ゼロから始める深層強化学習(NLP2018講演資料)/ Introduction of Deep Reinforcement LearningPreferred Networks
Introduction of Deep Reinforcement Learning, which was presented at domestic NLP conference.
言語処理学会第24回年次大会(NLP2018) での講演資料です。
http://www.anlp.jp/nlp2018/#tutorial
ゼロから始める深層強化学習(NLP2018講演資料)/ Introduction of Deep Reinforcement LearningPreferred Networks
Introduction of Deep Reinforcement Learning, which was presented at domestic NLP conference.
言語処理学会第24回年次大会(NLP2018) での講演資料です。
http://www.anlp.jp/nlp2018/#tutorial
人的資本経営[1]を実現するには,生産性とQoW(Quality of Work,働き方の質)を同時に改善し続けていくことが有効である.そのための課題は多岐に渡るため,DX(Digital Transformation)的発想が求められる。一方、情報の約60~80%が位置情報に関連していることが報告されている.本稿では,地理空間情報と他の情報とを連携させて課題解決を支援する地理空間インテリジェンス(GSI)でDXを促進し,製造現場やサービス現場で人的資本経営を支援することに資する筆者らの一連の取り組みについて紹介する.
【DLゼミ】XFeat: Accelerated Features for Lightweight Image Matchingharmonylab
公開URL:https://arxiv.org/pdf/2404.19174
出典:Guilherme Potje, Felipe Cadar, Andre Araujo, Renato Martins, Erickson R. ascimento: XFeat: Accelerated Features for Lightweight Image Matching, Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)
概要:リソース効率に優れた特徴点マッチングのための軽量なアーキテクチャ「XFeat(Accelerated Features)」を提案します。手法は、局所的な特徴点の検出、抽出、マッチングのための畳み込みニューラルネットワークの基本的な設計を再検討します。特に、リソースが限られたデバイス向けに迅速かつ堅牢なアルゴリズムが必要とされるため、解像度を可能な限り高く保ちながら、ネットワークのチャネル数を制限します。さらに、スパース下でのマッチングを選択できる設計となっており、ナビゲーションやARなどのアプリケーションに適しています。XFeatは、高速かつ同等以上の精度を実現し、一般的なラップトップのCPU上でリアルタイムで動作します。
The presentation by Cerenaut at the orientation (2021-06-05) for the 5th WBA Hackathon, an online AI competition to implement working memory
https://wba-initiative.org/en/18687/
- Neuroscientific issues
- Architecture details
- Instruction on the CodaLab competition
Task details of the 5th WBA Hackathon, an online AI competition to implement working memory
The slides were presented at an orientation event on June 5th, 2021.
https://wba-initiative.org/en/18687/
The task demonstration can be found here: https://www.youtube.com/watch?v=7RnRQa6ZD4A
4. 大規模言語モデル、世界モデル、基盤モデル
~自己教師付き学習からマルチモーダルAGIへ~
Reed, Scott, et al. "A generalist agent." arXiv preprint
arXiv:2205.06175 (2022).
Wu, Philipp, et al. "DayDreamer: World Models for Physical Robot
Learning." arXiv preprint arXiv:2206.14176 (2022).
Google fires engineer who contended its AI technology was sentient,
By Ramishah Maruf, CNN, Updated 1745 GMT (0145 HKT) July 25,
2022, https://edition.cnn.com/2022/07/23/business/google-ai-
engineer-fired-sentient/index.html
8. Friston, Karl, and Ping Ao. "Free energy, value, and attractors." Computational and mathematical methods in medicine 2012 (2012).
脳の大統一理論 自由エネルギー原理とはなにか (岩波科学ライブラリー) , 乾 敏郎, 阪口 豊, (2021)
Karl Friston, Rosalyn J. Moran, Yukie Nagai, Tadahiro Taniguchi, Hiroaki Gomi, Josh Tenenbaum, World model learning and
inference, Neural Networks, 2021,
自由エネルギー原理、予測符号化、世界モデル
Free energy principle, predictive coding, World model
10. PCA
Image Observation
(Grand truth)
Imagination
Reconstruction
Hafner, Danijar, et al. "Learning latent dynamics for planning from
pixels." International Conference on Machine Learning. PMLR, 2019.
