Hold the Code #21
What's 9 + 10?
Hello!
Welcome to Hold The Code, edition #21.
This week, we cover news about how AI was used to restore a Rembrandt painting, LinkedIn's biased job-matching algorithm, an app that gathers open-sourced data, and an opinion piece titled "What Artificial Intelligence Still Can't Do."
AI Solves A 17th Century Problem: Restoring a Rembrandt
In 1715, one of Rembrandt's most well-celebrated works, The Night Watch, was destroyed when city officials tried to move it to Amsterdam's city hall. Unable to fit the painting through its doors, they cut large panels from the sides — but the fragments were lost after removal.
But now, centuries later, the painting has been made complete through the use of artificial intelligence. In 2019, the museum embarked on a multi-year, multi-million-dollar restoration project, referred to as Operation Night Watch, to recover the painting. The museum tapped its senior scientist, Rob Erdmann, to head the effort using three primary tools: the remaining preserved section of the original painting, a 17th-century copy of the original painting attributed to Gerrit Lundens that had been made before the cuts, and AI technology.
AI was used to solve a set of specific problems:
First, the copy made by Lundens is one-fifth the size of the original, which measures almost 12 feet in length.
Second, Lundens painted in a different style than Rembrandt, which raised the question of how the missing pieces could be restored to an approximation of how Rembrandt would have painted them.
Erdmann created three separate neural networks, training computers to learn how to do specific tasks to address the problems.
“The first [neural network] was responsible for identifying shared details. It found more than 10,000 details in common between The Night Watch and Lundens’s copy.”
For the second, Erdmann said, “Once you have all of these details, everything had to be warped into place,” essentially by tinkering with the pieces by “scoot[ing one part] a little bit to the left” and making another section of the painting “2 percent bigger, and rotat[ing another] by four degrees. This way all the details would be perfectly aligned to serve as inputs to the third and final stage. That’s when we sent the third neural network to art school.”
Comparing Solutions for Biased Job-Matching AIs
“You typically hear the anecdote that a recruiter spends six seconds looking at your résumé, right?” says Derek Kan, vice president of product management at Monster. “When we look at the recommendation engine we’ve built, you can reduce that time down to milliseconds.”
Years ago, LinkedIn discovered that the recommendation algorithms it uses to match job candidates with opportunities were producing biased results. The algorithms were ranking candidates partly on the basis of how likely they were to apply for a position or respond to a recruiter. The system wound up referring more men than women for open roles simply because men are often more aggressive at seeking out new opportunities.
LinkedIn discovered the problem and built another AI program to counteract the bias in the results of the first. Meanwhile, some of the world’s largest job search sites—including CareerBuilder, ZipRecruiter, and Monster—are taking very different approaches to address bias on their own platforms.
LinkedIn's Approach: their new AI ensures that before referring the matches curated by the original engine, the recommendation system includes a representative distribution of users across gender.
Monster's Approach: the marketing team at Monster focuses on getting users from diverse backgrounds signed up for the service, and the company then relies on employers to report back and tell Monster whether or not it passed on a representative set of candidates.
ZipRecruiter's Approach: the company’s algorithms don’t take certain identifying characteristics such as names into account when ranking candidates; instead they classify people on the basis of 64 other types of information, including geographical data.
Our takeaway
While it's hard to measure which of these solutions best resolves the issue of biased job-matching algorithms, it is essential to note the immense impact recommendation engines have on the job process. If you find that you're struggling to identify and connect with new opportunities, it's worth diversifying your efforts across multiple job-board websites, aiming to leverage the algorithm solution that best represents your profile.
Selling Intelligence
Need some pocket money? The San Francisco-based data company Premise is paying its app users (who are often from developing countries like Afghanistan) small amounts of money to complete menial tasks such as snapping photos, filling out surveys, or even just taking a walk. Why? These humble tasks provide the company valuable data, which they will then turn into business intelligence for its clients ranging from private companies seeking consumer information to the U.S. military.
Notably, since 2017 Premise has received more than $5 million from the U.S. military to work on projects like:
assessing U.S. information operations effectiveness
scouting & mapping out key social structures like mosques, banks, and internet cafes
covertly monitor cell-tower and wifi signals
Although Premise "fully informs" users of its public terms of service, it does not explicitly inform users of which clients their data is sent to.
