AI That Only Saw the Middle of Photos Just Got a Fix

Most AI today learns by being shown millions of examples and having its mistakes corrected mathematically. This paper explores a very different approach: AI that evolves, more like breeding than teaching. Previous work found something strange. When this evolving AI was asked to recognize handwritten digits, it kept focusing on a tiny patch in the middle of each image and ignoring the edges. Its accuracy stuck at 21% — roughly a coin flip with a few lucky guesses.

The obvious question was whether the AI was simply bad, or whether the setup itself was forcing it to look in the wrong place. To find out, the researchers borrowed an idea from modern AI called 'mixture of experts' — the same principle as sending different parts of a job to different specialists rather than one person doing everything. They sliced each image into 13 non-overlapping zones and let each zone evolve its own small network. Accuracy more than doubled to 43%.

Crucially, the improvement wasn't about cleverly combining the answers. Simply averaging all 13 specialists equally, with zero extra tuning, still delivered a 70% gain. The win came from the dividing itself. When the researchers examined which pixels each network actually reacted to, the picture was dramatic: coverage of the image jumped from 4% to 79%.

So what? Two honest caveats. First, 43% is still far below what everyday image-recognition systems achieve — this is not a product. Second, the paper's real contribution is a pair of general-purpose tools: a diagnostic that detects when an AI has gone blind to parts of its input, and a partitioning fix. Those could apply well beyond this one experiment, wherever AI is quietly missing half the picture.