Since the publication of an article in The Tech about allegations of plagiarism against HackMIT contestants on the Seamless team, HackMIT organizers and two members of Seamless have stated that the videos they presented as output of their own code in the hackathon’s final presentation were actually published by Microsoft Research. Both these two contestants and the organizers said that the misattribution was unintentional, while a third member of the Seamless team has sought to publicly distance himself from the project.
HackMIT organizers have since stated that they do not stand by a HackMIT blog post’s initial characterization of the incident as a “disqualification” of Seamless, who they said stepped down voluntarily.
Seamless, which had originally won second place at the hackathon, did not receive the $3,000 prize last month after HackMIT investigated accusations that the videos the members showed in their final presentation were not their own results. Another top eight finalist, AgileAssault, was also “disqualified” for submitting a similar hack the week before to HackUMBC, according to the blog post.
Seamless was composed of three MIT sophomores: Kevin Kwok ’17, Nikhil Buduma ’17, and Guillermo Webster ’17.
The team was seen referring to the looping videos in question as their own results during their final pitch. Webster introduced some of the videos, saying, “It’s a very difficult algorithm, and it was a big pain to implement… but here are some products of that.” Buduma refers to the process by which the videos were produced as “our algorithm” several times during the presentation.
In a written statement to The Tech, Webster said that in their final presentation, “failing to attribute the videos to Microsoft qualifies as plagiarism.” However, he said that the lack of proper attribution was unintentional.
Though Webster said they also tried to make an end-to-end implementation of the Microsoft Research algorithm capable of making short videos into seamless loops, both he and Kwok indicated to The Tech that they did not successfully generate such loops using their own code during the hackathon.
“Backstage, I voiced concern about correctly attributing everything, and we decided as a team that we would do so, but in this regard the presentation went very poorly,” Webster wrote.
“When we presented, we were just going to be like, ‘And this is what it might look like if your photos were moving,’ but I guess it morphed from an explanation into graph cuts to an explanation of the algorithm,” Webster said. “And then on stage, [it turned into] ‘this was our algorithm…’”
Webster said, “If we had really wanted to be cheaters, we could have just produced our own videos with the implementation that is freely downloadable from the website. Again, we weren’t trying to say that we made those.”
“I guess it’s probably a bad assumption for us to have made — an assumption that we made is that ‘isn’t it obvious that we didn’t make that?’” Webster added, referring to the looping videos.
There are two rounds of judging at HackMIT before the final round. Webster said that the pitches he made in the initial rounds of judging were very different from the final presentation and said he had made clear at that time that the videos were not produced by Seamless.
TechX associate director Ankush M. Gupta ’16 said he was among the judges who judged Seamless in one of the two rounds before final judging. “Their presentation from the initial round of judging and the final round of judging was actually quite different. In the initial round of judging, Kevin and Guillermo were the two people presenting it. They presented… here is a sample input, here is what the output would be. They cited the Microsoft paper,” and then demonstrated the web-based component of the project, he said.
During the final presentation, Buduma spoke the most, while Webster and Kwok contributed only minor explanations about their project.
“During our actual presentation, I was mostly oblivious,” Kwok wrote in a statement to The Tech. “We had mentioned the [Microsoft Research] paper, but failed to clearly articulate that their example data was simply used to showcase the video loop viewers.”
According to Kwok, during the final presentation, “There were phrases which seemed to be liable to misinterpretation, but I felt they were all within the bounds of true. It wasn’t until afterward, talking to Guillermo, that I began to realize the misleadingness of the sum of it all, and I began to feel uncomfortable with what we had done.”
Webster said that he realized that the audience overestimated the extent of Seamless’ contribution immediately after the presentation, when a judge implied that she thought the team had implemented a system to add motion to still images. “The first question we were asked by the judges made it clear that we were not being judged for what we accomplished,” he said. “[The first judge] clearly thought that not only were the videos ours, but that the algorithm was ours, and that it could work magic.”
However, in response to the first judge’s question, Buduma said, “Let’s say you were at Disneyland taking a picture with your family and people are moving past you or behind you. This algorithm would actually operate on that and remove the person in the first place. So here’s an example… if you look at the result of our algorithm, [the dancing person] gets removed,” while referring to a looping video displayed on-screen.
