NeurIPS 2020 has come to an end, but what we learned last week regarding AI and ML will be applied in 2021 and beyond.
Wow! What a week it has been for our engineering crew. NeurIPS 2020 was seven days of amazing insights and deep, thought-provoking presentations. We met a lot of excited engineers looking for ways to change the world, and it has definitely been an invigorator for our whole team. It was also an excellent learning experience, and our team wanted to share a few of our key themes and learnings from the NeurIPS 2020 conference.
AI is reshaping the way we build software and conduct research
We would like to start by highlighting an informative talk by Dr. Chris Bishop, the Laboratory Director at Microsoft Research in Cambridge and Professor of Computer Science at the University of Edinburgh, who dedicated his talk to the artificial intelligence (AI) revolution that is happening right now. AI has helped us accomplish great things and solve many challenging problems. However, Bishop argues that the real revolution exits in the way we are approaching software development and research.
One example that he gave is that in the last 40 years, engineers and software developers were busy programming computers. This means that in order to accomplish a specific task using a computer program, developers had to explicitly define each part of the software by writing many lines of code that list out all of the actions that the computer needs to take in order to achieve its final goal. However, Bishop highlights that in the next 40 years, instead of programming computers, engineers and software developers will be busy training them.
This is the core notion of Machine Learning, the highly regarded AI methodology behind our successful product EnsoSleep.
One of the main advantages of machine learning (ML) is that it provides the ability to solve many non-programmable problems by training the software on a large amount of data which allows it to learn the task in hand without being explicitly programmed with many lines of code. Consequently, it is not a surprise to us that before Machine Learning was used to solve challenging problems in healthcare, robust automation of Sleep Staging seemed impossible.
Machine Learning Enables Further Customization of AI Algorithms
Furthermore, Bishop highlighted that Machine Learning allows a level of customization that wasn’t available before. With Machine Learning, each AI model can be fine-tuned and tailored to the user’s needs and preferences. Many AI solutions work to improve the user experience and increase efficiency. In the past, this level of customization would have been extremely expensive to implement and maintain. Bishop also highlighted the following quote by Dr. Felix Nensa, consultant at the Institute of Diagnostic and Interventional Radiology and Neuroradiology at University Essen, Germany:
“AI will not replace radiologists, but radiologists who do not use AI will be replaced by those who do,” said Dr. Felix Nensa at NeurIPS 2020.
This is an extremely important quote that emphasizes the potential of AI to empower clinicians, allowing them to work more efficiently and dedicate more of their time to their patients and their treatment plans. This is one of our main goals at EnsoData, making our AI technologies accessible to clinicians worldwide and help them adopt our AI tools so they can efficiently use them in their practice. You can find Bishop’s full talk on the NeurIPS 2020 site.
Creating bias free, fair, and generalizable AI is on everyone’s mind
Another main recurring theme during NeurIPS 2020 was the emphasis on the importance of creating robust AI models and algorithms that are free of any bias or discrimination across race, gender, and different social groups. During the conference, Dr. Charles Isbell, Dean of Computing at Georgia Tech, gave an excellent talk dedicated to highlighting how these types of issues are affecting various AI research and products.
Moreover, Isbell highlighted that bias and discrimination has existed in technology throughout history. For example, in the past, photography companies chose to use specific photographic film material and contrast calibration techniques that produced the best quality images for light skin subjects which, in the past, used to be the dominant market. Furthermore, audio recording devices targeted lower frequency ranges that captured best deep low frequency male voices thus, generating discrimination against the female gender. Although these types of devices have drastically improved throughout the years and the digital age enabled the generalizability towards different race and gender groups, newer AI technologies, in many cases, still lack this much needed advancement.
One recent example of an AI that failed the bias test is an AI algorithm called PULSE. The PULSE algorithm uses AI methods to up-sample low resolution images. During the summer of 2020, this algorithm gained a lot of attention on social media after users started to post examples that raised concerns for this algorithm. In one example, a low-resolution photo of former president Barack Obama was used as the input to the algorithm, and the output received was an image of a white man. You can read more about this example in The Verge. While this is just one example, many other developers have fallen victim to unforeseen biases in their algorithms. Cognizant organizations put a constant focus on paying attention to these possible biases.
Addressing the Bias issue in AI head on
Luckily, Isbell highlighted several approaches that could help with solving these types of issues. First, he emphasized the importance of having a deep understanding of the problem needed to be solved and the data used to train the AI models. Second, Isbell mentioned that having deep knowledge of the AI methods used to achieve a specific task can help develop different approaches that would be able to correctly handle bias and increase generalizability in a rigorous manner. Lastly, Isbell highlighted the importance of collaboration between AI researchers, domain experts, and product users.
The experts and the users themselves both have different views and domain knowledge that can greatly assist with identifying the areas where bias may exist and help mitigate them. The talk about bias and discrimination didn’t end with Isbell’s talk though. At EnsoData, we were very glad to observe how important these issues are for other AI engineers and researchers throughout the conference. It’s equally exciting to see the vast amount of research papers that are dedicated to solving these types of issues in AI.
For our EnsoData AI team, it is important to us to be extremely conscious of the subject of bias and generalizability. We constantly validate our algorithms while controlling not only for different age and gender populations, but also different disease severity groups, and constantly explore new ways to mitigate issues of bias. We highly encourage all of our readers to check out Isbell’s extremely informative and well-done virtual talk.
Symposium: COVID-19 Symposium
It wouldn’t be a 2020 conference if COVID-19 wasn’t a featured topic. The year of the coronavirus has required organizations of all industries to adopt their strategies or risk going out of business. For many scientists and researchers, what we can expect in 2021 is a major question. That’s what brought the COVID-19 symposium speaker to NeurIPS 2020. Dr. Michael Mina, MD, PhD, of Harvard Medical School, spoke at length on the strategies we’re using to battle the virus in the NeurIPS presentation kicking off the symposium: “Covid-19 Testing Strategies for Disease Surveillance and Control.”
According to Dr. Mina, the most recent wave, featuring the late November, early December spike in cases and fatalities, was expected. Per Dr. Mina, the coronavirus is a seasonal disease, one that is at its strongest from October through December. Does that mean we’re in the clear? Of course not! With the holiday season in full swing, Americans are traveling at much larger volumes, meaning the natural coronavirus increase due to seasonality will likely be buoyed by an equally inopportune travel time period.
From there, Dr. Mina re-emphasized what we know and what we still don’t know about testing and screening for COVID. Below are the four types of tests that we might use to see if someone has COVID or not: medical diagnostics, surveillance testing, entrance screening and public health screening. Each group of tests has a designed purpose, and it’s up to us to better use our available resources to determine COVID infection rates.
If you’d like to dive into more of Dr. Mina’s thoughts, this press conference transcript and audio is chock full of information on the current coronavirus testing situation.
That’s a wrap on NeurIPS 2020!
While there’s obviously hundreds of other key takeaways from the weeklong virtual NeurIPS 2020 conference, the above three examples are topics we thought you really NEEDED to know about after this week.