Talha Nazar

#Master's thesis: infographic understanding

Resumes increasingly use infographics, bar meters, circular progress rings, icon scales, to communicate skill levels instead of prose. My MSc thesis built a pipeline to automatically find those infographics inside a document, tie each one to its associated skill label, and convert its visual fill level into a number, treating the resume as a visual scene rather than a block of text.

1. Detect. A YOLO model localizes each infographic in the document and returns a bounding box.
2. Associate. Padding is added to the top and left of each box to capture the nearby skill name, which OCR then reads.
3. Clean & threshold. A 3×3 Gaussian filter smooths the crop; Otsu and adaptive thresholding then separate filled pixels from empty ones.
4. Quantify. Contours isolate the filled region, and the skill level is scored as filled area ÷ total infographic area.

Fig. 1: detection, skill association, and quantification pipeline (YOLO + OCR + contour analysis).

The thesis conclusion proposed this as a baseline for a more brain inspired successor, incorporating models of human visual attention such as saliency mapping and eye tracking style priors, so infographic interpretation looks less like template matching and more like how a person actually scans a document. That's the thread that pulled me toward NeuroAI.

Published at IMCOM 2023 · under review at IJDAR, full citations under Publications.

#Where the work happened

#Representational drift: Hebbian vs. backprop

Both models are trained on the same binary task until performance stabilizes, confirmed stable across a three epoch window, then tracked for five more epochs of continued training. No new task, no injected noise.

Fig. 2: core MLP validation accuracy, 3 seeds. Both rules converge before the post stability tracking window begins.

Drift is quantified four ways, RSA, CKA, centroid movement, and class axis geometry, plus a decoder trained on held out images at the stable epoch and retested on separate images from later epochs, to confirm class information survives the drift.

Fig. 3: RSA drift, CKA similarity, centroid shift, class axis angle, cross epoch decoding, and validation accuracy · backprop vs. hybrid · averaged across 3 seeds.

Both learning rules preserve task information throughout. But backprop shows a small, consistently greater amount of post stability drift than the hybrid rule, across all three metrics and all three seeds.

CNN and KAN runs used the same two learning rules but weren't repeated across seeds or controlled as tightly as the MLP, so they're exploratory checks, not part of the main conclusion. Across all three architectures, backprop and Hebbian RSA similarity degrades over training, most sharply in the MLP and more gradually in the CNN and KAN.

Fig. 4: backprop and Hebbian RSA similarity over training, MLP vs. CNN vs. KAN (exploratory).

Fig. 5: CNN representational dissimilarity matrices, backprop vs. Hebbian, conv2 layer (exploratory).

#Certifications & languages