According to Gupta, the members of Seamless “did not explicitly come forward to HackMIT as an organization” after their presentation. “Personally, I felt bad enough that I had planned on donating the money to charity when I got it,” Webster said.
Kwok wrote, “To be honest, it felt like good closure when we agreed to step down from second place. Even though it was all an honest misunderstanding, the fact that our descriptions were ambiguous enough to leave that up to interpretation was unacceptable, and for that I am personally sorry.”
None of the members of Seamless commented for the initial article in The Tech about the matter. Webster said the team members initially decided to ignore interview requests in the hopes that “maybe if we ignore it, it’ll go away.” Many comments were left on the original Tech article, some proclaiming the presentation to be completely faultless while others accusing the members of intentional deception.
Seamless Member Roles
After the first Tech article was published, Buduma sent in a statement distancing himself from his teammates, saying that he had no involvement in the team’s attempt to implement Microsoft Research’s algorithm. He also posted similar statements on a public Facebook page and in a named comment on the original Tech article, adding that he felt the original article “framed [him] as complicit” and that he experienced “harassment” and “blackmail” when asked for comment by a reporter for The Tech.
In his statement, he wrote, “I had zero involvement in building the ‘magical’ back-end of project (which could be easily confirmed by the HackMIT team by taking a look at the Git repository and revision history) and… I was not involved in any misconduct that might have occurred.”
The team’s Git repository indicates that Buduma did commit back-end code as part of the team’s attempt to implement Microsoft’s algorithm, which seemed at odds with his statement that he only worked on a front-end interface for the videos.
“As can be verified in the Git logs, Nikhil did commit at least once to the MATLAB code, of which the C++ code is a direct port,” said Webster. “He wrote the function spatula, which is the spatial energy function,” he said, referring to an aspect of the video processing algorithm the team attempted to implement.
Kwok said, “For the first few hours of the hackathon, Guillermo and Nikhil worked on porting the energy functions specified in the paper into MATLAB, as we had found bindings for a graph cut optimization library.”
Buduma also stated, “[Webster and Kwok] just needed me to do a bit of front-end work and help them pitch.” In an interview with The Tech, however, Webster expressed some concerns with Buduma’s account.
“I personally didn’t ask for help with pitching the project at the beginning of HackMIT, and it was my impression that we had recruited Nikhil to help with the project in general, especially given that we had started working on the algorithm together,” said Webster. “I’m not sure what interactions Kevin and Nikhil had without my knowledge.” Kwok had pitched a HackMIT project himself the previous year with Webster, now called Project Naptha, and won second.
Buduma wrote, “If I had known that there was something fishy about their work, I would have never agreed to be a part of it.”
Webster said all of the team members knew that the videos were taken from Microsoft’s page. In the video of their final presentation, Buduma refers to the Microsoft videos as the results of “our algorithm” multiple times.
The HackMIT organizers said they did not know what information Buduma had or did not have about the team’s project during the competition and therefore couldn’t confirm or deny his claim that he knew less about the internal workings of the project than his teammates.
The Tech had previously reported that HackMIT organizers had conducted a hearing with members of Seamless. Siegel had previously stated that they had decided to step down without being “incriminated or found guilty of anything.”
When asked why Seamless didn’t demonstrate the functionality of their implementation in the hearing, Gupta said, “It is a multiple-hour-long process to generate the videos, which is part of the reason they explained for why they had to use Microsoft’s videos instead of their own.”
HackMIT co-director Katie Siegel ’16 said that during the hearing with herself and Gupta, Seamless team members demonstrated the functionality of parts of their code. “They did implement, and we saw that they implemented… all the parts of the algorithm that were necessary to generate the output,” she said.
But Siegel also said that she had not seen an end-to-end run of Seamless’s code, i.e. a non-looping video processed and turned into a looping video.
Webster said that he had never successfully run the code end-to-end until he modified the code from the hackathon recently. He said that this past weekend, “I changed some stuff, fixed some errors,” adding that with the modifications “it worked and it took 8 hours.”
