Judging AI Competitions: How Mathematics for AI/ML Unlocks AI Learning's "Vital Channels"?

Recently I had the honor of serving as an academic judge for IAI2O and a few other youth and university AI innovation competitions. After doing deep reviews of dozens of papers spanning macroeconomic policy control, spatial transcriptomics, medical time-series prediction, and multispectral pathology diagnosis, I came away with some strong impressions.

Many students preparing for AI competitions who've bought math4aiml (Mathematics for AI/ML) often ask me: "The formulas in the book look heavy and deep — do competitions actually need them? Can't I just call a package (scikit-learn, PyTorch) and get results just as well?"

In this diary entry, combining real cases from this round of judging, I want to dig into why every top-ranked paper, without exception, has deep grounding in mathematical principles, and lay out an alignment guide between "math topics and academic competitions" for readers of the book, pointing toward where to focus next.

Judging the papers

The Core Pain Point: Why "Pure Package-Calling" Doesn't Score High With Top Judges

In this competition's official rubric, "AI Core Application & Research Contribution" carries a full 20% of the weight.

Looking at high-scoring papers such as MAKOTO, SpaCoEx, and DermaSpectra, one iron rule emerges:

  • Low-scoring projects ("black-box package-calling"): usually take an off-the-shelf model as-is (XGBoost for a classification task, an LSTM for a forecast) with no mathematical modification, and often can't even give a mathematical justification for their hyperparameter choices. These projects typically score only 5-7 on the AI core dimension.
  • High-scoring projects ("mathematical innovation / deep customization"): define a custom optimization objective for the specific scientific problem, derive their own solving algorithm, or design an innovative probability measure. They understand the underlying model logic deeply enough to describe the system's boundaries and limitations in mathematical language. These projects typically score 9-10, a near-perfect mark, on the AI core dimension.

In short: mathematical principle is the decisive force separating mediocrity from excellence.

With Dr. Haribabu Arthanari, Associate Professor at Harvard University

Aligning Math Topics: Which math4aiml Chapters Keep Showing Up in Top Papers?

Let's line up the paper methods that kept recurring in this round of judging against the chapters of math4aiml.

1. Linear Algebra and Spatial Geometry (Ch. 1, 2, 7, 8): From Embedding Vectors to Manifold Geometry

In an era where high-dimensional feature representation and representation learning are everywhere, linear algebra is not just a computational tool — it's the language of spatial modeling.

  • Case 1: SpaCoEx, sparse gene selection in spatial transcriptomics
    • Method: rather than treating the spatial expression matrix as an ordinary table, the study rigorously modeled the local gene co-expression matrix as a symmetric positive-definite (SPD) matrix on a Riemannian manifold. Using a log-Euclidean representation, it converted distance measurement on the manifold into a Frobenius norm in tangent space.
    • Maps to math4aiml: Ch. 1 (vector spaces and inner products), Ch. 7 (symmetric matrices), and Ch. 8 (SVD). Understanding the geometric properties of matrices and how norms are defined is the foundation for this kind of high-level, cross-disciplinary geometric modeling.
  • Case 2: Semantic Fracture, measuring media bias
    • Method: represents text semantics via sentence embeddings (the contrastive model SimCSE), then applies ZCA whitening to address anisotropy in high-dimensional embedding vectors, stretching the feature space so semantic distance is computed more accurately.
    • Maps to math4aiml: Ch. 2 (matrix fundamentals)'s treatment of covariance matrices and matrix decomposition — whitening is, at its core, decorrelating and rescaling a covariance matrix.

2. Matrix Calculus and Numerical Optimization (Ch. 11, 12): The Soul of Solving and Searching

Every model's training and algorithmic optimization is, at bottom, a search for an optimum in the hyperspace defined by a loss function.

  • Case 1: SpaCoEx
    • Method: to perform row-and-column-wise sparse selection (a group-lasso L1 constraint), the authors fully derived and implemented a projected proximal gradient ascent algorithm in the appendix, including the gradient step, one-sided soft-thresholding operator, and box projection.
    • Maps to math4aiml: Ch. 11 (matrix calculus) and Ch. 12 (optimization methods). Proximal gradient methods are a classical advanced tool in non-smooth convex optimization; only by mastering Ch. 12's fundamental theorems on convex optimization, regularization, and constrained optimization can one understand and independently derive a solver at this level.
  • Case 2: MPC-GWO, grey wolf optimization for feature selection
    • Method: proposes a multi-population cooperative grey wolf optimizer (MPC-GWO). Parallel populations, elite migration, and dynamic leadership mechanisms address the tendency of a single population to fall into local optima, applied to high-dimensional ASD gene feature selection.
    • Maps to math4aiml: Ch. 12 (optimization methods). The book discusses the optimization landscape, the tradeoff between local and global optima, and the balance between exploration and exploitation in detail — the theoretical bedrock for understanding and designing metaheuristic optimization algorithms.

3. Probability, Statistics, and Model Calibration (Ch. 13): Measuring Uncertainty for Real-World Decisions

In AI4Science (medical diagnosis, weather forecasting, macroeconomic control), a model's "probability" output can't just be a number — it needs real physical/statistical meaning.

