Import AI 210: Satellite collisions & ML; helping the blind navigate with Lookout; why Deepfakes are the most worrying AI threat

Deepfakes are the most worrying-crimes – UCL researchers:
Deepfakes, specifically audio/video impersonation of someone for criminal purposes, are the greatest AI-driven crime threat, according to research from UCL published in the journal Crime Science. The research is based on a two-day workshop that occurred in early 2019, which had 31 attendants from the public sector (including the UK’s National Cyber Security Center), academia, and the private sector. At the workshop, attendees shared research on a variety of different AI-driven crime threats, then got together and ranked 20 of them from low to high threats, across four dimensions for each crime (harm, profit, achievability, defeatability).

The top threats: The things to be most worried about are audio/video impersonation, followed by tailored phishing campaigns and driverless vehicles being used as weapons.
The least worrying threats: Some of the least worrying threats include forgery, AI-authored fake reviews, and AI-assisted stalking, according to the attendees.
Things that make you go ‘hmmm’: Some of the threats that required decent text generation capabilities were ranked as being fairly hard to achieve – I wonder how much that threat landscape has changed, given the NLP advancements in past year and a half (e.g, GPT2, GPT3, CTRL, et cetera).
  Read the research: AI-enabled future crime (BMC Crime Science, open access).
  Read more: ‘Deepfakes’ ranked as most serious AI crime threat (UCL News).

###################################################

SpaceCraft collision detection – surprisingly hard for ML:
…Competition results mean ML isn’t the greatest fit for averting Kessler Syndrome, yet…
How well can machine learning approaches predict the possibility of satellites colliding with one another – not well, according to the results of a competition, the Spacecraft Collision Avoidance Challenge, hosted by the European Space Agency. In a writeup of the competition, ESA-affiliated researchers describe the challenge (try to predict satellite collisions via a dataset of satellite-specific data files called “conjunction data messages” that store data about satellite events).

The results: First, out of 97 teams that entered, only 12 managed to beat a time series prediction baseline, illustrating the difficulty of this problem. Many of the teams experimented with ML, but it’s notable that the top-ranking team eschewed a standard machine learning pipeline for something far more involved, combining some ML with a series of if/then operations. The team that ranked third overall used a purer ML approach (specifically, a ‘Manhattan-LSTM’ based on a siamese network). A nice thing about this competition was the inclusion of a reassuringly hard baseline, which should give us confidence in techniques that beat the baseline.

What do to for next time: “Naive forecasting models have surprisingly good performances and thus are established as an unavoidable benchmark for any future work in this area and, on the other hand, machine learning models are able to improve upon such a benchmark hinting at the possibility of using machine learning to improve the decision making process in collision avoidance systems,” they write.
  Read more: Spacecraft Collision Avoidance Challenge: design and results of a machine learning competition (arXiv).

###################################################

Have poor vision? Use the ‘Lookout’ app to help you:
…How machine learning can help the partially-sighted…
Google has used machine learning to improve an application targeted at partially-sighted people, named Lookout. The new features mean that ‘ ‘when the user aims their smartphone camera at the product, Lookout identifies it and speaks aloud the brand name and product size’. This may be particularly useful to partially sighted people trying to accomplish daily tasks, like shopping.

What goes in it? Lookout relies on MediaPipe, a Google-developed ML development stack. Each instance of a Lookout app will ship geographically-curated information on around two million popular products to users’ phones, so they can get help when they’re walking around.

Why this matters: world navigators: Apps like ‘Lookout’ are part of a genre of AI applications which I’ll call ‘world navigators’ – they make it easier for a certain type of person to navigate the world around them. Here, it’s using ML to make it easier for partially sighted people to get around. In other use cases, like Google’s Translate app, the same technology can make it easier for people to speak in other languages. In a few years, I think AI tools will have made it easier for us to generally ‘translate’ the world for different people, helping us use ML to improve the lives of people.
  Read more: On-device Supermarket Product Recognition (Google AI Blog).
  Get the app here (Lookout by Google, official Play store listing).

