Erica Chenoweth and Zoe Marks Named Pfoho Faculty Deans
Harvard SEAS Faculty Reflect on Outgoing Dean, Say Successor Should Be Top Scholar
South Korean President Yoon Talks Nuclear Threats From North Korea at Harvard IOP Forum
Harvard University Police Advisory Board Appoints Undergrad Rep After Yearlong Vacancy
After Meeting with Harvard Admin on ‘Swatting’ Attack, Black Student Leaders Say Demands Remain Unanswered
Jon L. Noronha ’11 and Eric N. Hysen ’11 have invented a tool to track the calories, fat, protein, and carbohydrates in food by taking pictures. “PlateMate,” as the tool is called, allows users to take pictures of their meals and send them to other people who estimate calorie counts and nutritional information with great accuracy.
The people making the estimates are part of Amazon Mechanical Turk—a crowdsourcing community—and are referred to as Turkers. They are online workers who are paid a few cents for each short task they complete. In this case, Turkers examine PlateMate photographs and provide calorie estimates.
When testing PlateMate, Noronha and Hysen found that the results “overestimated calories by about 7 percent, compared to 5 percent to 9 percent for the best experts.”
According to Noronha, when PlateMate was used by students at Harvard, “a majority preferred our system and found it more accurate and easier to use than doing their own estimates.” The PlateMate results were also much more accurate than those of the students’ personal food diaries.
Noronha said there are two reasons why the system is appealing.
“First, it’s way easier because all you have to do is take a picture, without painstakingly recording or memorizing what you ate and looking up nutrition facts” he said. “Second, it’s less biased. Prior studies show that people seriously underestimate their own eating, remembering their portions as smaller and healthier than they really were. Having a third party do the estimate removes that bias.”
According to Noronha, if the questions are structured properly, results can be received within minutes.
Hysen and Noronha noted that the development process has been difficult. Early on in the testing, the pair had trouble convincing Turkers to complete PlateMate tasks truthfully. At first, they asked Turkers to both identify the types of food in the photographs and then determine the nutritional value of the meal by searching a database.
“We found that the results from that were very bad because we asked workers to do two things at once,” Hysen said. “We replaced the single task by a two-step process where one set of workers identifies what’s in the box, and another set matches those items to the food database, and got much better results.”
PlateMate arose as a final project for a computer science course last fall, according to Hysen. Currently neither Hysen nor Noronha are working on the project, but Hysen said they “hope [their] research inspires someone to turn this into a publicly-available product. Making a smartphone app would be one of the first steps if that happens.”
Want to keep up with breaking news? Subscribe to our email newsletter.