20171207 domain-adaptation

Dec 7, 20172 likes675 views

This document introduces a study that uses simulated data and domain adaptation techniques to efficiently train policies for robotic grasping. The study trains a neural network policy for grasping using large amounts of simulated data. It then adapts the features and visual appearance of the simulated data to match the real world domain using domain adversarial training and a GAN. In experiments, it finds that domain adaptation improves the grasping policy trained on simulated data, leading to more successful real world grasps.

Introduction of a paper by Bousmalis et al.
for efficient training of grasping by simulation
ROS Japan UG #19 機械学習?AI勉強会
7th December 2017
1

Confidential
センスタイムジャパンAbout me
? Name: Taku Yoshioka
? Interests: Bayesian inference, machine learning,
deep learning and robotics
? Robot and ROS: 9 months
? Affiliation: SenseTime Japan
§ Computer vision and deep learning
§ https://www.sensetime.jp
§ https://blog.sensetime.jp (lunch blog)
§ Kyoto, Tokyo
§ We are hiring!
2

Confidential
センスタイムジャパンPrevious study
3
? Learn visual servoing policy with a large-scale data
obtained with 14 manipulators
? Train CNN for evaluation of success probability of
grasping (target)

Confidential
センスタイムジャパンIdea of the study
? Using simulated environment to learn visual servoing
policy for grasping in real environment
? Make simulated features close to real ones by domain
adaptation (DA) on:
§ Feature vector in the evaluation network
§ Visual appearance of the simulated environment
5
Feature vector
Visual appearance

Confidential
センスタイムジャパン
? Fitting distributions based on random samples, without
likelihood function (implicit learning)
? Adversarial training (e.g., GAN for generative model)
6
Fitting distributions

Confidential
センスタイムジャパンDA in feature space
? Domain-adversarial neural net (DANN) [3]
7

Confidential
センスタイムジャパンDA on visual appearance
8

Confidential
センスタイムジャパンDA on visual appearance
9
? GraspGAN: GAN + Label + Semantics (pixel level)
GAN Label Semantics

Confidential
センスタイムジャパンExperiment
10
? 8M simulated data + 94K?9.4M real data
? Sim-only, Real-only, Sim + Real
? Na?ve mix, domain-specific batch normalization (DBN) +
visual randomization, DBN + DANN, DBN + DANN +
visual randomization, GraspGAN

Confidential
センスタイムジャパンResults
11
DANN(-R): Feature DA
GraspGAN: Feature DA + Visual DA

Confidential
センスタイムジャパン
12
? Findings:
§ Efficient training by simulated data
§ Improvement of policy by visual appearance DA
? Future study
§ Consider physical discrepancy
§ Other sensor (e.g., depth)
§ Further improvement of success rate of the task

Confidential
センスタイムジャパンReferences
13
[1] Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M.,
Kalakrishnan, M., ... & Levine, S. (2017). Using simulation and
domain adaptation to improve efficiency of deep robotic
grasping. arXiv preprint arXiv:1709.07857.
[2] Levine, S., Pastor, P., Krizhevsky, A., Ibarz, J., & Quillen, D.
(2016). Learning hand-eye coordination for robotic grasping with
deep learning and large-scale data collection. The International
Journal of Robotics Research, 0278364917710318.
[3] Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H.,
Laviolette, F., ... & Lempitsky, V. (2016). Domain-adversarial training
of neural networks. Journal of Machine Learning Research, 17(59),
1-35.

The document discusses probabilistic programming in robotics. It introduces probabilistic programming and recent advances in probabilistic inference techniques. It then provides an example of using PyMC3 to perform SLAM by formulating the SLAM problem probabilistically and using variational inference to estimate unknown variables like landmark and robot locations. The summary concludes by mentioning some technical issues for applying probabilistic programming to real robotics.

