Want Smarter AI? Scientists Found a Way to Tune It Automatically
Less human guesswork could make AI brains faster and cheaper to build.
Using a tree-structured Parzen estimator (TPE), researchers optimized ES-HyperNEAT’s hyperparameters for an MNIST handwriting task, exploring a search space of over 3 billion possible combinations. TPE clearly beat random search in mean, median, and best accuracy, and the top setup reached 29.00% accuracy on MNIST—better than previous studies, even with a smaller population and fewer generations. When tested on other tasks, those hyperparameters transferred well to the more complex Fashion-MNIST image problem, but transfer was limited for simpler logic operations. The work highlights a path to unlocking more from neuroevolutionary algorithms and offers insight into how hyperparameters transfer across tasks of different complexity.
- The paper investigates how to automatically tune the 'dials' of evolution-based AI models, testing over 3 billion possible settings.
- The new method, TPE, beats random guessing — boosting accuracy on handwritten digits to 29%, a record for this type of AI.
- Tuned settings worked on a harder fashion-image task but failed on simple logic puzzles, showing that one-size-fits-all tuning doesn't exist.
Why It Matters
This work brings us closer to AI that designs itself, saving time, money, and human expertise in building specialized systems.