4. 今日のお話
‣ Bayesian models of cognition
- 一部で論争(?)が起きている
- Bayesian just-so stories in psychology and neuroscience
(Bowers & Davis, 2012)
5. 今日のお話
‣ Bayesian models of cognition
- 一部で論争(?)が起きている
- Bayesian just-so stories in psychology and neuroscience
(Bowers & Davis, 2012)
- 心理学におけるBayes理論は科学ではなく,単なる作り話である
6. 今日のお話
‣ Bayesian models of cognition
- 一部で論争(?)が起きている
- Bayesian just-so stories in psychology and neuroscience
(Bowers & Davis, 2012)
- 心理学におけるBayes理論は科学ではなく,単なる作り話である
- How the Bayesians got their beliefs (and what those beliefs actually are):
Comment on Bowers and Davis (2012) (Griffiths, Charter, Norris & Pouget)
- 上の著者は,この分野を正しく理解していない!という反論
7. 今日のお話
‣ Bayesian models of cognition
- 一部で論争(?)が起きている
- Bayesian just-so stories in psychology and neuroscience
(Bowers & Davis, 2012)
- 心理学におけるBayes理論は科学ではなく,単なる作り話である
- How the Bayesians got their beliefs (and what those beliefs actually are):
Comment on Bowers and Davis (2012) (Griffiths, Charter, Norris & Pouget)
- 上の著者は,この分野を正しく理解していない!という反論
- Is that what Bayesians believe? Reply to Griffiths, Chater, Norris, and Pouget
(2012) (Bowers & Davis, 2012)
- さらに元の著者からの反論 …
8. 今日のお話
‣ Bayesian models of cognition
- 一部で論争(?)が起きている
- Bayesian just-so stories in psychology and neuroscience
(Bowers & Davis, 2012)
- 心理学におけるBayes理論は科学ではなく,単なる作り話である
- How the Bayesians got their beliefs (and what those beliefs actually are):
Comment on Bowers and Davis (2012) (Griffiths, Charter, Norris & Pouget)
- 上の著者は,この分野を正しく理解していない!という反論
- Is that what Bayesians believe? Reply to Griffiths, Chater, Norris, and Pouget
(2012) (Bowers & Davis, 2012)
- さらに元の著者からの反論 …
- 何が起きているのか?
- 議論が起こっているということは,誤解が生じやすいということ.その部分を
明らかにしたい
9. 認知科学と教師なし学習
‣ 自然言語処理と先ほどの議論は,関係が0ではない?
‣ Computational linguistics Where do we go from here?
(ACL2012, Mark Johnson)
- 現在の計算言語学はサイエンスではない
- 精度(f値)を1%あげても,言語の本質に近づいたとは言えない
- 計算言語学として言語の本質に近づくためにはどうすれば良いか?
‣ 特に言語の教師なし学習に関して…
- ベイズで認知モデルを組み立てることと,自然言語の教師なし学習モデルを作る
ことと,やっていることは同じ(どちらも生成モデルと推論法を考える)
- 認知モデルへの非難を受けて,教師なし学習はどのようなことを考えて進めば
良いか?
28. computational model can identify which information sources
ce to do something
I
赤ちゃんが音素列から単語をどのように得るか?
word segmentation is first step to learning a lexicon
y Mu Nw Ma Mn Mt Nt Mu Ns Mi ND Me Nb MU Mk (Johnson, 2012)
I using distributional information and syllable structure achieves
about 90% token f-score
‣ 赤ちゃんが言葉をどのように学習するか?の最初の問題
nergies in acquisition:
‣ 母親の話す言葉は,単語に切れてはおらず,連続している
I learning word segmentation and syllable structure jointly learns
both more accurately than learning each on its own
‣ 赤ちゃんは連続した音素のみから,単語の切れ目を見つけている,と考える
I learning word object mapping together with word segmentation
‣ 生成モデル
improves word segmentation accuracy
- p(H) DirichletProcess(α,P0) : 各単語の出現確率 ex) p( dog ) = 0.01
Animals don’t move on wheels”
- p(D¦H) = Multinomial(H) : 独立に生成された単語がくっついて,母親が発する確率
– Tom Wasow
‣ 推論
e: Fleck, Goldwater, Swingley and many others p(H) と文の区切りを見つける
- 単語が繋がった文の集合のみから,単語の集合
‣ このモデルが,正しい単語の区切りを見つけることが出来たら,赤ちゃんは
21/43
このような確率モデルを頭の中に持っていると言える(?)