[GTC 2019] Bringing Personal Robots Home: Integrating Computer Vision and Hum...Preferred Networks

The document discusses requirements and technologies for personal robots. It describes two projects from Preferred Networks - an interactive picking robot that can be controlled by spoken language instructions, and an autonomous tidying robot. For picking, they achieved 92.7% accuracy in matching objects to instructions through an interactive clarification approach. For tidying, their system can tidy at a rate of 1.9 objects per minute with a 90% grasp success rate using computer vision and a tablet user interface. Remaining challenges include handling unlimited items and unseen environments.

課題ドリブン、フルスタックAI開発術 [MOBILITY:dev]DeNA

础滨の精度向上において、学习データやモデルの出力の可视化や解析は非常に重要です。我々は、これまで手动で行っていたそれらのプロセスをマイクロサービスに置き换え自动化することで大幅な効率化を実现しました。本セッションでは、そうしたフルスタックな础滨开発术についてお话しします。

PFNにおける研究開発（2022/10/19 東大大学院「融合情報学特別講義Ⅲ」）Preferred Networks

IRJET- Object Detection in Real Time using AI and Deep LearningIRJET Journal

This document summarizes research on object detection techniques using AI and deep learning. It discusses how object detection can be used to enhance e-commerce by recommending products seen in videos. The document reviews several existing object detection algorithms and methods, including YOLO and topological maps. It also identifies limitations in existing systems for multi-class detection efficiency and handling similar backgrounds. The researchers propose using object detection with video input for product recommendations and improving current systems.

Template_for_Presentation_january25 (1).pptarchanadeiva

MILA DL & RL summer school highlights Natalia Díaz Rodríguez

Introduction to Matsuo-Iwasawa Lab (HP)Matsuo Lab

Working With ImageVicky Kumar

Introduction to Matsuo Laboratory (ENG).pptxMatsuo Lab

Matsuo Lab is comprised of over 200 members including researchers, staff, students, and teaching assistants affiliated with the Graduate School of Engineering at the University of Tokyo. The lab specializes in artificial intelligence and web engineering research with a focus on deep learning and its applications in areas like computer vision, natural language processing, and robotics. The lab aims to both conduct fundamental AI research and contribute to society by supporting the creation of startups through entrepreneurship education and business incubation programs.

IRJET- Object Detection in an Image using Deep LearningIRJET Journal

The document summarizes object detection in images using deep learning. It introduces common object detection methods like convolutional neural networks (CNNs) and regional-based CNNs. CNNs are effective for object detection as they can automatically learn distinguishing features without needing manually defined features. The document then describes the methodology which uses a CNN with layers like convolution, ReLU, pooling and fully connected layers to perform feature extraction and classification. It concludes that CNNs provide an efficient method for real-time object detection and segmentation in images through deep learning.

A proposal of XBNaoka MISAWA

IoT-Fundamentals-And-Digital-Tranformation-Repaired.pptxAurelia JQ

The document provides an overview of an emerging technologies curriculum focused on IoT fundamentals and digital transformation. It discusses issues like massive youth unemployment and growing skills shortages as context for the opportunities around digital transformation and IoT. The curriculum aims to develop skills like critical thinking, collaboration and communication through hands-on learning activities using tools like prototyping labs and Packet Tracer. It also outlines several courses that make up the curriculum, including courses on connecting things, big data and analytics, and running a hackathon. The goal is to provide a foundation for various IoT-related career paths.

CLEAR: A Fully User-side Image Search Systemjoisino

This document describes CLEAR, a fully user-side image search system developed by Ryoma Sato at Kyoto University. CLEAR allows users to build and publish their own image search engines without backend servers by formulating image search as a multi-armed bandit problem and implementing the system using only JavaScript on the client-side. This overcomes limitations of traditional search engines which require extensive resources to operate. CLEAR demonstrates that ordinary users can now develop customized image search tools.

Visualizing the Importance of Floor-Plan Image Features in Rent-Prediction Mo...Okamoto Laboratory, The University of Electro-Communications

A Smart Target Detection System using Fuzzy Logic and Background SubtractionIRJET Journal

The document describes a smart target detection system that uses fuzzy logic and background subtraction to detect targets in static videos. It first separates the background from input video frames using a fuzzy logic model. It then identifies moving objects and determines if they are cars or humans based on aspect ratio. The system was able to accurately detect and identify cars and humans as targets by applying fuzzy logic to background modeling and subtraction. Some limitations included lack of detection if there is low contrast between the target and background.

