Learning as Denoising Diffusion Process - A Journey of Unlearning and Reconstruction
Learning is truly a denoising diffusion, an immersion process. Over the last few weeks, it has been a profound denoising—a deep self-retrospection and a humbling experience.
My six-plus years in industry gave me tremendous learning, but it was of a different school, a different magnitude and spectrum. After coming to Columbia, I’ve spent considerable time in solitude, learning and figuring things out with no external constraints. This is a completely different experience.
In industry, learning happened at breakneck speed as we built solutions, met commitments, and raced against timelines. There was little time to turn back and observe the true gaps in understanding. These last few weeks have marked a new milestone in my learning practice—what I call the denoising process. I’ve gained clarity on my real interests, my natural pace, potential capacity, and how effectively I spend time with concepts.
The Art of Patience
I’ve come to value that learning and developing certain ideas must take their own time, their own pace. I need to develop my mental model, reconstruct it from first principles—like reconstructing from the latent space in diffusion models.
Through this process, I’ve had to acknowledge many self-sabotaging beliefs and traits: an egoistic point of view that pushed me to bet on something unique at every step. This was draining me, consuming excessive time even on simple tasks. But this realization, this retrospection, has shown me a humble way to move forward.
I see this as genuine progression because I don’t want to be a prisoner of the past. This is a new normal where I must be patient, trust the process, and make consistent progress.
When Art Guides Logic
Through this time of reflection, I’ve discovered something crucial: my artistic brain should guide the logical brain to play a subtle game. To trust the process. To not feel alone or long for something from the past.
It will be a painful effort, but we must trust ourselves. Sometimes our priors are very difficult to update with new beliefs. We need a longer run, a warm-up period with a better learning strategy.
This is where I’ve landed now. This is also what differentiates my journey—the prior that I hold and the joint distribution that I want to learn or match is completely different and unique from the usual path.
If this diffusion process is given its due time and respect, it should yield flying colors.