###################################################

How hard is it to be an ethical AI developer these days? Pretty hard, says researcher:
…A tale of two APIs…
Roya Pakzad, an independent AI researcher, has written about some of the ethical challenges developers face when using modern AI tools. In a detailed post, Pakzad imagines that Unilever wants to analyze the emotions present in social media tweets about the company, following its (unsuccessful) debut of a ‘fair and lovely‘ skincare campaign in India (which advertised a skincare product on the basis of it being good at lightening skin town).

Pakzad tries to do what a developer would do. First, she finds some as-a-service AI tools that could help her do her task – IBM Tone Analyzer and ParalletDots Text Analysis systems – then tests out those APIs. After registering for both services, she looked at how the different services classified different tweets using the ‘fair and lovely’ term in reference to the campaign – discouragingly, she found massive divergence between the IBM and ParallelDot API results, highlighting the brittleness of these systems and how capabilities vary massively across different APIs. Pakzad also looks into the different privacy and security policies of the different services, again highlighting the substantial differences between them.

Pakzad’s top tips for providers and developers:
– Providers should ensure their APIs are well documented, communicate about issues of bias and fairness and security directly, and develop better systems for preserving developer privacy.
– ML practitioners should carefully analyze APIs prior to using them, test against benchmark datasets that relate to potentially discriminatory outcomes of ML projects, share ethical issues about the API via opening pull requests on the dev’s GitHub page (if available), and be clear about the usage of the API in documentation about services it is used within.
  Read more: Developers, Choose Wisely: a Guide for Responsible Use of Machine Learning APIs (Medium)

###################################################

Top tips for AI developers, from A16Z:
…It’s the long tail distributions that kill you…
Venture capital firm Andreessen Horowitz – famous for its co-founder Marc Andreessen’s ‘software is eating the world’ observation – thinks that AI companies are becoming more important, and has written some tips for people trying to put machine learning techniques into production in a startup context.

A16Z’s tips:
– ML is as much about iterative experimentation as standard software engineering.
– It’s the long-tail part of the distribution that you’ll spend most of your time tuning your ML for. “ML developers end up in a loop – seemingly infinite, at times – collecting new data and retraining to account for edgy cases,” they write. And trying to solve the long tail can lead to AI firms exhibiting diseconomies of scale.
Break problems into sub-components: Big all-in-one models like GPT3 are a rarity – in production, most problems are going to be solved by using a variety of specialized ML models targeted at different sub parts of large problems.
Your operation infrastructure is your ML infrastructure: Invest in the tools you run your ML on, so consolidate data pipelines, develop your own infrastructure stack (without reinventing too many wheels), test everything, and try to compile and optimize models.

Why this matters: Machine learning is transitioning from an academic-heavy artisanal production and development phase, to a scaled-up process-oriented phase; posts like these tips from A16Z illustrate that this shift is occurring as they try to take implicit knowledge from ML practitioners and make it explicit for a wider audience.
Read more: Taming the Tail: Adventures in Improving AI Economics (Andreessen Horowitz blogpost).

###################################################

AI governance – lessons from history:
…What might the AI equivalent of the 1957 space treaty look like?…
What does space, the ethics of IVF, and the Internet governance org ICANN have in common? These are all historical examples of some of the surprising ways countries, companies, and individuals have collaborated to govern emerging science and technology. In a blog post, Verity Harding, a researcher at the University of Cambridge, lays out some of the ways we can learn from history to make (hopefully) fewer mistakes in the governance of AI.
  Given that it’s 2020 and multinationalism is struggling under the twin burdens of COVID and the worsening tensions between the US and China, you might expect such research to have a grave tone. But the post is surprisingly optimistic: “The challenges are great, and the lessons of the past cannot be simply superimposed onto the present. What is possible geopolitically, however, is one example where AI scientists, practitioners and policymakers can take heart from historical precedent,” writes Harding.

