2025 Contest Days

🔲 Pixel

Choose the most informative pixels under a budget — probe what models actually look at.

CVCreativeModel probing★★★☆ open-ended thinking

📜The task, in plain English

Under a strict pixel budget, choose which pixels of each image to keep (masking the rest) so that a classifier can still do its job. Score depends on how well classification survives your masking.

InputImages + a pixel budget
OutputA pixel mask per image
Really testsCreative reasoning about where information lives in images and how to probe a model's attention

🔧The baseline you're given

Naive masks: random pixels, or a fixed central crop. Limitations: ignores image content entirely.

🚀Baseline vs. solution

⚠️ The baseline (what you are given)
  • Random / central pixels for every image
  • Content-blind
✅ The winning approach
  • Saliency-guided selection: use gradients (∂output/∂pixel) or occlusion tests to find pixels the classifier relies on; spend the budget there
  • Cheap classical proxy: edges/high-contrast regions (opencv) carry most information
  • Evaluate empirically: try 3 strategies on a validation slice, keep the winner — measurement beats theory in contests

🧒Explain it like I'm brand new

The intuition: not all pixels are equal — a blank sky pixel tells you nothing; the pixel on a cat's ear tells you a lot. "Saliency" is asking the model: if I nudge this pixel, how much does your answer change? Pixels with big answers matter. Tasks like this reward experimenters: nobody knows the best strategy in advance, so the winner is whoever tests the most ideas cheaply.

💬The Gemma 4 playthrough (2000-token limit)

Chat 1 · Gradient saliency
YOU
PyTorch classifier `model`, input image x (1,3,H,W), requires_grad. Give ONLY code computing pixel saliency = abs gradient of the top logit wrt x, summed over channels, then indices of the top-k pixels. Max 15 lines.
GEMMA 4
x = x.clone().requires_grad_(True)
logit = model(x).max()
logit.backward()
sal = x.grad.abs().sum(dim=1).squeeze(0)     # (H,W)
flat = sal.flatten()
topk = flat.topk(k).indices
ys, xs = topk // sal.shape[1], topk % sal.shape[1]

🎯Takeaways & what Day 1 might do with this

  • Pattern family: open-ended probing task — appears each year in some form; strategy beats recipe.
  • Budgeted-resource thinking again: pixels here, LLM calls in Concepts, questions in John Wilkins.
← 🚻 Restroom