Dream2Control paper reviewtaeseon ryu

The document describes the Dreamer model for reinforcement learning. Dreamer learns a world model from images of its experience. It then learns behaviors by imagining future sequences predicted by the world model and backpropagating value gradients through the imagined sequences. Experiments show Dreamer outperforms prior model-free and model-based methods on a variety of visual control tasks, demonstrating it can learn behaviors purely from latent imagination to solve challenging problems.

PMF_AAT.pptxssusera3f83d

Google Chrome and Mozilla Firefox have different business models. Chrome is developed by Google and funded through advertising revenue, while Firefox is developed by the non-profit Mozilla Foundation and funded through donations and commercial partnerships. Both browsers are open-source and free to users. According to market share statistics, Chrome has the largest share while Firefox's share has declined in recent years.

International pbl conf_5b-c_kizakiSatoru Kizaki

The document summarizes a project-based learning activity on collaborative software development using an agile process in a distributed environment. Students worked in teams over 4 months in 2010 and 2011 to develop software projects. In 2011, improvements were made such as introducing ticket-driven development, participation of industry professionals, and agile coaching to address challenges in 2010 related to project manager workload and student motivation. While some goals were achieved, maintaining student motivation and balancing work assignments remained problems.

The power of deep learning models applicationsSameera Sk

I have performed a detailed analysis of recent advances in deep-learning based research efforts applied in the domains of Image Processing, Audio Processing and Video Processing. I have identified 25 relevant papers and 17 web resources, examining the particular area and problem they address, models employed, datasets used and the images from respective papers are reproduced with proper citation. i hope this information is useful who are doing research in this area.

pdsa new ppt for subject marking and pytssuser72b8e8

Context Aware Recommendations at NetflixLinas Baltrunas

At Netflix, we try to provide the best personalized video recommendations to our members. To do this, we need to adapt our recommendations for each contextual situation, which depends on information such as time or device. In this talk, I will describe how state of the art Contextual Recommendations are used at Netflix. A first example of contextual adaptation is the model that powers the Continue Watching row. It uses a feature-based approach with a carefully constructed training set to learn how to adapt to the context of the member. Next, I will dive into more modern approaches such as Tensor Factorization and LSTMs and share some results from deployments of these methods. I will highlight lessons learned and some common pitfalls of using these powerful methods in industrial scale systems. Finally, I will touch upon system reliability, choice of optimization metrics, hidden costs, risks and benefits of using highly adaptive systems.

Week1- Introduction.pptxfahmi324663

This document provides an agenda and overview for a deep learning course. The agenda includes an introduction to program and course learning outcomes, the syllabus, class management tools, and an introduction to week 1 of deep learning. The syllabus outlines 15 weekly topics on deep learning concepts and algorithms. Example student projects are provided showing applications of deep learning to areas like computer vision, natural language processing, and games. The introduction to week 1 discusses artificial intelligence, machine learning, and deep learning definitions and provides an overview of programming assignments and deep learning in action.

JupyterCon 2017 - Collaboration and automated operation as literate computing...No Bu

Jupyter is useful for DevOps. It enables collaboration between experts and novices to accumulate infrastructure knowledge, while automation via notebooks enhances traceability and reproducibility. Yoshi Nobu Masatani shows how to combine Jupyter with Ansible for reproducible infrastructure and explores knowledge, workflow, and customer support as literate computing practices. Session type: Session Topics: Usage and application

How To Think Like A ProgrammerKassandra Keeton

3D Environment : HomeNavigationYechan(Paul) Kim

202212APSEC.pptx.pdfHiroshi Maruyama

Machine Learning Systems Engineering (MLSE) is a collective effort started 5 years ago in Japan to address challenges in developing and deploying machine learning systems. Key activities included panel discussions at conferences to raise awareness among software engineers, workshops identifying gaps between ML and software engineering practices, and forming a special interest group to organize further discussions. Working groups studied challenges such as fairness, infrastructure, and development processes. International collaborations helped spread ideas to other countries. Research projects explored techniques for requirements engineering, testing, debugging and assuring quality in machine learning systems to develop the new field of machine learning systems engineering. Guidelines and books were also created to establish best practices.

