In 2015, I chaired the last of our departmental Research in Progress seminars, in which PhD students from our department presented their research. In preparation for the seminar, I asked recent PhD graduates from our department to reflect on “…one thing that you wish you would have done during your PhD or postgrad training time, because you feel it would have helped you.”

 

Three of the major regrets were ‘not having attended more seminars, ‘not having done more to build connections’, and ‘not having focussed enough on developing soft skills, such as the ability to present well’.

 

Seven years earlier, in 2008, I started our Research in Progress seminars to address precisely these goals. However, as apparent from what students regretted after their PhD, most students did not prioritise these goals during their PhD. This is the main reason why after seven years, I eventually gave up organising our Research in Progress seminars.

 

Back in 2007, I joined the Department of Biochemistry at NUS as an Assistant Professor. When I started my position, my expectation was that I would focus on establishing my research group and developing a research programme. What I did not foresee is that I would spend a large chunk of my time over the next years to promote graduate student life in our department.

 

Soon after joining the Department of Biochemistry, I was put in charge of hosting our departmental Monday research seminars. At the same time, I noticed that there was no opportunity for our graduate students to present their work in front of the department. I had witnessed as an undergraduate and postdoc at Tufts University’s Biology department how the weekly graduate seminars were an important academic and social highlight of departmental life.

 

While at Tufts, I was impressed that the vast majority of professors made it point to show up at these seminars, ask questions and encourage students, even if their research interest was entirely different from that of the student presenters. As such, the seminars were a great opportunity for students showcase their work and obtain feedback, as well as to get to know students from other labs other over sandwiches after the seminar. The realisation that no similar structure existed in our department prompted me to start our Research in Progress seminars.

 

My appointment at the department coincided with the recruitment of Prof. Fu Xin-Yuan to take over the leadership of the department. One thing that really impressed me about Prof. Fu was his genuine passion for scientific research and fostering a departmental atmosphere of scientific exchange. As our Head of Department, he had the vision to create a culture where faculty, research staff and PhD students build connections and engage in scientific exchange. As such, he was highly supportive of our Research in Progress seminars as well as many other departmental activities.

 

Thus, with the strong backing and support from our Head of Department, I made it my mission to help promote a departmental culture of scientific exchange and a strong departmental community. Within nine months of my arrival, we launched our biweekly Research in Progress meetings, in which two PhD or Masters presented their research work. For the next seven years, the meetings became a fixture of our departmental life, allowing students to practice their presentation skills and interact over snacks after the meetings. The Research in Progress meetings likely also motivated students to make progress in their project in order to give a good presentation.

 

However, the major obstacle to the success of the meetings was student participation. There certainly was no lack of attempts and initiatives from my side to motivate students to attend. Thus, we routinely offered snacks after the seminars. We also let attending students vote on the quality of every presentation and awarded prizes for the best presenters at the end of each year. We even gave away “participation prizes” for students who attended the most seminars.

Ultimately, none of these measures really made a difference, because the two most important motivating factors for students were lacking: encouragement from the students’ supervisors to participate in the meetings, and participation by the supervisors themselves to lead by example. While I was at Tufts University’s Biology department, the majority of professors attended the graduate seminars, even though most talks were outside their area of expertise. This is clearly the best way to signal to students the importance of these seminars.

 

It is also important for supervisors to actively encourage students to participate. For example, as organiser of these meetings, I communicated to my students the benefits of attending and that I expected them to turn up. As a result, all my graduate and undergraduate students made it a priority to attend our Research in Progress seminars. Although this may not always have been convenient, my students told me that in the long run it was beneficial for them, because they were exposed to different areas of science from early on in their career and learned a lot about how to present well and how to use presentation tools effectively.

 

In contrast, most professors in our department likely did not actively encourage their students to participate. In fact, some professors even discouraged their students from attending or presenting, pointing out that it is more important for students to focus on their experiments. Moreover, only few professors showed up at the meetings, although there were some exceptions (most notably Assoc. Profs Deng Lih Wen and Chen Ee Sin and Asst. Prof. Takao Inoue!).