Historical examples: In the post, Harding discusses examples like the 1967 UN Outer Space Treaty, the UK’s Warnock Committee and Human Embryology Act, the Internet Corporation for assigned Names and Numbers (ICANN), and the European ban on genetically modified crops.

Why this matters: AI governance is going to receive a lot of attention in the coming years as the technology gets cheaper, more capable, and more widely available (see: predictions about AI surveillance, geopolitics, and economics here), we’re going to need to deal with increasingly hard challenges for the management of the technology. If we study more historical examples, we’ll hopefully make fewer mistakes as we muddle our way forward.
  Read more: Lessons from history: what can past technological breakthroughs teach the AI community today (University of Cambridge, Bennett Institute for Public Policy).

###################################################

Tech Tales:

Teenage drone phone:
[2025 The outer, outer suburbs of Boston, Massachusetts.]

Her Dad didn’t want her to talk to the boy, so he blocked her internet and installed a bunch of software onto her phone so she couldn’t access certain websites or contact numbers not ‘whitelisted’ by her father. She sat in her room staring at the calendar app on her phone, counting down the days till she turned 16 and she’d get the run of her own phone.

It was about a month before her birthday when she was tidying her room and found the drone in her closet. She’d got it for christmas a couple of years earlier. She took it out and stared at it. Then looked at her window.

The next day in school she asked the boy where he lived.
Why are you asking? He said
So I can break into your house, she said.
He told her his address.

That night, she looked at her house and his house on Google Maps, while browsing the specifications of her drone.
I’m going to do some night photography! She told her Dad
Let me know if you see any birds, he said, and don’t go beyond the fence.
Sure thing, Dad, she said.

Outside, she went to the edge of the garden, by the fence. Then she turned the drone on and pointed it at the trees, and her house, and then her. She hit record.
She told the drone some things, but really she was speaking to the boy. With the footage recorded, she sent the drone on a pre-planned route to the boy’s house. She hoped he’d be smart and go out. She could see occasional images beamed back to her via the drone, which was piggybacking on a bunch of networks. She saw the boy come out. Saw his eyes focus on the part of a drone where she’d taped a post-it note that said “insert cable here, or I really will break into your house”. Watched him go inside and come back out with a cable. A minute later, the drone told her someone had access its data and made a copy. She told it to flash its takeoff lights, and get up when it was safe to do so.

Any birds? Her Dad said when she came back in.
No, but I only checked a couple of the trees, I’ll do the others tomorrow, she said.
Sounds good to me, he said.

And just like that, she had a way to talk to the boy outside of school. The next day she brought the drone into school and had him pair his phone with it when he was nearby. “Now you can record on it when it comes to your house,” she said.
Where do you live? He said.
Are you crazy? What if you were some kind of criminal. It knows where I live, she said, then winked.

That night, she went out into the garden, and hovered the drone again. It went over to the boy’s house and it dropped down and the boy recorded a message on it and the drone came back, across the city. Later that night, she found a bird nest in one of the trees. She told her Dad about it and he said she’d need to be careful, but if she could film the baby birds when they were in the nest, that would interest him. She said yes, so she and the boy could keep sending drones to eachother. 

When she got older she reflected on all of this – the drone working like a tin can on the end of a string, and of herself filming the baby birds in their nest, and of her dad monitoring her in phoneworld – which by then was as much a part of reality for teens as anything physical. She resented some of it and was thrilled by other parts. And she knew she fit into it, using her drone to go and study the boy, and in a way study herself as she learned to use her own intelligence to use the tools of the world to study those around her, so she could have something that seemed like control or connection.

Things that inspired this story: The consumerization of drones; miniaturization of battery power overtime; consumerization; reductions in prices of display screens and onboard AI computation devices.