Reflection on the 2020 Virtual World Summit AI
I recently got the chance to attend World Summit AI Americas … which means that as a mere high schooler, I was able to spend two days sitting in the basement listening to some of the leading experts on AI talk about their work and got the chance to talk to a few of them!!
Due to the COVID-19 pandemic, the conference was moved to a virtual format and I want to take a moment to thank the InspiredMinds team who made that possible. In fact, the experience was almost richer because it was online. Instead of only having a few minutes to talk to people in between talks, I was able to schedule 30 minutes one on one calls with the speakers and other conference-goers.
The topics covered, as you can imagine, were highly diverse and approached from many different angles given that AI is so intersectional. Climate change, fighting bias in AI, robotic sensing, and ML-driven bug fixes were only a few of the research areas.
Something I was really struck by was the emphasis on social good and ethical responsibility.
It can be easy with such a ubiquitous and powerful technology to simply leverage it for profit or power, but I appreciated everyone’s commitment to protecting the privacy and using AI to make a positive difference in the world.
I want to share just a snapshot of the crazy things people are using AI for right now.
Climate Change:
Climate change is obviously a major pressing issue and a prime use case for AI because there is an abundance of climate and weather data.
Alexis Hannart, CSO at Axionable, an AI consulting company supporting sustainable businesses discussed two current applications. Wind and solar renewable energy fluctuate as does energy consumption needs. Current weather models generally predict 24–36 hour time frames leaving an opportunity gap for short term forecasting. Being able to predict short term wind and solar fluctuates is necessary for us to be able to reliably use a renewable grid.
He also talked about using CNNs to identify early hurricanes and other storms and predict their severity and location. While key now, this will become increasingly important as climate change causes severer storms.
At BrainBox AI, CTO Jean-Simon Venne talked out their work in implementing a proactive HVAC system. The current heating and cooling system simply kick in whenever the temperature goes outside of its normal range, however, this forces it to overcompensate and causes the temperature to fluctuate between too low and too high and thus expends a lot of energy. Their solution uses data already collected by the sensors, runs it through an LSTM, and proactively kicks in the HVAC system. Their fully autonomous system was highly accurate and saved significant amounts of energy.
Sasha Luccioni, a postdoctoral researcher and Director of Scientific Projects in AI for Humanity, also talked about numerous applications, many of them focused on raising awareness and getting more people involved.
She worked to identify crossways between problems and technology (information at climatechange.ai) and provided resources and ideas for jumping-off points. Some areas of research she pointed out were: optimizing grid distribution, improving battery technology, optimizing transportation flow and supply chains, mapping emissions on farms, precision agriculture, planting seeds using drones based on humidity and wind, and tracking climate change coverage in the media.
She is also working to turn models’ predictions for what the future will look like into images that people can see. Searching your location in 50 years and seeing what 5 inches of flooding looks like makes the impacts more tangible.
And finally, she talked about the GHG emissions caused by these ML algorithms. There’s a website here that enables you to calculate your algorithm’s emissions so you can be more aware and maybe spend less time training your algorithm or use a more sustainably powered server.
Ethics:
AI relies on data and there are so many things that can go wrong.
Privacy — Where the data comes from matters as does keeping the data secure. What should be the rules or policies in place to protect us? Is it acceptable to scrape images or should you have to pay for them?
Bias — Since the algorithm is trained on data, if that data is biased or not diversified enough then the algorithm will be inaccurate and perpetuate biases. If you train a model to predict whether or not you will want to hire someone, it may accept fewer women as that is what your data portrayed. Audrey Boguchwal, Senior Product Manager at Samasource, talked about a study demonstrating that a computer vision algorithm that had a hard time identifying pedestrians with dark skin simply because it wasn’t given enough pictures of people with dark skin to be able to recognize them as people. That’s a major problem if our self-driving cars can’t identify some types of people as people. Thus, it’s critical to have high quality, diversified data, and continual learning for an algorithm to limit bias.
Trust — What will it take for us to be able to trust AI? The point many of the speakers made was that AI is such a broad term that it really doesn’t make sense to classify and make decisions on AI as a whole. We already trust AI with many things from navigation to Siri and Alexa to routing data packages and no one has a problem with that. However, there are still many concerns, some logical some not that need to be addressed.
