Richard Cotton covers four AI stories: Johns Hopkins researchers use AI to predict cardiac arrest from heart scans up to ten years early. The indie thriller Fall used deepfake technology to replace over thirty curse words with milder ones, shifting the film from R to PG-13 without reshoots. Princeton researchers expose a reproducibility crisis in machine-learning studies, finding 329 errors across fields.
Chapters
Hi there everyone. Welcome to this week's episode of the Weekly Roundup your go to for the latest and greatest in data science. This week we're talking about predicting cardiac arrest 10 years in the future, the reproducibility crisis in research, and how deepfakes are helping movies get to PG13 ratings in post production. Make sure you're subscribed so that you never miss out on another update and let us know in the comments what your favourite story of the week is. In our first story this week, researchers at Johns Hopkins University have developed an artificial intelligence approach they say can help predict if and when a person could suffer cardiac arrest. By analyzing the patterns of scar tissue that can develop in the cardiac muscle over time due to heart disease, the researchers found that AI algorithms could offer accurate predictions of a person's risk up to 10 years in advance. After gathering years of heart scans from hundreds of Johns Hopkins patients with cardiac scarring, the researchers said the AI was able to outperform physicians predictions that relied on simple visual estimates of scar features like their total volume and mass. Dan Popescu, a former Johns Hopkins doctoral student, says the images carry critical information that doctors haven't been able to access. The scarring can be distributed in different ways and it says something about a patient's chance for survival. There's information hidden in it. These kinds of predictions help doctors give the right care to the right people, and we're super excited about how this AI might directly improve and possibly lengthen people's lives. While many tech companies are feeling the economic pinch at the moment, our next story takes a look at how hiring is still full steam ahead for machine learning roles in the American pharmaceutical industry. Global Data, a company that tracks job listings from key companies in various sectors, says that North American roles make up 63% of the total worldwide pharmaceutical machine learning job market. Key hotspots identified in the data were Cambridge, Massachusetts, Boston and San Francisco. This fits with recent data that shows that the pharmaceutical industry as a whole is ramping up its data science hiring plans, with 39% of pharmaceutical companies hiring for data science related roles, according to Global Data. We've also seen an uptick in job postings for data science roles, and hiring managers across a number of industries, including pharmaceuticals, are using datacamp Recruit to meet their demand for data skills. We love seeing industries taking an interest in the ways data science can improve their processes, and it looks like the pharmaceutical industry is certainly on the right track. Our next story takes a look at what scientists are calling a reproducibility crisis in modern research in recent years, many research findings have offered spectacular results thanks to machine learning. Often, these findings are really exciting and newsworthy and grow our understanding of the world and the phenomena in it. However, Princeton professor Arvind Narayanan and his PhD student Sayash Kapoor have found that while machine learning is a useful tool, uninformed usage of it is widespread in the scientific community. For example, a recent line of research claimed to be able to predict when civil war would break out with 90% accuracy. On closer inspection, however, researchers in the team hadn't set up their testing and training sets correctly and ended up testing their model on the same data with which it was trained. When Narayanan and Kapoor surveyed areas of science where machine learning was applied, they found that other researchers had identified errors in 329 studies that relied on machine learning across a range of fields. The pair then went on to organize a workshop aimed at discussing and addressing some of these issues. Norman Malik, a data scientist at the Mayo Clinic, says the general lesson is that it's not appropriate to approach everything with machine learning. Despite the rhetoric, the hype, the successes and hopes, it's a limited approach. Have you ever spotted a game changing machine learning mistake? Let us know in the comments. Our final story this week takes a look at how AI was used to change a movie's rating From R to PG 13. When Lionsgate wanted to buy theatrical release rights for the indie action thriller movie Fall, the filmmakers faced a big ratings problem. The film is about two young women who are in danger of plunging from the top of a 2000 foot tall radio tower, and when it was filmed, the movie's rating hadn't been decided yet. As a result, the original version of the film is full of F bombs, a dialogue problem that would significantly limit the movie's box office performance. Reshooting the scenes wasn't within the budget, so to solve this problem, the film's director, Scott Manny, turned to the AI dubbing company Flawless. The team used AI to change more than 30F bombs in post production into PG13 acceptable epithets like freaking along with a few other lines of dialogue using the same take that powers deepfakes. Flawless normally helps out by changing mouth movements to better accommodate dubbing in different languages. That's all for this week's roundup. Thanks for watching. Don't forget to let us know what your favorite story of the week is in the comments and we'll see you same time, same same place next week.
