티스토리 뷰

Research (연구 관련)

Ad hoc categories

홍돌 2019. 5. 23. 21:13

Ad hoc: "이것을 위해" 즉 "특별한 목적을 위해서" == "for this purpose"

Object Categorization (Bags of Words Model) 수업 중 Ad hoc Category가 잠깐 언급됐다. 
일반적인 카테고리들(Common Categories)로는 과일, 바지 등이 있고 그 외의 카테고리들로는 functional categories, ad hoc categories가 있으며 각각 
chairs = “something you can sit on”,  “something you can find in an office environment”가 예시로 들어졌는데 둘이 무슨 차이인지, 특히 Ad-hoc category 자체의 정의가 무엇인지 헷갈렸다. 

그래서 구글링을 엄청했는데 한글은 물론이고 영어 결과도 별로 없었다. 알고보니 심리학 용어인데, 1983년 논문인 <Ad hoc categories>에서 설명을 해놓았다. 대충 내가 감 잡을 수 있을 정도로만 읽었는데, 이 논문에 따르면 원래 '카테고리'에는 Common categories와 Ad-hoc categories가 있다. Common category는 정말 우리가 흔히 아는, 또한 보편적으로 공유되는 category(집, 가구, 동물, 등등)이고 Ad-hoc category는 보편적이지 않으며 사람마다 특정한 목적을 위해 개개인이 가지고 있는 범주(category)인 것 같다. "things to take from one's home during a fire." - 강아지, 지갑, 통장... "things to pack in a suitcase." - 양치도구, 수건... 때문에 무수히 많은 Ad hoc category가 있으며 보통은 많은 사람들이 Ad hoc이라는 카테고리를 머릿속에 가지고 있지 않다.

Common categories
Ad hoc category: 여행 갈 때 챙길 것

Common category는 두 가지 특성을 가지는데, 하나는 "Graded Structure"이고 다른 하나는 "Well Establisehd Category Representation"이다. 결론부터 말하면 Ad hoc category는 Graded Structure가 있지만, Well Established Category가 없다. 논문에 따르면 Ad hoc categories는  Graded Structure가 있어서 대상들의 분류가 가능한(?) 범주라고 한다. 일단 두 가지의 의미를 알아보자. Well Established Category, Correlatinoal Structure, sharing correlated properties 다 같은 말 같다. 쉽게 말해 새, 차량, 물고기는 각각 굉장히 다른 특성을 가지고 있고 자기들끼리는 그런 특성을 굉장히 비슷하게 가지고 있다. 새는 날고, 날개가 있고 물고기는 헤엄치고 지느러미가 있고 등등. 그런데 Ad hoc category 중 하나인 ""things to take from one's home during a fire." 를 생각해보자. 강아지와 지갑이 무슨 특성을 공유하는가?

Category representation. A central difference between Common and ad hoc categories appears to be that common categories have well established category representations in memory, whereas ad hoc categories do not. Ad hoc categories are not wen established simply because people rarely, if ever, think of them. This precludes the development and strengthening of associations between the nodes representing them.

Graded Structure는 세가지 특성이 있다고 한다. 하나는 어떤 객체들은 다른 객체보다 그 카테고리에 대해 더 좋은 예라는 것. 의자는 가구란 카테고리에 대해 책장보다 더 좋은 예이다. 더 좋다는 건 '가구'하면 먼저 떠오르는 객체가 의자라는 말 같다. 둘째는 불분명한 케이스들이 카테고리 범주 안의 객체 중에 있다는 것이다. 라디오는 가구인가? 셋째는 그 카테고리 범주에 안들어가는 객체마다 유사성이 다 다르다는 것이다. 타자기는 강아지보다 스테레오장치와 더 비슷하다. 참고로 타자기는 음향장비가 아니므로 스테레오장치에 속하진 않는다. 아무튼 종합하면 Graded Structure란 객체마다 카테고리에 속하는 정도에 차이가 있다는 걸 의미하는 것 같다. 아래는 Graded structure에 대한 원문이다.

First, some instances are better examples of a category than are others; "chair" is a more typical example of "furniture" than is "bookcase." This aspect of graded structure has been found in all common categories investigated so far. Rosch (1973, 1975b) found typicality in color categories (e.g., red, green). Rips, Shoben, and Smith (1973) and Rosch (1973, 1975a) found typicality in common semantic categories (e.g., fruit, clothing). Rosch and Mervis (1975), Rosch, Simpson, and Miller (1976), Smith, Shoben, and Rips (1974), and Tversky (1977) have since argued that the typicality of a category member increases as it becomes more similar to other category members. The second aspect of graded structure is the presence of unclear cases, items whose category membership is uncertain (McCloskey & Glucksberg, 1978); people are not sure whether "radio" belongs to the category of "furniture." The third aspect of graded structure is that the nonmembers of a category (i.e., its complement) vary in how similar they are to the concept of the category; "typewriter" is more similar to the concept of "stereo equipment" than is "dog." This aspect has commonly been cited as the reason some false items take longer to reject in the category verification task than do others (McCloskey & Glucksberg, 1979; Smith et al., 1974); "bat" takes longer to reject as a member of "birds" than does "chair." No evidence bears on whether such similarity gradients cause subjects to reliably rate nonmembers of a category for typicality within the complement of the category.

논문에서는 실험을 통해서 Ad hoc category가 graded structure를 가지고 있고 이것을 통해서 분류를 한다 뭐 이렇게 결론 내리는 것 같다. 확실하지 않다. 또한 참고로 functional category는 common category의 common property로 function이 들어갈 수 있기 때문에 common category의 하나로 볼  수 있을 것 같다.

그럼 다시 비전으로 돌아와서. Ad hoc category를 가지고 Bags of Word Model을 사용할 수 있나? 그게 의미가 있나? 인공지능이 집에 불이 났을 때 자동적으로 집에서 가져고 나와야할 리스트를 정리해준다면 유용하다. 의미가 있다. 그런데 Bags of Word Model을 적용하려면 특성들이 있어야하는데.. 이쯤되면 비전의 영역인지도 모르겠다 하하 논문에 따르면 Ad hoc category는 common properties가 없으니 Bag of Words Model이 적용안될것같은데, 그럼 이 문제는 어떻게 해결하지? 개개인마다 다르니, 개인의 선택을 강화학습해야하나

 

 

객체의 특징들을 Words로 만든다
다른 두 객체에 대해 어떤 "Words"를 얼만큼 가지는지를 가지고 차이를 계산할 수 있다
어떻게 구별하는지는 또 다른 문제로 Linear model(선그어서 구분), Nearest Neighbor가 있다.

 

'Research (연구 관련)' 카테고리의 다른 글

sudo w  (0) 2019.06.01
Binocular Disparity  (0) 2019.05.27
Why do we need validation set?  (0) 2019.05.15
netD(fake.detach())의 이유  (7) 2019.05.11
Transposed Convolution  (0) 2019.05.10
공지사항
최근에 올라온 글
최근에 달린 댓글
Total
Today
Yesterday
링크
«   2024/05   »
1 2 3 4
5 6 7 8 9 10 11
12 13 14 15 16 17 18
19 20 21 22 23 24 25
26 27 28 29 30 31
글 보관함