supervised_learning – AI in Media and Society

‘Ground truth’ and labeled data

May 4, 2021 Mindy McAdams

Cassie Kozyrkov, who wrote this article, is head of decision intelligence at Google. It starts out with what looks like a standard explanation of an image-recognition system — which she deprecatingly refers to as the “the cat/not-cat task.” But don’t be fooled — Kozyrkov communicates with clear, sharp precision, and very quickly she asks us to consider circumstances in which we would want a tiger to be considered a cat and those in which we would want it to be not-cat.

This leads to a discussion of ground truth. This is “an ideal expected result” — but for whom? Well, for the people who originally built the system. Kozyrkov notes that ground truth is NOT an objective, perfect truth like something studied in a philosophy class (Truth with a capital T). It’s whether a tiger is a cat in your reality or not-cat in mine.

I am reminded of one of my favorite lines in the rock opera Jesus Christ Superstar: “But what is truth? Is truth unchanging law? We both have truths. Are mine the same as yours?”

“When such a dataset is used to train ML/AI systems, systems based on it will inherit and amplify the implicit values of the people who decided what the ideal system behavior looked like to them.”
— Cassie Kozyrkov

It also brings to mind the practice of testing for intercoder reliability — standard practice in research that relies on qualitative data. (More about that here.)

Say you are using an existing labeled dataset — not one you yourself have created — which is often the case. The labels attached to the data items are the ground truth for that dataset. If it’s a dataset of images, and some labels applied to photos of people are racist, then that’s the ground truth in that dataset. If it’s a dataset for sentiment analysis, and a lot of toxic comments are labeled “not toxic,” then that’s the ground truth you’re adopting.

It’s essential for developers to test systems extensively to uncover these flaws in ground truth.

“You wouldn’t want to fall victim to a myopic fraud detection system with sloppy definitions of what financial fraud looks like, especially if such a system is allowed to falsely accuse people without giving them an easy way to prove their innocence.”
— Cassie Kozyrkov

In a video embedded in the same article, Kozyrkov pithily proclaims: “There are only actually two real lines there. Here’s what they are: This objective. That data set.” (At 9:16.) Of course there’s a ton more code than that (she’s talking about the programming of the system that creates the model), but in terms of what you want the system to be able to do, that’s it in a nutshell: How have you framed your objective? And what’s in your dataset? More important, in many cases, is what’s NOT in your dataset.

She says this is where the core danger in AI lies, because in traditional programming “it might take 10,000 lines of code, a hundred thousand lines of code maybe, and some human being has to worry about every single one of those lines, agonize over it.” With supervised machine learning, you’ve only got the objective and the (gigantic) dataset, and the question is, Have enough people with expertise really agonized over each of those things?

My other favorite bits from the video:

“A system that is built and designed for one purpose may not work for a different purpose.” (6:17)
“Remember that the objective is subjective.” (6:31)
“And if you take those two parts really seriously, that is how you are going to build a safe and effective and kind AI system.” (20:16)

AI in Media and Society by Mindy McAdams is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Include the author’s name (Mindy McAdams) and a link to the original post in any reuse of this content.

Comment moderation as a machine learning case study

September 7, 2020September 7, 2020 Mindy McAdams

Continuing my summary of the lessons in Introduction to Machine Learning from the Google News Initiative, today I’m looking at Lesson 5 of 8, “Training your Machine Learning model.” Previous lessons were covered here and here.

Now we get into the real “how it works” details — but still without looking at any code or computer languages.

The “lesson” (actually just a text) covers a common case for news organizations: comment moderation. If you permit people to comment on articles on your site, machine learning can be used to identify offensive comments and flag them so that human editors can review them.

With supervised learning (one of three approaches included in machine learning; see previous post here), you need labeled data. In this case, that means complete comments — real ones — that have already been labeled by humans as offensive or not. You need an equally large number of both kinds of comments. Creating this dataset of comments is discussed more fully in the lesson.

You will also need to choose a machine learning algorithm. Comments are text, obviously, so you’ll select among the existing algorithms that process language (rather than those that handle images and video). There are many from which to choose. As the lesson comes from Google, it suggests you use a Google algorithm.

In all AI courses and training modules I’ve looked at, this step is boiled down to “Here, we’ll use this one,” without providing a comparison of the options available. This is something I would expect an experienced ML practitioner to be able to explain — why are they using X algorithm instead of Y algorithm for this particular job? Certainly there are reasons why one text-analysis algorithm might be better for analyzing comments on news articles than another one.

What is the algorithm doing? It is creating and refining a model. The more accurate the final model is, the better it will be at predicting whether a comment is offensive. Note that the model doesn’t actually know anything. It is a computer’s representation of a “world” of comments in which some — with particular features or attributes perceived in the training data — are rated as offensive, and others — which lack a sufficient quantity of those features or attributes — are rated as not likely to be offensive.

The lesson goes on to discuss false positives and false negatives, which are possibly unavoidable — but the fewer, the better. We especially want to eliminate false negatives, which are offensive comments not flagged by the system.

“The most common reason for bias creeping in is when your training data isn’t truly representative of the population that your model is making predictions on.”
—Lesson 6, Bias in Machine Learning

Lesson 6 in the course covers bias in machine learning. A quick way to understand how ML systems come to be biased is to consider the comment-moderation example above. What if the labeled data (real comments) included a lot of comments offensive to women — but all of the labels were created by a team of men, with no women on the team? Surely the men would miss some offensive comments that women team members would have caught. The training data are flawed because a significant number of comments are labeled incorrectly.

