Intro

2019-07-03_10h45_14.png

Above one is subjective, could vary some what, and its easy for humans to do.

Cases Revisited

  • We earlier saw, kNN, and discrimination attempts to solve same domain navigation problems. A new problem is in same domain as previous cases.

2019-07-03_10h46_37.png

2019-07-03_10h46_58.png

So how to handle cross domain problems?

Need for Cross Domain Strategy

2 Stories:

  1. A strong laser beam could kill the patient. What doc can do?
  2. A fort on way has mines which gets detonated if army crosses over. Instead army split, attacked and win.

From 2nd story, one could abstract the strategy and apply to first. Use lesser intense multiple lasers.

Spectrum of Similarity

2019-07-03_10h57_16.png

Process

RHS is case based reasoning, given for comparison.

2019-07-03_11h03_24.png

Retrieval

2019-07-03_11h05_40.png

In Raven's problem,

Superficial similarity

  • Features refer to size of the square, whether it is filled or hollow etc
  • Counts refer to no of object
  • Objects refer to circle, diamond etc

Note

  • Superficial similarity is kinda unary relationship
  • Relationship between objects (RBO) is kind of binary relationship
  • Relationship between relationship (RBR) is kind of tertial relationship

Deep Similarity

  • Dot is on right side of circle (RBO)
  • Dot is above circle (RBO)
  • The square inside circle in A has come out side in B (RBR)
  • The dot in A has disappeared in B (RBR)

As we go from unary to tertiary and so on, the simialarity goes deepe in context.

Types of Similarity

2019-07-03_11h15_48.png

  • Semantic example: Women climbing up a ladder, Women climbing up a step ladder - both conceptually similar
  • Pragmatic: doctor wants to kill tumor, army wants to capture fort - similarity in goals
  • Structural: we will see shortly

Note one might give different weightage to different similarity here, that is why our answers could differ in our 1st question above.

Quiz Analogical Retrieval I

2019-07-03_11h23_32.png

Sentence Reason
A woman climbing a set of stairs
A womain climbing a step ladder
Both objects are similar, and their relationshipts also.
That is why both deep and superficial, and ranked toppest
An ant walking up the wall Only deeply similar, as only relationshipts match
A womain painting a ladder Only superficial because though there are similar objects, no common relation
A woman climbing corporate ladder I would argue both because objects are also similar apart from relation
A water bottle.. nothing is common
A plan atking off in to the sky again only relationships same, so its deep

Analogical Retrieval II

2019-07-03_11h30_47.png

Analogy Mapping

2019-07-03_11h38_46.png

  • if you look at first half alone it does not matter if sun maps to nucleus or electron maps to nucleus and remaining object to planet.
  • but including the lower part, where planet revolves aruond sun, it would become clear, electron should map to planet insetad of sun and so on.

Analogy Transfer

2019-07-03_11h42_53.png

  • First, analogical retrieval has given the source case of fort for given target problem.
  • Analogical Mapping is also done.
Source Target
King Tumor (not patient)
Army Laser beam
Mine Healthy tissue

For source we also know solution.

Source: Army divided in smaller groups and they all attack the king at fort at same time.
Target: Laser beam is divided in lesser intense onse, and attack the tumor at same time

Note, source and target has similar goals - Its an example of pragmatic similarity.

Quiz - Structural Similarity

2019-07-03_11h56_28.png

Why?

Note neither the objects or goals between both systems are same. Only relationships are structurally similar.

Evaluation & Storage

Same as we saw earlier. Encapsulate and store, so this could become source for new problems. Thus agent is also learning incrementally here.

Design by Analogy

2019-07-03_12h02_17.png

The shock wave created by high speed train was solved by using beak design of the bird above, which was observed to be reduced the similar waves, when it captures prey from air to water medium.

Design by Analogy : Mapping & Transfer

Imagine we have done a complex analogical retrieval of studying an insect (source) to make a robot(target) walk on water.

analogy_1.gif

analogy_2.gif

Looping

also possible.. a transfer may lead to new mapping and so on..

2019-07-03_12h15_20.png

There can be looping between transfer and retrieval also.. In below example first slow speed is solved, till transfer, and then for fast speed, we again went back to retreival mode..

2019-07-03_12h17_58.png

2019-07-03_12h18_55.png

If evaluation fails, there is also looping possible as below. You may want to try different transfer or mapping or retrieval..

2019-07-03_12h20_30.png

Advanced questions

  • Common vocabulary If I describe planet rotates in an orbit around the sun, and electron revolves around nucleus, then we cannot map both. So what to do in case of no common vocabulary?

  • Abstraction and Transformation So far we saw, target is fixed, but source only abstracted. Often target may also need to be abstracted and transfered.

  • Compound analogies Imagine designing a car with engine from one source, chasis from another etc. how to make it work?

  • Compositional analogies Compare 2 biz org. We might do first at level of people, next at level of processess and so on.

  • Visuospatial so far we only talked via causal relation. We will see visuospatial later.

  • conceptual combination Learning new concept from parts of different concept. Oh wait this sounds simialr to compound analogy as well.