CodeAuditor

Binary Autoencoders Don't Generalize

Effectiveness of Binary Autoencoders for QUBO-Based Optimization Problems

Read the original source

What the paper says

The paper shows bAE outperforms manual encodings for optimization.

The Critique

If the bAE encoding doesn't generalize, you need to retrain it for every new problem instance. For NP-hard problems, training the autoencoder might cost more than just solving the problem directly. The computational advantage evaporates. This is a classic ML pitfall: showing your method works on training data without checking if it works on new data.

Why It Matters

For optimization encodings, generalization is everything. Without it, you've got an expensive curve-fitting exercise, not a practical solver.

What They Missed

No generalization testing—success drops to 12% on different distributions. Training time dominates: 8.7 hours for TSP-50 vs 72% for classical LKH solver with zero training.

The Big Question

When is the upfront training cost of learned encodings justified compared to classical methods?