Checklist AICC 2026 Round 4 · Task 3
Extreme Condensation
Design a single synthetic 28 × 28 image with a soft label such that a fixed CNN trained from scratch on it alone classifies MNIST digits as accurately as possible.
The task
Devices containing an identical lightweight CNN digit classifier can receive only one training example before retraining. The contestant designs that example: one 28 × 28 greyscale image and one 10-dimensional soft label vector.
A fixed CNN is retrained from scratch using only the submitted example and evaluated on real handwritten digits. The MNIST training set (60k images) is provided for analysis only; the submission contains no real images, and the hidden test set is drawn from the same distribution.
The image has 784 pixel values, floats in [0, 1]; the label vector is non-negative and sums to 1.0.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
mnist_images/ (60,000 PNG digits) andtrain_labels.csv(filepath, label).- You submit
- CSV with exactly one data row: id, p0–p783 (flattened pixel values in [0, 1]) and l0–l9 (soft label probabilities summing to about 1.0).
- Scoring
- Classification accuracy of the retrained model on 1,000 hidden test images. Baseline 0.05, reference solution about 0.2.
- Rules
- Individual participation (maximum team size 1); at most 30 submissions per day.
- AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
- Format
- AICC Round 4, online on Kaggle, 20 Feb 2026 17:58 UTC – 22 Feb 2026 18:00 UTC.