11. Jeopardy! に関する補足説明
• Clue (ヒント)
– REGARDING THIS DEVICE, ARCHIMEDES
SAID, "GIVE ME A PLACE TO STAND ON,
AND I WILL MOVE THE EARTH"
• Correct Response
– WHAT IS A LEVER? (てこ)
12
41. sequence-to-sequence learning
framework with attention
LSTM, stacked LSTM, residual connections, bidirectional LSTM, sub-word units, …
知识就是力量
Knowledge is power
それまでに読み込んだ全ての単
語の意味を表現しているベクトル
1単語ずつ出力
どこに注目
するかを変
化させつつ
42. word2vec
単語
one-hot vector
周りの単語
one-hot vector
300次元とかの
vector
vector X = vector(”biggest”)−vector(”big”) + vector(”small”)
cosine distanceでXに一番近い単語→smallest
Paris – France + Japan = Tokyo
[Mikolov+, Efficient Estimation of Word Representations in Vector Space, 2013]
「Parisという単語の周りに出てくる単語」と「Franceとい
う単語の周りに出てくる単語」の違いは、「Tokyoという単
語の周りに出てく単語」と「Japanという単語の周りに出
てくる単語」の違いと似ている。
68. Formal Theory of Creativity & Fun &
Intrinsic Motivation (1990-2010) by Jürgen
Schmidhuber
http://people.idsia.ch/~juergen/creativity.html
• (A) an adaptive predictor of the growing
data history as the agent is interacting with
its environment
• (B) a reinforcement learner selecting the
actions that shape the history
• (B) is motivated to learn to invent
interesting things that (A) does not yet
know but can easily learn.
69. (つづき)
• To maximize future expected reward, (B)
learns more and more complex behaviors
that yield initially surprising (but eventually
boring) novel patterns that make (A)
quickly improve.
70. (つづき)
• O(t): the state of some observer O at time t
• H(t): its history of previous actions &
sensations & rewards until time t
• Beauty B(D,O(t)) of any data D: the negative
number of bits required to encode D
• Interestingness I(D,O(t)) of data D for
observer O at discrete time
step t>0: I(D,O(t))= B(D,O(t))-B(D,O(t-1))
72. (つづき)
Implementations
• Intrinsic reward: prediction error
• Intrinsic reward:
improvements in prediction error
• Intrinsic reward: relative entropies
between the agent's priors and posteriors
73. 参考資料
• Second Interdisciplinary Symposium on
Information-Seeking, Curiosity and Attention
https://openlab-flowers.inria.fr/t/second-
interdisciplinary-symposium-on-information-seeking-
curiosity-and-attention-neurocuriosity-2016/187
• Information-seeking, curiosity, and attention:
computational and neural mechanisms
http://www.pyoudeyer.com/TICSCuriosity2013.pdf