Medal-winning scholar takes machine learning to new heights

Edmund Hofflin learned of his University Medal while hiking the Larapinta Trail.

Students

Medal-winning scholar takes machine learning to new heights
Medal-winning scholar takes machine learning to new heights

As he reached one of the highest lookouts on a ridgeline stretch of the Larapinta Trail, Edmund Hofflin paused to take in the 360° panorama.

“Oh, I have Internet!” said a fellow hiker, who had powered up his phone. Hofflin did the same, and, after taking some photos, checked his email.

Standing on a desert peak, 1,389 metres above sea level and more than 2,000 km from the lakeside campus where he’d recently completed his undergraduate studies, Hofflin received some unexpected news: he had been awarded the University Medal for academic excellence.

“I was absolutely delighted”, he said. “Shouted my lungs off.”

Hofflin had agreed to join the 200 km hiking expedition before realising he’d be missing his graduation ceremony. Each time he and his friends reached a summit, news from The Australian National University was beamed into his palm.

Edmund Hofflin on the Larapinta Trail

At a previous perch, he had received his marks for his final semester courses and his Honours research project.

“I was like, ‘Oh, yes, I’ve graduated, thank God.’”

Now, the Mount Giles lookout had supplied the altitude necessary for word of his University Medal.

And as for the highest peak on the Larapinta Trail? “We were on Mount Sonder while the graduation ceremony was on”, he said.

Shouting of a University Medal from a mountaintop heritage site had not been part of Hofflin’s plan, but making plans and then discovering better ones has been a central part of his 5 ½ year journey at ANU.

A slightly steeper curve

As a child, Hofflin was infamous for repeatedly trying to crawl under parked cars to see how they worked. Born with a hunger for learning, his performance at school was hindered by limited reading speed and speech impediments.

He would stutter, and he couldn’t pronounce ‘r’ or ‘th’ sounds along with a smattering of other syllables from kindergarten through Year 3.

A resulting shyness and reluctance to speak in public compounded the issue. He attended speech therapy classes and, by all appearances, has now put his speech issues behind him.

“I still stutter if I’m not careful”, he said. “And I try to carefully choose my words to make sure I am not misspeaking or misrepresenting anything.”

When Hofflin arrived at ANU, he brought with him a strong preference for meticulous planning, both on a day-to-day and year-to-year scale, when it came to his studies. But in his third semester, a sudden illness forced him to let go of that preference.

As he began to prepare for mid-semester exams, he fell ill. And when he found he couldn’t focus his mind enough to properly study, he went to a clinic on campus.

“One of the doctors looked at me and said, ‘Oh, I’ll take some bloods’. Came back the next day, and said, ‘Please go to hospital.’”

Hofflin was diagnosed with a rare condition similar to lupus. His immune system was attacking his tissues and organs. He missed his exams and was forced to drop 3 of his 4 courses. Friends came to visit him at the hospital and offered moral support, but at times during his six-month battle to regain his health, he questioned his decision to leave his home in Sydney for university.

“I was still sort of getting used to being an adult and being by myself,” he said. “I felt it would have been a bit easier to have done this with my family in Sydney.”

The academic path he had charted for himself was completely upended. He would need to wait for the courses he had dropped to come around again, but, before long he came to appreciate the fluidity made available by the School of Computing and his flexible double degree.

“I always discovered new things when I did a course and I was like, ‘Oh, that’s fascinating. Let’s do that.’”

He didn’t complete one of his second-year courses until fifth year because he couldn’t resist opportunities to take unique courses that were one-offs.

“It kept getting delayed,” he said. “And then finally I realised if I don’t do it now, I won’t graduate.”

These days, Hofflin’s academic plans are far less rigid in order to allow for any further medical problems or changes in his academic interests.

His newfound comfort with open-ended exploration has dovetailed nicely with the School of Computing’s teaching philosophy. And, when it came time to write the Honours research project that would lead to his University Medal, intellectual adaptability would be an integral part of his methodology.

