Skip to content

Latest commit

 

History

History
26 lines (24 loc) · 1.69 KB

online_vs_offline.md

File metadata and controls

26 lines (24 loc) · 1.69 KB

Fast vs. slow (online vs. offline) learning

#draft/aitp_thoughts


Humans process information at vastly different rates

  • Our reflexes respond before the signal even gets to our brains.
  • We can complete a sentence or recognize an object without any thought.
  • We can store a list of objects in our short term memory.
  • We can remember things long term
  • Important information we remember our whole lifetime (and dreams seem to be an important mechanism here for preserving this information).
  • Human knowledge is built over generations, slowing being refined.
  • Our brains come with certain instincts built-in via millennia of evolution.

In the same way, information can come into a learned model at different speeds.

  • A neural network or machine learning model can respond to a short input instantly.
  • Sequence models can take in very large inputs with a lot of information, and hold that information in memory long enough act on it.
  • kNNs and other retrieval mechanisms directly store data which can be quickly retrieved through various mechanisms.
  • Small neural networks can be trained really quickly, even in real time.
  • Pre-trained medium and large neural networks can be quickly fine-tuned (seconds, minutes, hours, days depending on the size of the network and size of the data).
  • Medium size neural networks can be trained in minutes, hours, or days depending on their size and the amount of training data, but they can be hour or day.
  • Large neural networks hold a huge amount of information but require many resources to train from scratch.

Different aspects of fast verse slow:

  • Permanent vs. forgettable
  • Old knowledge vs new knowledge
  • Fixed vs adaptable
  • Development vs production environments