Kwok told The Tech, “Toward the end of the hackathon, we managed to get C++ code to a state where a subset of the optimization could be run. We noticed that it was extremely slow, taking about a minute to run on a single candidate period, a task that needs to be run several hundred times before a single video loop can be generated.” He added, “On top of that, we weren’t even sure the data was formatted correctly.”
According to Gupta, the investigation concluded before the full functionality of that implementation could be determined. “Essentially our investigation concluded once we reached our mutual conclusion that the prize would not be awarded to them,” he said.
“We did not find them guilty through… verifying or not verifying that the code [worked],” Siegel said.
It was unclear from the organizers’ statements in the interview how the outcome of the hearing would have changed if Seamless had produced a fully functional implementation during the meeting. After their interview, the organizers emailed The Tech asserting that the state of the code’s functionality did not affect their decision to make Seamless ineligible for prizes.
Fellow HackMIT organizer Richard Z. Ni ’15, Gupta, and Siegel sought to draw a distinction between declaring the team “guilty” of “plagiarism” and allowing them to “step down” as a result of a “misrepresentation.” They further distanced themselves from a HackMIT blog post that said that they “disqualified” the hacks of two teams at the competition.
Ni said that Seamless not receiving the second place prize was part of a “mutual decision.” Siegel said that based on the hearing, “We decided that there was not enough evidence to find them guilty, and we talked about it a little bit more and they decided that instead of going on and digging further into the code… they were just going to step down because of the misrepresentation we discussed during the meeting that we had.”
While Ni said, “We technically didn’t disqualify them because they agreed to step down,” there is no such distinction in the HackMIT rules, nor any set procedure for dealing with alleged rule violations.
Of the blog post, Siegel said, “I don’t think at the time we thought too closely about the exact word ‘disqualification,’” adding that it was written by a member of the development operations team rather than those involved in the hearing.
Siegel suggested that “disqualification” was referenced in the blog post because Seamless was “lumped in” with AgileAssault, another project that was made ineligible for prizes by the organizers and which she did say was “disqualified” by HackMIT. But from Siegel’s description of the process used to remove AgileAssault from competition, the difference between disqualification and stepping down remained unclear.
“We did not investigate thoroughly,” she said, in reference to AgileAssault. She said she approached the project’s creator with the information that organizers of another hackathon said that he had recently presented a very similar project at a previous hackathon, while HackMIT only allows projects created entirely during the competition. Though the creator agreed, he defended himself by saying that it was a re-implementation of a previous concept with new code. According to Siegel, she said, “‘I still think that the two projects were too similar and that we can’t give you the prize,’ and he said, ‘Ok, I accept that.’”
The organizers also asserted that they did not find the Seamless team members “guilty” of any specific charge, citing a lack of evidence of malicious intent, although the HackMIT rules also do not reflect such a distinction.
Ni said that HackMIT had never described the incident as plagiarism, but after being read the MIT Academic Integrity Handbook definition of plagiarism, he said “I would think that it is plagiarism [by MIT’s] definition… If we didn’t think there was any plagiarism we wouldn’t even have talked to them [about stepping down].”
Siegel said, “The reason why we decided that what they showed us was not conclusive enough to find them guilty was because they didn’t [finish] the end-to-end [implementation]… to the point that they were able to present the change in the Microsoft videos.”
According to Ni, “‘Guilt’ implies a lot. It implies some sort of moral failure. And for us, we didn’t want to assign any sort of moral failure to their actions.” Gupta also said, “While we agree that they plagiarized, we have no evidence to attribute that to some sort of malicious intent or moral failure on their part. And that is why we… didn’t make a statement that would imply as such.”
Siegel said that only she and Gupta were involved in the investigation and that neither was friends with the members of Seamless. “In fact, all of the HackMIT team members who did know them took a big step away… when it came to reviewing this,” she said.
She also said that investigation was conducted internally because the HackMIT leadership was better positioned to make the decision than those outside the organization, who she said might be less familiar with the structure of the hackathon.
They said they had not made any decisions about whether members of Seamless would be allowed to enter HackMIT in future years.