Senior Spotlight: Ting (Justin) Jiang’s Path from Music to Machine Learning

Ting (Justin) Jiang is graduating with a double major in electrical and computer engineering (ECE) and computer science. Originally from Beijing, China, he transferred to Duke as a sophomore and later joined Yiran Chen‘s lab to work on generative models. This fall, he will begin his PhD at Carnegie Mellon University’s School of Computer Science, where he will continue working on generative AI, diffusion models, and tools that support human creativity. Read on to learn how music shaped his interest in engineering, the research questions driving his work, and where he wants to take generative models next.

How did you get to Duke?

I came in as a transfer student. I’d spent my first year at UNC-Chapel Hill, but I knew I wanted to study engineering, so I transferred to Duke. I was admitted as a math major but switched into computer engineering after arriving, and I’m really glad I did. ECE gave me a solid foundation in digital signal processing, which has shaped the way I think about machine learning systems today.

What interested you in ECE in the first place?

Music. One of my biggest hobbies since middle school has been music production, and that’s actually what pulled me toward ECE. Music technology is deeply connected to signal processing: convolution reverb and impulse response, additive synthesis and Fourier series, FM synthesis and the Fourier transform. Before coming to Duke, I built a Max/MSP patch that lets me browse synthesizer presets by walking through a 2D map of timbres and plugged it into my Ableton Live setup. Projects like that require real signal processing knowledge, and that’s how I got drawn into ECE.

What have been some of your favorite ECE classes?

ECE 110 with Professor Hisham Massoud was an important one. He’s known for being tough and grading strictly, but his standards really help cultivate better engineers. I’d also mention ECE 661 with Professors Yiran Chen and Helen Li. For anyone interested in machine learning, it’s one of the most useful courses you can take, covering everything from backpropagation to diffusion models.

How did you get involved in research?

I joined Professor Yiran Chen’s lab in my sophomore spring. I still remember our first meeting: I came in nervously with an early-stage idea called cache-assisted pruning, and he engaged with it seriously. He gave me freedom to pursue any direction I was excited about, with sharp feedback on what mattered and what didn’t. I would not be where I am today without him.

What I valued most was the room to find my own questions. I became fascinated by two problems the field tends to treat separately: why sampling from diffusion models is so slow, and why unconditional training has a fidelity gap over conditional training. The more I worked on them, the more they felt like two sides of the same foundational problem.

What has your research focused on?

My research so far has been on diffusion models, working on two questions that initially looked separate but increasingly feel like two sides of the same problem.

The first is sampling: why generating an image with these models is so slow. Different samples have very different stability properties, but most acceleration methods apply the same shortcut to all of them. The first paper I led, with two close collaborators, reframes this as a stability-prediction problem: at each step, the sampler decides whether it’s safe to skip computation, and corrects the error when it does. It was accepted at ICML 2025.

The second question is more foundational—how understanding and generation relate inside a model. The way I think about it: understanding compresses a high-dimensional image into a compact representation, which is many-to-one, while generation has to invert that and expand a low-dimensional embedding into a distribution of possible images, which is one-to-many. So they’re really two directions of the same kind of transport, and there’s a lot of room for them to inform each other once you stop treating them as separate tasks. The paper I co-led pushes on exactly that.

What’s next for you after graduation?

After graduation, I’m heading to Carnegie Mellon for a PhD in computer science. The way I see it, generative AI has spent the last few years answering one question: can a model produce a photorealistic image, a coherent piece of music, a navigable 3D scene? Diffusion models have largely answered yes. The question I want to spend my PhD on is the next one: can a model do that inside a creative loop, responding to a human collaborator in real time? A diffusion model powerful enough to simulate worlds still can’t keep up with a live saxophone solo.

Two things have to happen. The first is on the audio side: music, speech, and environmental sound are still mostly handled by separate, task-specific encoders, and the unified encoders that try to cover all three lag behind. I want to close that gap. The second is on the understanding-and-generation side. When I make music, I don’t just generate: I play something, listen back, change my mind, play again. Creativity is iterative, and the two directions need to inform each other constantly.

What I want to build isn’t a model that hands you a finished piece, but a differentiable layer underneath the tools musicians already use: something that lets traditional DSP reach places it was never able to express.