Teaching Statement – Sean I Young, PhD

Page 1 Back to the top

Teaching Statement. I believe that the role of a good educator should be to help students find, in themselves,

the joy of learning. Carl Friedrich Gauss once said:

“It is not knowledge, but the act of learning,

not possession but the act of getting there,

which grants the greatest joy.”

Machines learn, too, but do not derive joy from learning. That only humans can derive joy from it suggests

that learning is what makes us ultimately human, rendering the role of educators more important than ever

in an age where people are inundated with knowledge. I believe that to find this joy, learning must be made

multi-scale—knowledge is first transferred at a coarse scale, where it is more digestible, and progressively

refined through a series of deltas until a clear picture emerges—much like a multi-scale transform of images.

I now discuss some of the techniques that I have used in my own teaching to achieve this:

Lie, and then reveal the lie. The multi-scale approach involves “lying” to students on purpose at first, and

then revealing the “lie” when the students are ready for deltas. A toy example of this is when I visualize the

logistic regression equation using the model of a perceptron and tell my students that the two models are

identical (a lie). I then tell students the choice of non-linearity is actually different for the perceptron model

(revealing the lie). This approach has two benefits. First, it is true that the two models are identical for the

most part, and the relationship between the two is clear from the offset. Students are not missing out on so

much even if they were not to consume the delta. Second, the delta is explicitly communicated such that

students are clear about the difference between the two models if they were to consume the deltas.

Cover intuition first and details second. Many textbooks tackle proofs of theorems by laying out a proof

sketch prior to the actual proof. This idea of covering intuition before going through the details with a fine-

toothed comb is similar to the “lying on purpose” one above, except the mechanism used for conditioning

is different. Here, the intuition acts as the coarse-scale knowledge and the details are the deltas. No lying

takes place. Again, the benefits are obvious: it is better for students to understand the whole proof sketch

and none of the actual proof, as opposed to, say, if students were to understand only the first few lines of the

actual proof.

Communicate at the level of students. There is no use in communicating finer and finer deltas in hopes

of clarifying the picture to a resolution that students cannot handle. A level of abstraction is always required

and the “right” level of abstraction depends on the level of the students. For students working on medical

image segmentation using deep learning, quantum-mechanical aspects of MRI can perhaps be abstracted

out and MR images simply presented as a grid of voxels. While this seems obvious, students have different

backgrounds and levels of understanding, posing difficulties on deciding the right level of abstraction for all

students. An approach I have used to deal with this is to “scaffold” the course content, so the main material

is digestible by the majority of students, while those wanting more are pointed at the extra materials.

TEACHING EXPERIENCE

Contrary to the adage that those who cannot do teach, I believe that regularly meditating on fundamental

results leads to clearer thinking and keeps our brains nimble. I also consider it an academic’s duty to distill

knowledge and pass it on to the next generation of researchers in hopes that they will carry on our legacy

further. Here, I briefly list my part-time teaching activities over the years.

Multi-dimensional Signal Processing, Tutor (2017). A 10-week 4th year undergraduate course on multi-

Page 2 Back to the top

dimensional signal processing taught while my advisor (David Taubman) was on sabbatical. Each tutorial

was 1.5 hours. Course content included continuous and discrete LSI systems, review of native programming

in C and memory organization and management + multi-dimensional filtering; imaging systems, aliasing,

resampling and introduction to multi-dimensional filter design; multi-dimensional filter design, correlation

and the Discrete Fourier Transforms in multiple resolutions; multi-resolution processing and transforms;

shape and morphological processing; segmentation and image feature analysis; color processing, conversion

and analysis; motion and optical flow; and introduction to media compression.

Machine Learning and Artificial Intelligence, Lecturer (2022–2023). An introductory 12-week machine

learning course, delivered online, for four trimesters to Australian students from diverse backgrounds. Each

lecture is 2.5 hours duration. Course content included supervised and unsupervised learning, reinforcement

learning, introductory computer vision using CNNs and natural language processing using BERT, model

interpretability and ethics. Mandate also included developing new course content, designing and grading

student assessments.