Steady rise, stochastic subgradient descent

Hofflin’s course at ANU fascinated him from the start.

“Even from first year, first semester, I remember taking the introductory level course (COMP1100/1130) and the final question on the final assignment was ‘Do whatever you want, identify a problem, solve it however you want or do anything related to it.’”

Hofflin opted to create a computerised new Sudoku solver. He programmed it to randomly guess solutions, get marked on its performance, and then try to improve its next score by. learning from its mistakes.

“Theoretically my solver should have worked, however it was so terribly slow that it never did solve a single Sudoku,” he said. “However, my enjoyment in making this solver did lead me to take further interest in AI, machine learning, and, eventually, optimisation.”

Two years later, in the Advanced AI course (COMP4680), Hofflin realised that his attempt at machine learning had independently arrived at what he called a “poor version” of the Monte-Carlo machine learning methods.

Armed with new knowledge, he was able to fix the Sudoku solver he’d attempted in his first year.

Professor Stephen Gould was Hofflin’s advisor for his Honours research project, which focused on the editing of code to allow programs to run more efficiently and effectively, a process known as optimisation.

His thesis was a theoretical and experimental analysis of an algorithm called stochastic subgradient descent, which had been designed to level out non-smooth functions in neural networks caused during the optimisation process.

“What set his thesis apart was not just the deeply technical material that he needed to research but also Edmund’s ability to balance insights with mathematical rigor so as to make the thesis accessible to a more general computer science audience”, Gould said, adding that Hofflin’s work would have applications in the training of “large-scale machine learning models and, in particular, deep learning”.

“One reason I really like optimisation is that it has applications pretty much everywhere,” Hofflin said. “From chemical models to genome decoding, from a production system for manufacturing cars to just how your phone runs messenger.”

Asked how he measured success, Hofflin cited advice he’d received from his parents to “do what you love, and do it to the best of your ability”.

At each turn, he chose the path that most interested him. This meant that he spent an unusually long time delving into previously published research to scope out what had already been achieved in the optimisation field.

As his friends and classmates began writing their theses, Hofflin was still deep in the literature review process.

“Rather than having a very clear idea at the start, and sort of forcing my way through that, I allowed it to change, which did mean that later on I had less time to write my thesis, but it meant that I had less time to write something that I was really excited about.”

Hofflin wrote his thesis in six weeks, drawing upon detailed notes from his extended research, as well the courses he had taken throughout the previous five years. He didn’t do much editing or proofreading until the first draft was complete.

“I remember reading through my thesis and just being excited,” he said. “After spending an entire year researching, and the last six weeks quickly trying to write up a thesis, I was still so excited just to turn the next page and see what was happening. I think that was the moment for me when I was really proud of the work I’d done, because I’d set out to do something that I was really interested in, and as I read through it, I thought I’d achieved that.”

Child-rearing tip for parents of future software engineers

Hofflin said that his “do what you love” methodology was not the only life lesson for which he thanks his parents.

The second of four children, and the first of three boys, Hofflin said that very early in his childhood, his parents, Rebekah Jenkin and Philipp Hofflin, introduced a nightly tradition as they tucked their kids into bed.

“We could ask one question and they would answer it,” Hofflin said. “We really enjoyed trying to trick our parents, trying to find out what they didn’t know—mostly failing because it turned out that they had Google.”

The questions began with things like “How do you make a chair?” and “How does a volcano erupt?” and then increased in complexity over the years.

“By the time we were teenagers, my parents were having to read books on quantum mechanics to keep up with our inquiries”, Hofflin said.

This month, Hofflin has been spending time with his parents and siblings before saying goodbye as he embarks on his next adventure.

He will soon depart for the United States for a four-week intensive on applying computer science to ethics and safety issues hosted by The University of California at Berkeley.

From there, Hofflin will travel to Lausanne, Switzerland to begin a Masters in Mathematics Optimisation at the Swiss Federal Institute of Technology in Lausanne (EPFL).

arrow-left bars search times