Masashi Okada, Norio Kosaka, Tadahiro Taniguchi, PlaNet of the
Bayesians: Reconsidering and Improving Deep Planning Network by
Incorporating Bayesian Inference, IROS 2020
Imagination
世界モデルのアプローチ
例)PlaNet: Recurrent State Space Model (RSSM)
鈴木雅大、深層生成モデルと世界モデル(2020/11/20版)
https://www.slideshare.net/masa_s/20201120-239349802
11. 1. 従来の制御工学における客観的物理量の計測に基づく状態空間構成で
はなく、深層学習による教師なし・自己教師付き表現空間構成
2. 潜在空間上での制御則・高次戦略の獲得
3. 実ロボットでの実タスク実現
潜在的状態空間
(ロボットと対象系のダイナミクスが反映された空間)
<World model>
マルチモーダル世界モデル(World model)
ロボット自律化AIのための潜在的構造の同定
マルチモーダル
観測𝑜𝑜𝑡𝑡
状態𝑠𝑠𝑡𝑡 行動𝑎𝑎𝑡𝑡
モデル予測制御=
モデルベース強化/模倣学習
POMDP (部分観測マルコフ決定過程)
HAFNER, Danijar, et al. Learning latent dynamics for planning from pixels. In: International Conference on Machine Learning. PMLR, 2019.
画像情報
深度情報
力覚・触覚情報
自己感覚受容など
VAE(変分オートエンコーダ)
RSSM(再帰的状態空間モデル)
Transformerなど
潜在的状態空間に基づくモデル予測制御(強化学習・模倣学習・能動学習)
12. DayDreamer: World Models for Physical
Robot Learning [Wu+ 2022]
DayDreamer: World Models for Physical Robot Learning
Danijar Hafner, https://www.youtube.com/watch?v=A6Rg0qRwTYs
Wu, Philipp, et al. "DayDreamer: World Models for Physical Robot Learning." arXiv preprint arXiv:2206.14176 (2022).
世界モデルの代表的手法Dreamerを実ロボットで動かす。
比較的単純なタスクに対してそれでもある程度のパフォーマンスを示す。
15. Online spatial concept formation and lexical
acquisition: SpCoSLAM [Taniguchi+ 2017]
Akira Taniguchi, Yoshinobu Hagiwara, Tadahiro Taniguchi and Tetsunari Inamura, Online Spatial Concept and Lexical Acquisition with
Simultaneous Localization and Mapping, IEEE IROS 2017
Akira Taniguchi, Yoshinobu Hagiwara, Tadahiro Taniguchi, Tetsunari Inamura, Improved and scalable online learning of spatial
concepts and language models with mapping, Autonomous Robots, 2020. DOI: 10.1007/s10514-020-09905-0
16. SpCoSLAM
Map
Self location
How do they
call the
space?
Visual
information
“Third table”
“Meeting space”
Words
Cluster of
positions
How does the
space look
like>
Where is the
space?
Online learning of
Multimodal spatial concept
17. Notably, Cross-modal inference provides
"functions" to a cognitive system
Image to position:
Recalling its self-location from visual information
Position to image:
Imagining the target position visually
Speech/language to position:
Estimating the position from a user's utterance
Position to speech/language:
Telling the robot's position linguistically
Image to speech/language:
Recognizing a image representing a place
Speech/language to image:
Recalling an image represented by a word
18. Modeling cognitive systems with PGMs
and their inference is a general view
Probabilistic modeling and
inference
(Multimodal unsupervised
learning)
Supervised learning
Reinforcement
learning
Unsupervised
learning
Thomas Bayes
Multimodal UL: 𝑃(𝑥, 𝑦)
x
z
y w
Obs.
Lat.
Posterior distribution can be calculated.
𝑃 𝑦 𝑥 =
𝑃 𝑥, 𝑦
∑ 𝑃(𝑥, 𝑦)
𝑥
Agent
(robot/human)
x z y
SL: 𝑃(𝑦|𝑥)
Obs. Obs.
Lat.
Intelligence
Input Output
Control as Inference
Many modern RL algorithms can be
considered as a probabilistic
inference of future action
sequence
Inference
19. Spatial Concept-Based Navigation with Human Speech
Instructions via Probabilistic Inference [Taniguchi+ 20]
Inference
Control as Probabilistic Inference (CoI)
We can reformulate reinforcement
learning as Bayesian inference on
(PO)MDP.[Levince 18, Okada 20]
SpCoNavi [Taniguchi+ 20]
In the same way as CoI, Bayesian
inference on the PGM of SoCoSLAM can
perform path planning/navigation.
Sergey Levine, Reinforcement Learning and Control as Probabilistic Inference: Tutorial and Review,
arXiv:1805.00909 [cs.LG], 2 May, (2018) https://arxiv.org/abs/1805.00909
Masashi Okada, Norio Kosaka, Tadahiro Taniguchi, PlaNet of the Bayesians: Reconsidering and Improving
Deep Planning Network by Incorporating Bayesian Inference, IROS2020
Akira Taniguchi, Yoshinobu Hagiwara, Tadahiro Taniguchi, Tetsunari Inamura, Spatial Concept-Based
Navigation with Human Speech Instructions via Probabilistic Inference on Bayesian Generative Model,
Advanced Robotics, 34(19), pp.1213-1228, 2020. DOI: 10.1080/01691864.2020.1817777
Go to "kitchen"
Discovered
lexical knowledge
Spatial concept
Planning as "inference"
20. SERKET: An Architecture for Connecting Stochastic Models
to Realize a Large-Scale Cognitive Model [Nakamura+ 18]
Nakamura T, Nagai T and Taniguchi T, SERKET: An Architecture for Connecting Stochastic Models to Realize a
Large-Scale Cognitive Model. Front. Neurorobot. 12:25. (2018) doi: 10.3389/fnbot.2018.00025
Taniguchi, T., Nakamura, T., Suzuki, M. et al. Neuro-SERKET: Development of Integrative Cognitive System
Through the Composition of Deep Probabilistic Generative Models. New Gener. Comput. 38, 23–48 (2020).