29. カテゴリーの獲得
‣ Categorization theory (認知言語学などと関連)
- 人間は, もの をカテゴリーに分類して,理解している
- 見た目が犬っぽい動物は,全て 犬 として認識する
- 種類が違うりんごは,全て りんご というカテゴリーとして認識される
‣ Exemplar Model と Prototype Model
35. Bowers & Davis (2012) のabstract
According to Bayesian theories in psychology and neuroscience, minds and brains are (near)
optimal in solving a wide range of tasks. We challenge this view and argue that more
traditional, non-Bayesian approaches are more promising. We make 3 main arguments. First,
we show that the empirical evidence for Bayesian theories in psychology is weak. This
weakness relates to the many arbitrary ways that priors, likelihoods, and utility functions can
be altered in order to account for the data that are obtained, making the models unfalsifiable.
It further relates to the fact that Bayesian theories are rarely better at predicting data
compared with alternative (and simpler) non-Bayesian theories. Second, we show that the
empirical evidence for Bayesian theories in neuroscience is weaker still. There are impressive
mathematical analyses showing how populations of neurons could compute in a Bayesian
manner but little or no evidence that they do. Third, we challenge the general scientific
approach that characterizes Bayesian theorizing in cognitive science. A common premise is
that theories in psychology should largely be constrained by a rational analysis of what the
mind ought to do. We question this claim and argue that many of the important constraints
come from biological, evolutionary, and processing (algorithmic) considerations that have no
adaptive relevance to the problem per se. In our view, these factors have contributed to the
development of many Bayesian “just so” stories in psychology and neuroscience; that is,
mathematical analyses of cognition that can be used to explain almost any behavior as
optimal.
39. Mark Johnson (2012)
‣ A computational model can identify which information sources suffice
to do something
‣ Synergies in acquisition:
- learning word segmentation and syllable structure jointly learns
both more accurately than learning each on its own
- learning word → object mapping together with word segmentation
improves word segmentation accuracy
‣ 生成モデルにこんな情報を組み込んだら,こういう現象が観測された,
ということが大事
- ベイズモデルだと,そのような情報を陽に記述するので,現象が理解しやすい
‣ optimal learner = 組み込んだ情報を最大限に使えることが保証されている
‣ 人間も,そこで組み込んだ情報を使っているのではないか?という手掛か
りになる
41. Dan Klein (2005)
‣ The unsupervised learning of natural language structure
- To be clear on this point: the goal of this work is not to produce a
psychologically plausible model or simulation. However, while success at the
tree induction task does not directly speak to the investigation of the human
language faculty, it does have direct relevance to the logical problem of
language acquisition, particularly the argument of the poverty of the
stimulus, and therefore an indirect relevance to cognitive investigations. In
particular, while no such machine system can tell us how humans do learn
language, it can demonstrate the presence and strength of statistical patterns
which are potentially available to a human learner.
‣ 同じような主張をしている
42. Yoav Seginer (2007)
‣ Learning syntactic structure
- Even when a computational model is clearly not psychologically realistic, its
success in learning syntactic structure has important implications to the study
of language and language acquisition because such successful learning
indicates a relation between the surface structure of a language and its
hidden syntactic structure. Even if the method by which this relation is
established is not actually used by children acquiring a language, the relation
is still an empirical property of the language and may be used by children in
some other way in the process of language acquisition.