The key to figuring out these problems is close collaboration between technical specialists, policy makers, and the public to make sure everyone is educated and that proper precautions are taken.
Robotics:
Robotics is a great tool to understand and collect data on the natural world. School of Computer Science Director at McGill, Gregory Dudek, talked about how in many ways we have covered size, deployability, and agility, but there are still many obstacles to a fully autonomous intelligent robot.
It is difficult for robots to choose what data they should collect. He discussed a few methods to tackle this problem:
- The scientist specifies what is interesting. This requires the scientist to clearly be able to articulate the reward and decision-making process so that the robot can implement Bayesian Reward Learning.
- Search for the most diverse data. In this, the robot is just allowed to wander and sends back the 6 most diverse images it finds. The issue with this is that not all diversity is good. An underwater robot taking footage of a coral sent back an image of the sand simply because it was different from the coral and an image virtually the same as another except it had a wire dangling in it. While both of these are unique, they are not useful data.
- Navigate to the most interesting thing. The robot was allowed to follow the most interesting part of a photo. Here it too followed novelty and after seeing one thing for a while, it got bored and moved on. However, not all of this was useful, for example, partway through it got distracted by a turtle. So the question is, how does the robot determine what is interesting in the context of its general mission?
- Patch detection. In the previous example when a robot followed something, it can identify the object across different frames using a CNN. Therefore, a scientist wanting to teach it what they find interesting can just label the one object interesting or not interesting and the robot will know that all of the other associated frames should have the same label.
On a completely separate note, Rami Wehbe, CTO and Cofounder of GlobalDWS, talked about their work using service robots to help with senior care. Their robots have “vision”, language processing, speech, and movement abilities to provide companionship and and care for seniors.
I also was able to talk to Farhad Rahbarnia, Cofounder of Notos Technologies. We talked about their work improving the efficiency of drones using biomimicry and ML to understand air currents and temperature gradients to determine the most energy-efficient flight path. Their technology allows drones to run for longer with less battery.
Some other fun things:
Danny Tarlow is an engineer on the Google Research Brain Team looking at using AI to provide bug fix suggestions to developers. We talked about how code has structure and thus is best analyzed using a graph network instead of compressing it into linear data. In that way, it’s kind of like molecules where the structure tells you something beyond the exact formula. I also learned that there are a lot of security issues around which code you’re allowed to train on and there’s a fine balance between getting a proof of concept version of the product out for testing and making it accurate enough that it is helpful rather than hurtful.
I talked with Celeste Kidd, a psychologist at UC Berkley, whose lab is focusing on understanding how we learn. She discovered that our confidence in something is mostly based on recent validation. If you search something, after watching 3 3-minute videos all saying the same thing, your confidence that it is true will be extremely high and you will treat it as fact. Once you reach that point, it is nearly impossible to convince you otherwise. This issue is particularly relevant in an age with so much information online and an increasing amount of fake news. Her lab now is working to figure out ways to prevent people from coming to these decisions as quickly.
On the more technical side, Samy Bengio talked about Google’s Brain Teamwork on understanding why deep learning works. He argued that this is a really important question so that we can understand how the algorithms work and also when they won’t work.
One question is is the algorithm remembering the training data or creating generalizations? The traditional idea has been that with more parameters, the training data error will decrease, but after a certain point, the test data error will increase due to over-fitting.
They ran a test in which one algorithm was given a set of labeled data and another replica was given the same set of data except with random labels. In both cases, the training data was identified 100% accurately while the test data as labeled accurately for the correctly labeled data and not for the randomly labeled data. This means that even when the training data accuracy moves to 0, the test data can still decrease and instead of the test data error curve looking like a U, it decreases after the interpolation threshold. This is good because it means our models can be more accurate and we don’t need to worry as much about over-fitting.
The other question he discussed was whether all layers in a Neural Net are equally important. By replacing some layer’s input parameters with the initially set ones, they were able to discover that some layers were critical while others were robust. This is promising because it means only the critical layers really matter and that the others have a large degree of flexibility.
Overall this was an amazing experience, both listening to the talks and meeting with people and I learned so much about the varied applications of AI and the important ethical and security issues that ML developers have to deal with.
Shoutout to Isabella Grandic, the TKS Toronto Team, and Michael Raspuzzi for giving me the opportunity to attend! And thank you to everyone who took the time to talk with me during the conference!
