Most of what I actually work on is deep learning, but the part that keeps my attention isn't the output, it's what's happening underneath. How a network ends up representing whatever you feed it, and whether that representation survives being dragged somewhere it wasn't trained for. A lot of my projects start from adapters and small architectural changes, LoRA variants, routing between experts, frozen backbones with a bit of task memory bolted on, because that's usually where you can actually see what a representation is carrying. The writeups almost always start as me trying to figure out why something worked better or worse than expected, and they rarely stay as short as I planned.
That same question is what keeps pulling me into generative work, specifically the cross-modal kind. I've spent a while getting one signal to talk to a system built for something else entirely. EEG feeding into a frozen diffusion model's latent space is the one that actually worked, and I want to push the same idea toward audio, music, and video, not just static images. Lately I've also been reading into JEPA and world models, mostly because they skip reconstruction entirely. The model doesn't need to rebuild the input pixel by pixel, it just has to predict something useful about it. I suspect that's what I was already fumbling toward with the diffusion experiments, I just didn't have the vocabulary for it yet.
Neuroscience is where most of this ends up, EEG especially. Biological systems handle perception, plasticity, and memory in ways deep learning keeps rediscovering under a different name, and I like poking around in that overlap even when half the experiments don't go anywhere. I care more about the question than about landing a clean result, so these projects tend to grow into long ablations instead of tidy benchmarks. It's also the one place where I actually find out whether the representation learning ideas hold up outside a toy setup, or just fall apart.
Sepuluh Nopember Institute of Technology
Relevant coursework: Deep Learning, Machine Learning, Computer Vision, Data Mining, Natural Language Processing, Digital Image Processing, Computational Intelligence, Calculus I & II, Linear Algebra, Probability & Statistics, Discrete Math, Computer Graphics, Object Oriented Programming.
That curiosity doesn't stay contained to a notebook for long. I end up organizing, mentoring, or just lurking in whatever community will argue about model internals with me, which is where the list below comes from.
Schematics ITS
Avalon AI
Google Developers Group on Campus - Institut Teknologi Sepuluh Nopember
Bayucaraka UAV Research Team
Every so often that same curiosity gets pointed at a deadline instead of an open question. Competitions are where that happens, and what follows is less a trophy case than a record of the times the pressure actually worked in my favor.
Preliminary Round Datavidia 10
AXION by Avalon AI Community
ISE! Academy Python Programming for Data Science Intermediate Level
The tools followed whatever a project needed at the time, not the other way around. Grouped roughly by where each one tends to show up.
All of it points back to the same habit: pulling apart deep learning architectures to see why they work, sometimes with a neuroscience lens held up next to them. The rest of this page is just where that habit left evidence.