Towards an Agile approach to building application profilesPaul Walk

The document discusses taking an Agile approach to developing application profiles. It involves bringing potential users into the process early through techniques like prototyping and user testing. The goal is to iteratively develop application profiles based on evidence from user engagement to complement the Singapore Framework. By testing and receiving feedback early, costly mistakes can be avoided and user requirements can evolve with changing needs. Tools like MrVobi support collecting user input through activities like free listing, card sorting, and storyboarding to develop a shared understanding and iteratively refine domain models and functional requirements.

20191123 bayes dl-jpTaku Yoshioka

This document summarizes key aspects of variational autoencoders (VAEs): VAE is a generative model that learns a latent representation of data. It approximates the intractable posterior using an encoder network and maximizes a variational lower bound. Semi-supervised VAE models can incorporate unlabeled data by learning shared representations. VAEs have been extended for recurrent sequences, convolutional structures, disentangled representations, and multi-modal data. Importance weighted autoencoders provide a tighter evidence lower bound than standard VAEs.

20191026 bayes dlTaku Yoshioka

This document discusses variational Bayes methods for deep learning including: 1. Variational Bayes approximates the posterior distribution p(Z|X) with a variational distribution q(Z;ξ) by minimizing the Kullback-Leibler divergence between them. 2. For linear dimensionality reduction, it presents a model with observation model p(X|Z,W), priors p(Z) and p(W), and approximates the posterior p(Z,W|X) with variational posterior q(Z)q(W). 3. It derives an objective function L to maximize the evidence lower bound, which involves the expected log likelihood and KL divergences of the vari

More Related Content

Similar to 20171207 domain-adaptation (20)

Working With ImageVicky Kumar

Introduction to Matsuo Laboratory (ENG).pptxMatsuo Lab

IRJET- Object Detection in an Image using Deep LearningIRJET Journal

A proposal of XBNaoka MISAWA

IoT-Fundamentals-And-Digital-Tranformation-Repaired.pptxAurelia JQ

CLEAR: A Fully User-side Image Search Systemjoisino

Visualizing the Importance of Floor-Plan Image Features in Rent-Prediction Mo...Okamoto Laboratory, The University of Electro-Communications

A Smart Target Detection System using Fuzzy Logic and Background SubtractionIRJET Journal

Dream2Control paper reviewtaeseon ryu

PMF_AAT.pptxssusera3f83d

International pbl conf_5b-c_kizakiSatoru Kizaki

The power of deep learning models applicationsSameera Sk

pdsa new ppt for subject marking and pytssuser72b8e8

Context Aware Recommendations at NetflixLinas Baltrunas

Week1- Introduction.pptxfahmi324663

JupyterCon 2017 - Collaboration and automated operation as literate computing...No Bu

How To Think Like A ProgrammerKassandra Keeton

3D Environment : HomeNavigationYechan(Paul) Kim

202212APSEC.pptx.pdfHiroshi Maruyama

Towards an Agile approach to building application profilesPaul Walk

Working With ImageVicky Kumar

Introduction to Matsuo Laboratory (ENG).pptxMatsuo Lab

IRJET- Object Detection in an Image using Deep LearningIRJET Journal

A proposal of XBNaoka MISAWA

IoT-Fundamentals-And-Digital-Tranformation-Repaired.pptxAurelia JQ

CLEAR: A Fully User-side Image Search Systemjoisino

Visualizing the Importance of Floor-Plan Image Features in Rent-Prediction Mo...Okamoto Laboratory, The University of Electro-Communications

A Smart Target Detection System using Fuzzy Logic and Background SubtractionIRJET Journal