  • Case 1: MAKOTO, macroeconomic reinforcement learning
    • Method: to address the preference for reward stability under a non-stationary economic environment, the paper cleverly introduces an unsupervised IsolationForest anomaly/uncertainty score, embedding it into PPO as a multiplicative uncertainty-aware reward.
    • Maps to math4aiml: Ch. 13 (probability and random variables). How to quantitatively describe a system's uncertainty and fold it into an expectation is central to how an RL agent makes robust decisions in a complex environment.
  • Case 2: NBA Shot Quality shot-quality evaluation and SCRIBE continuous sepsis prediction
    • Method: both papers emphasize probability calibration. Raw sigmoid outputs from a classifier are often distorted; both papers plot reliability diagrams, apply temperature scaling / Platt scaling, and compute a Brier score, so the model's output probability genuinely reflects the real-world event probability.
    • Maps to math4aiml: Ch. 13 (probability and random variables)'s treatment of the CDF, conditional probability, and empirical probability distributions.

4. The Bellman Equation and Reinforcement Learning (Ch. 15): From Single-Step Decisions to Long-Term Planning

  • Case: MAKOTO
    • Method: studies macro fiscal policy optimization under multi-agent cooperative/competitive games, using a centralized-training decentralized-execution (CTDE) framework, evolving over long horizons in a macro simulator.
    • Maps to math4aiml: Ch. 15 (Bellman dynamic programming). Understanding Markov decision processes, state-transition probabilities, value-function iteration, and the Bellman optimality equation is a prerequisite for macro-policy RL modeling.

5. Dynamical Limits and Continuous Evolution (Ch. 19, 20, 21): The Cutting Edge

  • In time-series forecasting papers (sepsis prediction SCRIBE, or Colorado River flow prediction River AI), most competitors currently use discrete LSTMs or temporal convolutional networks (TCNs).
  • But at the academic frontier, neural ODEs and SDEs are becoming the standard tool for modeling physical and biological dynamical systems. This corresponds to math4aiml's Ch. 19 (stochastic differential equations), Ch. 20 (ODE/SDE and continuous limits of algorithms), and Ch. 21 (the Fokker–Planck equation and distribution dynamics). Anyone who applies this theory to model continuous-time systems in a future competition will be unmatched on the "originality" axis.
The score tallies left from this round of judging

The Gold-Standard Prep Guide: A Learning Roadmap for Readers of the Book

To turn math4aiml's theory into a competition weapon, here's a suggested staged learning path:

Stage One: Solidify the Math Foundation (Linear Algebra and Matrices, Ch. 1-10)

  • Focus on: Ch. 1 (inner products and similarity), Ch. 7 (eigenvalues of symmetric matrices and orthogonal projection), Ch. 8 (SVD for dimensionality reduction and denoising).
  • How to apply: this is the mathematical foundation for understanding today's mainstream multimodal alignment (e.g., cosine similarity on CLIP embeddings) and text-representation similarity algorithms.

Stage Two: Understand How Solvers Work (Calculus and Numerical Optimization, Ch. 11-12)

  • Focus on: Ch. 11 (the chain rule for gradients, Jacobians, and Hessians), Ch. 12 (gradient descent variants, constrained convex optimization, Lagrange multipliers, proximal gradient methods).
  • How to apply: when you need to modify a standard loss function, design a custom sparsity penalty, or design a custom weight-balancing strategy for multi-task learning, these tools let you derive the formulas confidently, instead of flailing at black-box testing.

Stage Three: Measure Uncertainty (Probability and Statistics, Ch. 13-14)

  • Focus on: Ch. 13 (distributions, covariance, and expectation of random variables), Ch. 14 (information theory fundamentals, KL divergence, and the nature of cross-entropy).
  • How to apply: when writing your paper, never report just one accuracy number. Use bootstrap to estimate confidence intervals, use a permutation test to verify significance (p-values), and rigorously calibrate the model's output probabilities. These are core measures of whether "research design" and "data and evidence" are solid in the rubric.

Stage Four: Discipline Your Research Design (Avoiding Rubric Pitfalls)

From a judge's point of view, students should pay special attention to the following when writing their papers:
1. Reject data leakage: when doing time-series or medical patient survival prediction, never use a random split. For time series, split chronologically (as in River AI); for medical patients, split by patient ID (as in SCRIBE, to avoid the same person's data appearing in both training and test sets).
2. Always run an ablation study: if you propose a multi-module system, you need a table showing performance degrading step by step as modules A, B, C are removed, to prove that none of your mathematical design choices are redundant (e.g., River AI's ablation on the regime feature, and MAKOTO's ablation on the uncertainty weight).
3. Don't over-claim: avoid words like "fully objective" or "perfect generalization." Judges place enormous weight on how deeply an author understands their own model's "limitations" (e.g., MAKOTO's candid discussion of its own welfare-estimate correction earned it high marks on the subjective scoring axis).

Closing

Mathematics is never mere decoration propping up an AI competition paper's formatting — it is the steel skeleton of the entire edifice of scientific reasoning.

If you own math4aiml, you're holding a map to the deepest layer of AI's core technology. Sit with it, and connect the book's matrix decompositions, optimization solvers, and probability dynamics to your specific scientific competition task — biology, meteorology, social science. Once you can derive and implement a numerical algorithm that belongs entirely to your own project, you'll already be standing on undefeated ground on the academic-competition stage.

In future posts, I'll go through specific math topics, and some of the finer places where points quietly get lost.

Wishing everyone great strides and great results on the path of AI learning!