 

I continued an uphill struggle over the years to keep the meetings alive, until I eventually gave up. In hindsight, I do think about the success of the seminars differently. I now feel that I should have cared less about the number of attendees and more about the benefit for individual students who did attend. This mindset could have helped keep the meetings alive for longer.

 

That said, these days postgraduate student affairs are no longer handled by departments but by our School and the individual research programmes, and so we cannot bring back the past. But it is useful and enjoyable to sometimes remember it, which I will continue to do next week.

 

 

Especially in education, there is a lot of talk about AI tools that can detect whether students have used generative AI to complete tasks. However, thus far, I have not come across efforts to train humans at getting better at distinguishing between what is AI- and what is human-generated content. That is until this week, when I read an interesting new paper in the scientific journal PNAS.

 

In the study, Dawel and colleagues tried to train humans to identify AI-generated faces.

 

The authors pointed out that commonly recommended strategies focus on detecting errors or anomalies (such as mismatched earrings or extra or missing fingers). But as generative AI gets better, such errors and anomalies are becoming increasingly less common.

 

In contrast to this, the researchers took a different approach. They argued that humans have an intrinsic ability to distinguish AI-generated and human faces, but are unable to unlock this ability without specific training.

 

In particular, the authors hypothesise that humans are likely to spot one specific difference between AI-generated and human faces: “… generative algorithms are inherently biased toward the mathematical average of the tens of thousands of faces on which they are trained. Consequently, AI-generated faces are perceived as more typical in appearance than real human faces.” According to the authors, humans are likely capable of detecting this “more average” appearance of AI-generated faces.

 

The researchers broke the appearance of human faces down into six specific impressions:

 

Distinctiveness (How much would this face stand out in a crowd?)

Memorability (How memorable is this face?)

 

Proportionality (How proportional is this face?)

 

Symmetry (How symmetrical is this face?)

 

Attractiveness (How pleasing or pleasant-looking is this face?)

 

Expressiveness (How emotionally expressive is this face?)

 

Previous research has shown that AI-generated faces are typically rated as more symmetrical, proportional and attractive compared to real human faces. On the other hand, AI faces are rated as less expressive, memorable, and distinctive. (Indeed, the researchers confirmed these differences in their study during the training phase, as discussed below.)

 

Hence, the researchers hypothesised that by training humans to detect differences in these six impressions, human ability to tell AI and human faces apart would improve.

 

To test this hypothesis, the researchers used the following study design:

They first established a baseline by conducting Single faces AI-detection and Triplets AI-detection tasks. In the Single faces AI-detection test, participants judged whether individually presented faces were AI-generated or human using a six-point confidence scale. In the Triplets AI-detection task, participants were presented with three faces. One of the faces was AI-generated, which the subjects had to select.

 

The baseline ability of participants to identify AI-generated faces was around 50% in the Single faces AI-detection task, and around 40% in the Triplets AI-detection task, and as such, appeared little different from chance.

 

After determining the baseline, the researchers then conducted the training intervention. Here the participants were exposed to images of AI-generated or human faces. Each image was clearly labeled as either AI-generated or human. To “learn” the association of AI-generated and human faces with different facial impressions, the participants rated each face on the six different dimensions (facial distinctiveness, memorability, proportionality, symmetry, attractiveness, and expressiveness).

 

After the training, subjects were then re-tested using the Single faces AI-detection and Triplets AI-detection task with new faces, and the percentage of improvement was determined.

 

The results were remarkable. Training was highly effective and nearly doubled the average prediction accuracy from 41.4% pre-training to 81.1% post-training. The authors pointed out that some ‘high performers approached near-perfect accuracy’.

 

The improvements were similar for Asian and White faces, irrespective of whether the subjects themselves were Asian or White. This is interesting because typically Whites tend to find it much more difficult to distinguish Asian faces and vice versa.

 

In conclusion, the results of the study demonstrate that human AI-face detection is highly trainable, which makes me wonder whether similar approaches are possible for other AI-generated content.

 

On the other hand, sadly, I also learned this week that AI is also being utilised to undermine the ability of humans to detect genAI usage. Thus, a recent commentary in Nature describes a ‘Humanizer’ tool that can erase signs of AI-written text. The tool works by “personalising” the tone of research papers and grant applications written with an AI programme. Unsurprisingly, as pointed out by the author, the tool is alarming scientists.