https://doi.org/10.1007/s00354-019-00084-w
Connecting cognitive modules developed as probabilistic generative models and
letting them work together as a single unsupervised learning system.
Having inter-module communication of probabilistic information and
guaranteeing theoretical consistency to some extent.
Neuro-SERKET supports deep generative models, i.e., VAE, as well.
21. Taniguchi, T., Nakamura, T., Suzuki, M. et al. Neuro-SERKET: Development of Integrative Cognitive System
Through the Composition of Deep Probabilistic Generative Models. New Gener. Comput. 38, 23–48 (2020).
https://doi.org/10.1007/s00354-019-00084-w
Example: unsupervised categorization of image and speech
[Taniguchi+ 2020]
26. Extension of BRA-driven development [Yamakawa+ 2021]
Yamakawa, Hiroshi. "The whole brain architecture approach: Accelerating the development of
artificial general intelligence by referring to the brain." Neural Networks 144 (2021): 478-495.
Taniguchi, Akira, Ayako Fukawa, and Hiroshi Yamakawa. "Hippocampal formation-inspired
probabilistic generative model." Neural Networks 151 (2022): 317-335.
Brain
architecture
PGM
27. Achievements of WB-PGM-based approach
Taniguchi, Akira, Ayako Fukawa, and Hiroshi Yamakawa. "Hippocampal formation-inspired probabilistic generative model." Neural
Networks (2022).
A. Taniguchi, M. Muro, H. Yamakawa, T. Taniguchi, Brain-inspired Probabilistic Generative Model for Double Articulation Analysis
of Spoken Language, ICDL 2022
28. (Example) Current prototype of WB-PGM
[Miyazawa+ 2019]
Miyazawa, K., Aoki, T., Horii, T., & Nagai, T. (2019). Integration of multiple generative modules for robot
learning. In Workshop on deep probabilistic generative models for cognitive architecture in robotics.
Miyazawa K, Horii T, Aoki T, Nagai T. Integrated Cognitive Architecture for Robot Learning of Action and
Language. Front Robot AI. 2019;6:131. Published 2019 Nov 29. doi:10.3389/frobt.2019.00131
29. 確率的生成モデルによる世界モデルの学習と
クロスモーダル推論による知能の表現
(World model learning and inference)
29
Sign
Agent 1 Agent 2 Agent K
Communication
(Sampling and acceptance)
Inference
Representation
learning
Collective representation learning
Internal
representation
Observation
Decomposition
Inference
Inference
Communication is Probabilistic Inference
Action/Planning is Probabilistic Inference
Perception/Categorization
is Probabilistic Inference
Intelligence as
"probabilistic inference"
Friston, K., Moran, R. J., Nagai, Y., Taniguchi, T., Gomi, H., & Tenenbaum, J. (2021).
World model learning and inference. Neural Networks, 144, 573-590.
36. T. Tangiuchi, D. Mochihashi, T. Nagai, S. Uchida, N. Inoue, I. Kobayashi, T. Nakamura, Y. Hagiwara, N. Iwahashi & T. Inamura,
Survey on frontiers of language and robotics, Advanced Robotics, 33(15-16), 700-730, 2019. DOI: 10.1080/01691864.2019.1632223
Survey on Frontiers of Language and Robotics
Vision Language Navigation (VLN) を始め
としたEmbodied AI*関連の研究が盛り上
がっているが、実世界での意味作用を光学
的に議論するLanguage and Robotics(言語
とロボティクス)の研究が言語学と人工知
能にとって重要なフロンティアとなる。
*注:Embodied AIといいつつ実世界の身体はない
37. Symbol Emergence in Cognitive
Developmental Systems: a Survey
[Taniguchi+ 2018]
Tadahiro Taniguchi, Emre Ugur, Matej Hoffmann, Lorenzo Jamone, Takayuki Nagai, Benjamin Rosman, Toshihiko Matsuka, Naoto
Iwahashi, Erhan Oztop, Justus Piater, Florentin Wörgötter, Symbol Emergence in Cognitive Developmental Systems: a Survey,
IEEE Transactions on Cognitive and Developmental Systems, 2018.
Shonan Meeting: NO.092: Cognitive Development and
Symbol Emergence in Humans and Robots, Shonan Village
Center, October 3 - 7, 2016