mechanisms? What inputs are we using? If you’re assessing risk from 1 to 6, does the decision-maker understand the difference between, say, a 2 and 3?

Zittrain: The data going in often is very reductive. You do an interview with a prisoner who doesn’t really answer so you take a stab at it … but the stabbiness of that data is not itself input. [No, Zittrain did not say “stabbiness”].

Griffin: The data quality issue is widespread. In part this is because the data sets are discrete. It would be useful to abstract ID’s so the data can be aggregated.

Zittrain: Imagine you can design mushrooms. You could design a poisonous one with the slightest variation from edible ones to game the system. A real life example: the tax system. I think I’d rather trust machine learning than a human model that can be more easily gamed.

Bavitz: An interviewer who doesn’t understand the impact of the questions she’s asking might be a feature, not a bug, if you want to get human bias out of the model…

Seltzer: The suspicion around machine algorithms stems from a misplaced belief that humans are fair and unbiased. The combination of a human and a machine, if the human can understand the machine’s model, might result in less biased decisions than either on their own.

Bavitz: One argument for machine learning tools is consistency.

Griffin: The ethos of our system would be lost. We rely on a judicial official to use her or his wisdom, experience, and discretion to make decisions. “Bias could be termed as the inability to perceive with sufficient clarity.” [I missed some of this. Sorry.]

Bavitz: If the data is biased, can the systems be trained out of the bias?

Dwork: Generally, garbage in, garbage out. There are efforts now, but they’re problematic. Maybe you can combine unbiased data with historical data, and use that to learn models that are less biased.

Griffin: We’re looking for continuity in results. With the prisoner system, the judge gets a list of the factors lined up with the prisoner’s history. If the judge wants to look at that background and discard some of the risk factors because they’re so biased, s/he can ignore the machine’s recommendation. There may be some anchoring bias, but I’d argue that that’s a good thing.

Bavitz: How about the private, commercial actors who are providing this software? What if these companies don’t want to make their results interpretable so as not to give away their special sauce?

Dwork: When Facebook is questioned, I like to appeal to the miracle of modern cryptography that lets us prove that secrets have particular properties without decrypting them. This can be applied to algorithms so you can show that one has a particular property without revealing that algorithm itself. There’s a lot of technology out there that can be used to preserve the secrecy of the algorithm, if that were the only problem.

Zittrain: It’d be great to be able to audit a tech while keeping the algorithm secret, but why does the company want to keep it secret? Especially if the results of the model are fed back in, increasing lock-in. I can’t see why we’d want to farm this out to commercial entities. But that hasn’t been on the radar because entrepreneurial companies are arising to do this for municipalities, etc.

Seltzer: First, the secrecy of the model is totally independent from the business model. Second, I’m fine with companies building these models, but it’s concerning if they’re keeping the model secret. Would you take a pill if you had no idea how it worked?

Zittrain: We do that all the time.

Dwork: That’s an example of relying on testing, not transparency.

Griffin: Let’s say we can’t get the companies to reveal the algorithms or the research. The public doesn’t want to know (unless there’s litigation over a particular case) the reasoning behind the decision, but whether it works.

Zittrain: Assume re-arrest rates are influenced by factors that shouldn’t count. The algorithm would reflect that. What can we do about that?

Griffin: The evidence is overwhelming about the disparity in stops by race and ethnicity. The officers are using the wrong proxies for making these decisions. If you had these tools throughout the lifespan of such a case, you might be able to change this. But these are difficult issues.

Seltzer: Every piece of software has bugs. The thought of sw being used in way where I don’t know what it thinks it’s doing or what it’s actually doing gives me a lot of pause.


Q: The government keeps rehiring the same contractors who fail at their projects. The US Digital Service insists that contractors develop their sw in public. They fight this. Second, many engineering shops don’t think about the bias in the data. How do we infuse that into companies?

Dwork: I’m teaching it in a new course this semester…

Zittrain: The syllabus is secret. [laughter]

Seltzer: We inject issues of ethics into our every CS course. You have to consider the ethics while you’re designing and building the software. It’s like considering performance and scalability.

Bavitz: At the Ethics and Governance of AI project at the Berkman Klein Center, we’ve been talking about the point of procurement: what do the procurers need to be asking?

Q: The panel has talked about justice, augmenting human decision-making, etc. That makes it sound like we have an idea of some better decision-making process. What is it? How will we know if we’ve achieved it? How will models know if they’re getting it right, especially over time as systems get older?

Dwork: Huge question. Exactly the right question. If we knew who ought to be treated similarly to whom for any particular classification class, everything would become much easier. A lot of AI’s work will be discovering this metric of who is similar to whom, and how similar. It’s going to be an imperfect but improving situation. We’ll be doing the best guess, but as we do more and more research, our idea of what is the best guess will improve.

Zittrain: Cynthia, your work may not always let us see what’s fair, but it does help us see what is unfair. [This is an important point. We may not be able to explain what fairness is exactly, but we can still identify unfairness.] We can’t ask machine learning pattern recognition to come up with a theory of justice. We have to rely on judges, legislators, etc. to do that. But if we ease the work of judges by only presenting the borderline cases, do we run the risk of ossifying the training set on which the judgments by real judges were made? Will the judges become de-skilled? Do you keep some running continuously in artesinal courtrooms…? [laughter]

Griffin: I don’t think that any of these risk assessments can solve any of these optimization problems. That takes a conversation in the public sphere. A jurisdiction has to decide what its tolerance for risk is, what it’s tolerance is for the cost of incarceration, etc. The tool itself won’t get you to that optimized outcome. It will be the interaction of the tool and the decision-makers. That’s what gets optimized over time. (There is some baseline uniformity across jurisdictions.)
Q: Humans are biased. Assume a normal distribution across degrees of bias. AI can help us remove the outliers, but it may rely on biased data.

Dwork: I believe this is the bias problem we discussed.

Q: Wouldn’t be better to train it on artificial data?

Seltzer: Where does that data come from? How do we generate realistic but unbiased data?

Be the first to comment »