So finally, based on the discussed criteria, can you tell which four of these faces are AI-generated? (The answer can be found at the bottom of the page.)

 

I am currently preparing for my first triathlon in 35 years, and my longest and most ambitious one ever. And so in this post, I am looking back at one week of training.

 

On Monday, I went cycling, trying to head for Lim Chu Kang. But seeing the dark sky ahead, I ended up doing five back and forth circles on Jalan Buroh along Bedok reservoir. This was not fun, but there are no major traffic lights and I still managed to do two hours of intense riding. Right after the cycling, I continued with a 13 km run to and around West Coast Park. This was my longest distance after cycling thus far and was really tough. What made it easier was listening to an amazing Rest is History podcast about Nelson Mandela and the South African National Anthem, which brought back memories of Clint Eastwood’s great movie Invictus.

 

Tuesday was swimming day. I usually swim during or after lunch at Clementi pool, even though it might be easier to use the NUS pool. But I prefer the anonymity of the public pool. Also, it feels nice to be away from work in a different environment during work time. The pool is very empty during the daytime and all this makes the swimming experience a little less painful. Surprisingly, I realised that I became faster compared to the previous week, which thus far I thought to be almost impossible.

 

On Wednesday at lunch time I went for a run at Rail Corridor, one of my favourites running spots, especially during the day on a weekday when there are very few people. Because it was very sunny, I cut my usual 15 km run short by 2 km. But clearly, of all the triathlon disciplines, running is the most fun and I actually look forward to it!

 

On Thursday I went swimming again. As always, I was dreading having my head mostly under water and feeling exhausted while counting down laps for 45 min. But I finished my 1600m of swimming, and afterwards, I felt happy that I did it.

 

On Friday afternoon at 5:30pm I headed off to cycle again, this time riding to Tuas. But this was a mistake because of the two and a half hours on my bike, I spent a large portion waiting at (unnecessary) traffic lights. Going forward, I will try to go to Lim Chu Kang or East Coast Park (if I am riding during the day when it is not as busy). After my bike ride, I followed up with another 13 km run, and although it was tough, my heart rate at the end of it felt lower than after previous cycling and running sessions, which is definitely a good sign.

 

Saturday was rest day, so I took a walk from NUS to Harbourfront, which was a lot of fun.

Thus far, it had been a good training week. And so on Sunday, my plan was to go for a long cycling session. However, I ended up not going. I simply could not motivate myself to go.

 

What made it so difficult?

 

It seems that I unconsciously knew that I did well this week and concluded that it was okay to skip this session and relax instead. Knowing this, I put a lot of things that I wanted to do into my morning schedule. And by the time I was finally ready to go out, it was already very late.

 

It goes to show that following through on plans when there is no urgency is difficult. In principle, it is, of course, okay not to follow one’s plans all of the time, because very few people do.

 

On the other hand, we usually plan the things we do because they make us feel good (or because we know that we feel good afterwards). In contrast, if I abandon my plans I often end up not relaxing or doing fun things, but to wasting time. This Sunday, I initially tried to take a walk, but it was cut short by rain. I then read and subsequently listened to music for a couple of hours, which was enjoyable. But eventually I wasted close to three hours searching for new music.

 

This is a strong argument for me to try to follow through on my plans. Looking back, the main problem was likely that I planned to do too many things in my office beforehand. By the time I finished them, I feel so tired that I wanted to relax. And then it was hard to find the motivation to go out and cycle.

 

In conclusion, I need to plan to do less!

 

Finally, one obvious question is why do I spend so much time and effort on training sessions that are tough and that I do not always enjoy in the first place?

 

Firstly, I feel that it is good for me to try new things. (Technically, I have tried triathlons before as a student, but that was so long ago that my experience now is likely going to be very different.)

 

Secondly, I know from experience that it is beneficial for my mental well-being to have goals. It provides a purpose as well as excitement.

 

Finally, challenging goals create meaning in my life in a bigger sense. When I look back at the two marathons I completed over the past two years, they feel like major achievements. Remembering them gives me great joy, which definitely offsets the difficult training. And I already feel excited about other marathon challenges in the future.