There’s a pretty good video attached to this lesson. It’s only 2.5 minutes, and it illustrates interaction bias, latent bias, and selection bias.

Lesson 6 also includes a list of questions you should ask to help you recognize potential bias in your dataset.

It was interesting to me that the lesson omits a discussion of how the accuracy of labels is really just as important as having representative data for training and testing in supervised learning. This issue is covered in ImageNet and labels for data, an earlier post here.

ImageNet and labels for data

August 19, 2020September 3, 2020 Mindy McAdams

Supervised learning is a type of machine learning in which a model is trained using labeled data. You begin with a very large collection of labeled data. (In the case of ImageNet, the data were all digital images. For the Iris Data Set, the data all refer to individual iris flowers, which can be divided into three related species. For the MNIST dataset, the data are images of about 70,000 handwritten numbers, 0 through 9.)

You divide the dataset into two parts, the training data and the test data. The split might be 70/30, or 80/20. You don’t choose which data goes into which group. Then you run the training data many, many, many times, adjusting certain parameters in the code along the way, until the code consistently returns good results — that is, the thing the code identifies (an object in an image, an iris species, a number) matches the label (which is hidden from the code).

At that point, you have a trained model. You feed the test data set to it and see whether the accuracy rate is also high. (It’s important that none of the test data were used to train the model.) Again, the proof is in the labels.

In a later post I will discuss how data come to be labeled. (Hint: It’s not elves.) In this post, I will discuss bad labels. Specifically, I want to highlight the work that AI researcher Kate Crawford and artist-researcher Trevor Paglen did around the famous ImageNet dataset.

In the video above, Crawford and Paglen present this work and show a lot of great examples. They also published a long article about the work, if you’d rather read than watch.

ImageNet is a huge collection of labeled images. More than 14 million images. They were labeled according to a set of categories and synonym groupings from WordNet, an English-language lexical database. The images were labeled by humans.

And that, it seems, is at the root of the problem.

Crawford and Paglen were interested in the ImageNet photos of people. Person is a category in WordNet. Within the category, there are many descriptive terms for people, such as “cheerleaders, scuba divers, welders, Boy Scouts, fire walkers, and flower girls.” So the photos of people in ImageNet are labeled with these terms. However, not all terms are neutral.

“A young man drinking beer is categorized as an ‘alcoholic, alky, dipsomaniac, boozer, lush, soaker, souse.’ A child wearing sunglasses is classified as a ‘failure, loser, non-starter, unsuccessful person.’”
—Crawford and Paglen

You might say, well, where’s the harm? They are only labels in a database, after all.

The ImageNet database has been used to train many convolutional neural networks used in image-recognition software.

When you feed a photo of yourself into an image-recognition application, you might be surprised at the labels that are applied to you. For example, an image of Paglen (a white man with a shaved head) was labeled as “Klansman, Ku Kluxer.”

Paglen built a web app called ImageNet Roulette so that anyone could upload a photo of themselves or a friend and see what labels were applied. (The app is no longer online.) It became clear that perfectly innocuous people in photos were being labeled as criminals or dangerous, or with racist or sexist terms.

About 952,000 of ImageNet’s 14 million images were in the person category as of 2010 (source). Many of those images — with their labels — were removed after the opening of Crawford and Paglen’s art exhibition, Training Humans, in Milan in September 2019.

ImageNet has been used to train countless image-recognition systems since 2010.

Additional information:

Leading online database to remove 600,000 images after art project reveals its racist bias (September 2019), The Art Newspaper.

Ask a computer to draw what it sees

August 12, 2020August 24, 2020 Mindy McAdams

If a computer can correctly identify an object (an apple, a tricycle) or an animal such as a zebra, can it produce a drawing of that object or animal? This is something most people can do, even if their drawing skills are minimal. After all, almost anyone can play Pictionary.

This 8-minute video shows us what happened when a programmer-artist reversed the process of an AI that recognizes objects and animals in digital images. I really admire the deft storytelling here.

Object recognition has improved amazingly in the past 10 years, but that does not mean these AI systems see the same way as a human does. In some cases, that might not matter at all. In other cases, it can mean the difference between life and death.

In yesterday’s post I mentioned the way a convolutional neural network (part of a machine learning system) processes an image through many stacked layers of detection units (sometimes called neurons), identifying edges and shapes that eventually lead to a conclusion that the image is likely to contain such-and-such an object, animal, or person. Today’s video shows a bit more about the training process that an AI goes through before it can perform these identifications.

Training is necessary in the type of machine learning called supervised learning. The training data (in this case, digital images of objects and animals) must be labeled in advance. That is, the system receives thousands of images labeled “tiger” before it is able to recognize a tiger in a random photo or video. If a system can identify 20 different animals, that system was trained on thousands of images of each animal.

If the system was never trained on tigers, it cannot recognize a tiger.

So today’s video gives us a nice glimpse into how and why that training works, and what its limitations are. What’s really fascinating to me, though, are the images produced by programmer-artist Tom White‘s system.

“I have created a drawing system that allows neural networks to produce abstract ink prints that reveal their visual concepts. Surprisingly, these prints are recognized not only by the neural networks that created them, but also universally across most AI systems which have been trained to recognize the same objects.”
—Tom White

In the video, you’ll see that humans cannot recognize what the AI drew. The rendering is too abstract, too unlike what we see and what we would draw ourselves. Note what White says, though, about other AI systems: they can recognize the object in these AI-produced drawings.

This is, I think, related to what is called adversarial AI, which I’ll discuss in a future post.