Jekyll2026-06-11T03:05:26+01:00https://xenxeon.github.io/feed.xmlTahir HasanComputer vision and machine learning researcher — MSc Computer Vision, Robotics and Machine Learning at the University of Surrey. Works on hierarchical medical image analysis, continuous sign language recognition, and explainable AI.Tahir HasanWhy a router network reduces inter-class confusion in GI imaging2026-06-08T00:00:00+01:002026-06-08T00:00:00+01:00https://xenxeon.github.io/blog/2026/router-networks-gi-confusionLook at the confusion matrix of any flat GI endoscopy classifier and the mistakes aren’t spread evenly — they pile up in small blocks of visually similar classes. The model almost never confuses an anatomical landmark with a polyp; it confuses kinds of similar findings with each other. That structure is an invitation: factor the prediction,

\[P(y \mid x) \;=\; \sum_{k} P(y \mid x, k)\, P(k \mid x),\]

where \(k\) indexes a coarse group. If \(P(k \mid x)\) is nearly free — and empirically it is — then all the hard work lives in \(P(y \mid x, k)\), and a specialist trained only within group \(k\) faces a dramatically easier decision boundary than one trained over all classes at once.

In code the inference path is almost embarrassingly simple:

group_logits = router(x)              # cheap, high-accuracy decision
k = group_logits.argmax(dim=-1)
fine_logits = specialists[k](x)       # capacity spent where confusion lives

Outline of the full post:

  1. The block structure of GI confusion matrices — with a real figure from my dissertation experiments.
  2. The mixture decomposition above, and what hard routing throws away (when \(\operatorname{argmax}\) is safe, and when it isn’t).
  3. Why specialists beat one big model at matched parameter count — an experiment, not a slogan.
  4. Failure mode: routing errors are unrecoverable; measuring the router’s ceiling on end-to-end accuracy.
  5. What I’d try next: top-2 soft routing, specialist abstention, calibration.
]]>Tahir HasanLessons from building a 9,010-video sign language dataset2026-05-18T00:00:00+01:002026-05-18T00:00:00+01:00https://xenxeon.github.io/blog/2026/lessons-9010-video-datasetPapers make dataset building sound like a procedure. It’s logistics: scheduling humans, fighting storage, and discovering that your labelling convention from week one fails on a sign you meet in week six. We collected 9,010 videos of continuous Bangla Sign Language across 530 word classes; the model that consumes them took a fraction of the effort the dataset did. This post is the honest accounting.

Outline of the full post:

  1. Picking 530 words — frequency lists vs. communicative coverage, and who should make that call (hint: signers, not engineers).
  2. Continuous vs. isolated recording — why we chose the harder option, and what it cost in annotation time.
  3. The annotation pipeline: tooling, label alignment, and the QC pass that caught the most errors.
  4. Storage, backup, and the boring infrastructure that saved the project twice.
  5. What I’d redo: things I’d standardise on day one, and the metadata I wish we’d captured.
]]>Tahir HasanFrom mechatronics to computer vision: what transfers2026-04-12T00:00:00+01:002026-04-12T00:00:00+01:00https://xenxeon.github.io/blog/2026/mechatronics-to-computer-visionI came to computer vision from a mechatronics degree — DH parameters and PID loops, not transformers. The retraining was real, but less of it was wasted than I expected: most of what an engineering education actually teaches is how to debug a system you don’t fully understand, and that skill transfers to machine learning embarrassingly well.

Outline of the full post:

  1. Kinematics is just geometry — and geometry is half of classical vision (camera models, homographies, my Modified-DH habit of writing frames down explicitly).
  2. Control loops → training loops: feedback, stability, and why “the loss went down” is the same trap as “the step response looks fine”.
  3. Hardware debugging discipline: change one thing, log everything, distrust the sensor first — i.e., distrust the dataloader first.
  4. What didn’t transfer: the statistics gap, and how I closed it.
  5. Advice for engineers crossing over (what I’d tell 2022-me).
]]>Tahir Hasan