Dream2Control paper reviewtaeseon ryu

PMF_AAT.pptxssusera3f83d

International pbl conf_5b-c_kizakiSatoru Kizaki

The power of deep learning models applicationsSameera Sk

pdsa new ppt for subject marking and pytssuser72b8e8

Context Aware Recommendations at NetflixLinas Baltrunas

Week1- Introduction.pptxfahmi324663

JupyterCon 2017 - Collaboration and automated operation as literate computing...No Bu

How To Think Like A ProgrammerKassandra Keeton

3D Environment : HomeNavigationYechan(Paul) Kim

202212APSEC.pptx.pdfHiroshi Maruyama

Towards an Agile approach to building application profilesPaul Walk

More from Taku Yoshioka (8)

20191123 bayes dl-jpTaku Yoshioka

20191026 bayes dlTaku Yoshioka

20191019 sinkhornTaku Yoshioka

20181221 q-traderTaku Yoshioka

This document describes using reinforcement learning for stock trading with Python. It discusses implementing Q-learning to learn the optimal trading policy, using a deep Q-network to represent the action-value function. The trading environment defines the state, actions, and rewards. States represent the stock price window, actions are buy, sell, or sit, and rewards incentivize profit. The agent trains on experience samples to learn the optimal policy for making trading decisions.

20181125 pybulletTaku Yoshioka

This document provides an overview of reinforcement learning concepts and frameworks. It discusses Markov decision processes, value functions, reinforcement learning algorithms like Q-learning and policy gradients, and Python modules like OpenAI Gym and Baselines for implementing reinforcement learning. It also includes an example of installing the necessary libraries and running a simple reinforcement learning task on the pendulum environment using PyBullet and TensorFlow.

20180722 pyroTaku Yoshioka

20160611 pymc3-latentTaku Yoshioka

自动微分変分ベイズ法の绍介Taku Yoshioka

20191123 bayes dl-jpTaku Yoshioka

20191026 bayes dlTaku Yoshioka

20191019 sinkhornTaku Yoshioka

20181221 q-traderTaku Yoshioka

20181125 pybulletTaku Yoshioka

20180722 pyroTaku Yoshioka

20160611 pymc3-latentTaku Yoshioka

自动微分変分ベイズ法の绍介Taku Yoshioka

Recently uploaded (20)

ESIT135 Problem Solving Using Python Notes of Unit-2 and Unit-3prasadmutkule1

AO Star Algorithm in Artificial Intellligencevipulkondekar

INTERNET OF THINGSSSSSSSSSSSSSSSSSSSSSSSSS.pptxbmit1

How to Build a Speed Sensor using Arduino?CircuitDigest

Instruction execution cycle _SwatiHans10

Design and Analysis of Algorithms Unit 5sureshkumara29

Algorithm design techniques include: Brute Force Greedy Algorithms Divide-and-Conquer Dynamic Programming Reduction / Transform-and-Conquer Backtracking and Branch-and-Bound Randomization Approximation Recursive Approach What is an algorithm? An Algorithm is a procedure to solve a particular problem in a finite number of steps for a finite-sized input. The algorithms can be classified in various ways. They are: Implementation Method Design Method Design Approaches Other Classifications In this article, the different algorithms in each classification method are discussed. The classification of algorithms is important for several reasons: Organization: Algorithms can be very complex and by classifying them, it becomes easier to organize, understand, and compare different algorithms. Problem Solving: Different problems require different algorithms, and by having a classification, it can help identify the best algorithm for a particular problem. Performance Comparison: By classifying algorithms, it is possible to compare their performance in terms of time and space complexity, making it easier to choose the best algorithm for a particular use case. Reusability: By classifying algorithms, it becomes easier to re-use existing algorithms for similar problems, thereby reducing development time and improving efficiency. Research: Classifying algorithms is essential for research and development in computer science, as it helps to identify new algorithms and improve existing ones. Overall, the classification of algorithms plays a crucial role in computer science and helps to improve the efficiency and effectiveness of solving problems. Classification by Implementation Method: There are primarily three main categories into which an algorithm can be named in this type of classification. They are: Recursion or Iteration: A recursive algorithm is an algorithm which calls itself again and again until a base condition is achieved whereas iterative algorithms use loops and/or data structures like stacks, queues to solve any problem. Every recursive solution can be implemented as an iterative solution and vice versa. Example: The Tower of Hanoi is implemented in a recursive fashion while Stock Span problem is implemented iteratively. Exact or Approximate: Algorithms that are capable of finding an optimal solution for any problem are known as the exact algorithm. For all those problems, where it is not possible to find the most optimized solution, an approximation algorithm is used. Approximate algorithms are the type of algorithms that find the result as an average outcome of sub outcomes to a problem. Example: For NP-Hard Problems, approximation algorithms are used. Sorting algorithms are the exact algorithms. Serial or Parallel or Distributed Algorithms: In serial algorithms, one instruction is executed at a time while parallel algorithms are those in which we divide the problem into subproblems and execute them on different processors.

