Neither convinced that a baby uses big data, nor for sure how a baby recognizes a cat much faster and more effective than the current deep learning neural networks with millions of cat pictures, but a wild guess would be,

  • Feature extraction: with a mysterious clever method
  • Pattern match: with a raw, inaccurate and fussy method

Both methods have been developed on top of an organ called brain of the baby for just a few months or years. The brain hardware has been developed/evolved for 6 million years, someone may say much more. It is almost empty in a new-born baby. However, it does get the simplest operation system with some underdeveloped sensors,

  • Sight
  • Hearing
  • Taste
  • Touch

and a quite sensitive sensor, feeling hungry. ;)<p>

The operation system gets only a few functions,

  • Crying
  • Sucking milk

However, the architecture of the brain is so optimized that it makes a so-called “baby learning” easily be implemented after birth. Opposite to the machine learning of current AI, baby learning needs only a small set of data working in a highly-optimized brain architecture. With baby learning function “implemented”, the two methods, feature extraction and pattern match, become add-on features.

I have to go again down to the road of reverse engineering of things happening in baby’s brain, which has the difficulty proportionally to a billion years of evolution according to Moravec’s paradox.