Skip to content

Latest commit

 

History

History
81 lines (56 loc) · 3.79 KB

L10 Incremental Concept Learning.md

File metadata and controls

81 lines (56 loc) · 3.79 KB
  • Incremental: show the agent one example at a time and the agent learns incrementally (instead of learning from a large set of examples at once)
    • Learning is often incremental: We learn from one example at a time
  • Examples are often labeled. e.g. Teacher tells us this a positive example or a negative ex�ample (supervised learning in ma�chine learning)
  • Examples can come in a particular order: first example is typically a positive example.
  • In case based reasoning, we store examples in their raw form in memory and reuse them. In incremental concept learning, we abstract concepts
  • Number of examples from which we’re extracting concepts is very small

We are trying to abstract concepts from examples, but what exactly to abstract? What exactly, to learn? What exactly to generalize? (tendency to often overgeneralize, or overspecialize)

Generalization or specialization

Variabilization

  • If we want to teach an AI agent to recognize arches, we can tell it about the definition of an arch (i.e. 4 bricks with a relationships shown in the right figure)
  • If the agent sees the example in the left, it is able to variabilize the bricks (mapping objects to variables, e.g. top brick is an instance of Brick A, left brick is an instance of Brick C, etc.)
    • right pic shows the variabilization of the example
  • If the agent sees a new example AND the current concept does not include or exclude the example:
    • it sees a positive example: generalizes if the positive example is not covered by the current concept
    • it sees a negative example: specializes current definition to exclude the negative example
  • A concept that an agent has depends on the agent's background knowledge (i.e. the definition and examples we gave to the agent)

Heuristics for specializing and generalizing

  • A concept can be characterized with many, many features... heuristics help to reduce the dimensions of the learning space

"Drop-link" heuristic

  • Generalization to ignore features
  • If we show the picture below and tell the agent it's a positive example:
    • it updates the current concept to accommodate the example:
    • update may be done using heuristics, e.g. drop-link heuristic (link indicated in the picture below is dropped)
    • the new concept can accommodate both the example and the previous concept
    • drop-link heuristic: useful when the new and old concepts overlap except an extra link

Require-link heuristic

  • Specialization to require features
  • e.g. show the agent the negative example below
  • links not in common between the example and the current concept must be required

Forbid-link heuristic

  • Specialization to exclude features
  • e.g. show the agent the negative example below

Enlarge-set heuristic

  • Generalization to abstract features
  • e.g. show the agent the positive example below

Climb-tree heuristic

  • Generalization with background knowledge
  • Update concept with background knowledge: brick wedge are both "blocks"
    • i.e. it "climbs" the "knowledge tree" (see the tree in the figure below)
  • Advantage: the more we know the more we can learn

Close-interval heuristic

  • expand range of values to be a positive example of the concept
    • e.g. update the concept of a dog to include large dogs by increasing the range of body length

Cognitive connection

  • Incremental concept learning (learning from one example at a time) is much closer to human's learning than e.g. machine learning with a large set of example