Take a photo of a cat. Wrap it in elephant skin.
A 3-year-old says "cat."
A billion-parameter vision model says "elephant."
That's the gap nobody talks about. We see shapes. Machines see texture.
So a team asked a strange question: what if we stopped feeding AI perfect images?
A newborn sees the world blurry. Almost no color. Almost no contrast. It takes years to sharpen.
They rebuilt that. Blurry first. Grayscale. Low contrast. Then slowly sharper — an infant's timeline, turned into a training curriculum.
No new data. Same images. Just a different order.
The result, published in Nature Machine Intelligence: the strongest shape bias ever recorded in a vision model, and far better resistance to noise and adversarial attacks.
لا توجد تعليقات بعد. كن أول من يعلّق!