Checklist MAIO 2025 Training and Selection Programme, Selection Test 2 · Task 1
Dieting Network
Make a fixed network whose weights are all 1 and biases 0 perform regression well by designing only its activation functions.
The task
A small fully connected network (8 -> 5 -> 5 -> 5 -> 1) has all its weights frozen to 1 and no biases. The contestant may only replace the three activation functions (act1, act2, act3) and may implement any activation function they like, subject to the restrictions below.
The statement frames the problem as a network that wants to 'lose weight'. Training, validation and test tensors are provided; the contestant saves predictions on X_test for scoring in a separate notebook.
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
- PyTorch tensors X_train,
y_train, X_val,y_valand X_test (downloaded from the organisers' Google Cloud Storage bucket by the notebook); the baseline DietNetwork with FixedLinear layers. - You submit
y_test_pred.pt(predictions on X_test) and the notebook with the reasoning.- Scoring
- 1 pt for explaining the approach; 1 pt for R2 >= 0.25 on X_test (sklearn.metrics.
r2_score) with valid activations; plus 0-3 pts = (your R2 - baseline) / (benchmark - baseline) x 3, where the benchmark is the best R2 among participants and the baseline is 0.25 (or the lowest participant R2 if that exceeds 0.25). - Rules
- Each activation function may contain at most 5 parameters.
- Activation functions must be stateless at inference (same input gives the same output).
- Weights stay fixed at 1 and biases at 0.
- Format
- Selection Test 2 of the Malaysian IOAI Training and Selection Programme 2025: five problems in two sittings (8 and 14 June 2025), about 2 hours per problem, used to pick the IOAI 2025 team. The problem was designed to be solved within 2 hours; all LLM assistance, including reasoning models, was allowed. Originally graded with an autograder via API submissions to the organisers' platform plus hand grading; slightly modified for distribution.