Hello, World!
Welcome (back) to Hold the Code. This week, we cover a wide range of topics from how AI can customize music, fly military jets, and improve the Netflix algorithm. With all of this variety, we are sure there is something for everyone in this edition.
As always, happy reading!
Apple’s AI DJ
Apple recently acquired the London-based company AI Music. Their products use AI to modify music to match your taste and your environment.
Contextual AI: An Introduction
AI Music uses contextual AI to power its music-making. Contextual AI interprets real-world, in-situ data to generate an output. AI Music uses information about the user and the environment to create a unique music experience from existing songs.
Imagine you listen to a song in the morning while you sip your coffee. The beat is slower, and the vibe may be more acoustic. Now, imagine you listen to the same song while you’re at the gym. Now you get a deep-house, upbeat version to motivate you during the last stretch of your cardio workout.
“[We are] shape-changing music to shift the way songs are consumed rather than generating music from scratch.” says AI MUsics CEO, Siavash Mahdavi.
What’s up at Apple?
In recent years, Apple has slowed down the number of acquisitions it has made. However, the last two companies they have purchased are focused on music. Prior to AI Music, Apple acquired classical music service Primephonic. Apple has not made a statement yet about their most recent acquisition, but many are speculating that they are looking to launch new auditory experiences across platforms like Apple Music, Fitness+, or TV+.
AI in the Air (Plus, Guns!)
With scores of Artificial Intelligence technologies making large strides, it’s no surprise that the military has taken notice. A group of companies have been developing an algorithm that will change aerial dogfighting forever by adding in AI pilots. The implementation of these AI agents, however, poses a problem.
Calibrated Trust
When a human fighter pilot was pitted against an algorithm called Falco in a simulation, they lost all five skirmishes fought. This seems to bode well for the technology. The goal of AI fighter agents, however, is not that they will be doing everything themselves, or at least not quite yet. The AI algorithm that is currently being tested is supposed to work with pilots, where the humans take a back seat and intervene when necessary. Using a the concept of calibrated trust, the algorithm feeds information to the human pilot up to four seconds before following through on their reported actions, letting their pilot decide whether or not those moves are actually a good idea and take control back if needed.
Ethics in Targeting
The US military is hopeful that AI will address the problem of target misidentification, but there are still biases within facial recognition software (specifically regarding recognizing the gender of women with darker skin) that suggests that these softwares need to be examined more closely before they should be used alongside lethal weapons. And the amount of trust that pilots and citizens have in these automated technologies has to be gained slowly. If it’s lost, even once, it takes a lot of time to build it back up again. Lethal autonomous weapons are also ethically worrying. We rely on machines in a lot of different situations, but is it a good idea to trust them to make decisions about what- or who- to shoot at in a war?
The trust of fighter pilots in their AI algorithms might fare differently on the actual battlefield, where life and death truly are on the line. Maybe the pilots are right about not putting an unwarranted amount of trust in these algorithms, maybe they will be a useful tool. But if these technologies do end up in the military, they could change the way that wars are fought forever.
Weekly Feature: Demystifying Data Science with Netflix
Data science has become a bit of a buzzword, with companies boasting their data science divisions and new articles coming out daily on how data science is the way of the future. But what exactly do data scientists do? Martin Tingley at Netflix’s Tech Blog has offered insight to that question, explaining how data scientists at Netflix use experimentation to determine where the company should go next.
So what do they do?
While they often have their own divisions, data scientists at Netflix hardly work alone. Partnering with product managers, engineers, and business teams, they determine what problems the company faces and then use their statistical knowledge to identify possible solutions. And while they may not have the same kind of work as traditional physical scientists, the method remains the same. Netflix data scientists employ the scientific method of observation, hypothesis, testing, and conclusion to maintain an intricate understanding of user interactions with the platform.
Why is this helpful? This detailed knowledge of customer patterns and desires allows Netflix to quickly determine what changes to the platform will truly improve user experience.
How do they do this?
The gold standard for drawing conclusions in data science and statistics is a method called A/B testing. This method works by first developing a hypothesis, such as “customers are more likely to watch a movie when we use algorithm A to recommend movies instead of algorithm B.”
The next step is to develop a test to determine whether there is sufficient evidence to support or reject the hypothesis. In A/B testing, a subgroup of Netflix users is split randomly into 2 groups, where one is given algorithm A and the other algorithm B. The key here is in the randomization. By randomly splitting the participants into groups A and B, any other factors that may affect movie watching, such as age, gender, or free time, will be evenly divided between the two groups.
Then the data scientists can look at the two groups and statistically determine whether they think any difference in movie watching is caused by the algorithm change.
What challenges are there?
While this is the general outline, data scientists at Netflix must also keep in mind the broader picture of what they are doing, as well as work through cases where traditional A/B testing may not be possible. Because many of the recommendations data scientists suggest would be costly to enact, they must be very cautious of Type-I errors, where there appears to be a relationship where one doesn’t actually exist. Consider advertising for example. If it is determined that running certain ads will increase membership when in reality the ads are ineffective, then the money spent on advertising will have been fruitless.
This is just one example of the many nuances that data scientists must be aware of when drawing conclusions from their data, with other examples including handling situations where the suspected impact of a change is very close to the current value, or situations where they’re not able to run true A/B testing such as in partnership opportunities. In all of these situations, the value of their statistical knowledge and understanding of the field they work in both become key factors in their success.
With data science being applied to everything, from advertising to payment methods, understanding what it is that data scientists actually do can be key to understanding how companies are continuing to move forward.
Love HTC?
Follow RAISO (our parent org) on social media for more updates, discussions, and events!
Instagram: @Raisogram
Twitter: @raisotweets
RAISO Website: https://www.raiso.org
Written by Molly Pribble, Hope McKnight, and Arielle Michelman
Edited by Molly Pribble