Common Network Architecture:X.25 Networks, Ethernet (Standard and Fast): fram...SnehPrasad2

Zero-Trust-Architecture-Reimagining-Network-Security.pptxyash98012

Sppu engineering artificial intelligence and data science semester 6th Artif...pawaletrupti434

Scientific_Thinking_and_Method_GE3791.pptxaravym456

Artificial intelligence based solar vehicle.pptxrrabin2

Biases, our brain and software developmentMatias Iacono

Wireless-Charger presentation for seminar .pdfAbhinandanMishra30

Turbocor Product and Technology Review.pdfTotok Sulistiyanto

Unit 1- Review of Basic Concepts-part 1.pptxSujataSonawane11

Defining the Future of Biophilic Design in Crete.pdfARENCOS

ESIT135 Problem Solving Using Python Notes of Unit-3prasadmutkule1

Environmental Product Declaration - Uni BellManishPatel169454

AI ppt on water jug problem by shivam sharmaShivamSharma588604

Designing Flex and Rigid-Flex PCBs to Prevent FailureEpec Engineered Technologies

Flex and rigid-flex printed circuit boards (PCBs) can be considered at the basic level some of the most complex PCBs in the industry. With that in mind, it’s incredibly easy to make a mistake, to leave something out, or to create a design that was doomed from the start. Such design failures can end up leading to an eventual failure by delamination, short circuits, damage to the flex portions, and many other things. The easiest way to circumvent these is to start at the beginning, to design with preventing failure in mind rather than trying to fix existing designs to accommodate for problems. In this webinar, we cover how to design flex and rigid-flex PCBs with failure prevention in mind to save time, money, and headaches, and what failure can look like. For more information on our flex and rigid-flex PCB solutions, visit https://www.epectec.com/flex.

ESIT135 Problem Solving Using Python Notes of Unit-2 and Unit-3prasadmutkule1

AO Star Algorithm in Artificial Intellligencevipulkondekar

INTERNET OF THINGSSSSSSSSSSSSSSSSSSSSSSSSS.pptxbmit1

How to Build a Speed Sensor using Arduino?CircuitDigest

Instruction execution cycle _SwatiHans10

Design and Analysis of Algorithms Unit 5sureshkumara29

Common Network Architecture:X.25 Networks, Ethernet (Standard and Fast): fram...SnehPrasad2

Zero-Trust-Architecture-Reimagining-Network-Security.pptxyash98012

Sppu engineering artificial intelligence and data science semester 6th Artif...pawaletrupti434

Scientific_Thinking_and_Method_GE3791.pptxaravym456

Artificial intelligence based solar vehicle.pptxrrabin2

Biases, our brain and software developmentMatias Iacono

Wireless-Charger presentation for seminar .pdfAbhinandanMishra30

Turbocor Product and Technology Review.pdfTotok Sulistiyanto

Unit 1- Review of Basic Concepts-part 1.pptxSujataSonawane11

Defining the Future of Biophilic Design in Crete.pdfARENCOS

ESIT135 Problem Solving Using Python Notes of Unit-3prasadmutkule1

Environmental Product Declaration - Uni BellManishPatel169454

AI ppt on water jug problem by shivam sharmaShivamSharma588604

Designing Flex and Rigid-Flex PCBs to Prevent FailureEpec Engineered Technologies