"[The company] said tasks needed to be designed to “safeguard true intent” — meaning contributors wouldn’t necessarily be aware they were participating in a government operation."
Mr.Blackman, Premise's CEO, says the open-source data gathered by such menial tasks is "available to anyone who has a cell phone" and neither unique nor secret. One user in Afghanistan said he's paid about 25 cents per task, a typical compensation.
Should people be able to know (and choose) who gets to use their data & for what purposes? What is the price of information?
Read the full article here.
Weekly Feature: AI-n't There Yet
AI has made incredible advancements over the past few decades, with AI-powered systems now able to compose music, accurately diagnose medical conditions, and write poetry. However, there are still a number of things that AI can’t do, which we will discuss in the following story. If you want to learn more about the limitations of AI, you can read the full article here.
Use common sense
Something that humans use every day that AI still struggles with is common sense. Common sense is derived from our own understanding of the world through persistent mental representations that we can understand and interact with.
For example, take the following sentences:
“A man went to a restaurant. He ordered a steak. He left a big tip.”
If you ask a person what the man ate, they would most likely answer with “steak.” However, an AI algorithm would struggle to answer this, since nowhere in the prompt does it explicitly say the man ate a steak.
The reason we, as humans, are able to come to this conclusion is that we have a mental representation of a restaurant, a place where people eat, and an order, something the customer intends to eat when they are at the restaurant. Without these understandings, AI models have trouble determining what the appropriate answer to the question would be.
“The absence of common sense prevents an intelligent system from understanding its world, communicating naturally with people, behaving reasonably in unforeseen situations, and learning from new experiences,” says Dave Gunning of DARPA. “This absence is perhaps the most significant barrier between the narrowly focused AI applications we have today and the more general AI applications we would like to create in the future.”
Learn continuously
Another thing that AI models struggle with is continuous learning and adapting. Currently, the development of an AI system is split into two distinct steps: training and development.
In the training phase, the AI is fed a static dataset, where it uses this data to inform its model.
In the deployment phase, the AI uses this model on new data to make predictions, find relationships, or produce categorizations.
This two-step process means that after the training phase, the AI model does not learn anymore and, therefore, cannot adapt to new or evolving situations. To update the model after the training phase, we would have to retrain the model with a new dataset and then redeploy it (which can be extremely time-intensive and computationally expensive).
Although work is being done on a solution to continuous learning, there are still many obstacles, the biggest of which is called catastrophic forgetting. This occurs when new information given to an AI system interferes with or overwrites previous learning the model has done.
Understand cause & effect
AI also has trouble understanding cause and effect. AI models are great at identifying patterns in data, but when it comes to developing causal relationships, these systems fall short.
This is reflected in the mathematical basis of AI. These systems, in essence, think in terms of X=Y, which can be mirrored to Y=X without any meaningful change between the two equations. However, saying X is caused by Y is not the same as saying Y is caused by X.
For example, a system could be fed data on when the sunrises and when a rooster crows. It would easily be able to determine that a rooster crowing is statistically likely to occur when the sunrises, however, it would not be able to distinguish between if the sunrise caused the rooster to crow or if the rooster crow caused the sun to rise.
Reason ethically
If you are at all familiar with our newsletter, this last one should come as no surprise. AI systems often fall short when it comes to ethical reasoning. This likely stems from two main factors:
AI systems are often amoral: While AI may not act on malice or be actually immoral, they have an apparent lack of moral guidelines. The most infamous example of this is probably Microsoft’s Twitterbot, Tay, which learned and used extremely racist and derogatory language in its tweets before it was shut down.
Human values are difficult to quantify: Our values are often nuanced, contradictory, and vague. This makes it difficult to encode these values into AI algorithms.
Promising work on this front has focused on building AI that figures out human values based on human behaviors. This idea is based on the concept of inverse reinforcement learning. This method allows algorithms to learn about our human values by analyzing our behavior patterns (the opposite of reinforcement learning where algorithms learn certain behaviors based on the existing environment).
Taking this concept one step further, AI theorist Eliezer Yudkowsky has introduced the term coherent extrapolated volition. This concept would produce AI systems that, in theory, would “act in our best interests according to not to what we presently think we want, but rather according to what an idealized version of ourselves would value.”
However it is done, developing ethical AI (or, even better, AI that can reason ethically itself) is of utmost importance as AI’s reach continues to expand through our society.
Written by Larina Chen, Molly Pribble, and Lex Verb