"Anime Generation with AI".
- Video: Generated Anime: https://youtu.be/X9j1fwexK2c
- Video: Other AI Solutions for Anime Production Issues: https://youtu.be/Gz90H1M7_u4
The document discusses recent advances in generative adversarial networks (GANs) for image generation. It summarizes two influential GAN models: ProgressiveGAN (Karras et al., 2018) and BigGAN (Brock et al., 2019). ProgressiveGAN introduced progressive growing of GANs to produce high resolution images. BigGAN scaled up GAN training through techniques like large batch sizes and regularization methods to generate high fidelity natural images. The document also discusses using GANs to generate full-body, high-resolution anime characters and adding motion through structure-conditional GANs.
This document discusses deepfakes, including their creation and detection. It begins with an introduction to face swapping, face reenactment, and face synthesis techniques used to generate deepfakes. It then describes several methods for creating deepfakes, such as faceswap algorithms, 3D modeling approaches, and GAN-based methods. The document also reviews several datasets used to detect deepfakes. Finally, it analyzes current research on detecting deepfakes using techniques like two-stream neural networks, analyzing inconsistencies in audio-video, and detecting warping artifacts.
Artificial Intelligence Workshop, Collegio universitario Bertoni, Milano, 20 May 2017.
Audience of the workshop: undergraduate students without neural networks background.
Summary:
- Deep Learning Showcase
- What is deep learning and how it works
- How to start with deep learning
- Live demo: image recognition with Nvidia DIGITS
- Playground
Duration: 2 hours.
Value Iteration Networks is a machine learning method for robot path planning that can operate in new environments not seen during training. It works by predicting optimal actions through learning reward values for each state and propagating rewards to determine the sum of future rewards. The method was shown to be effective for planning in grid maps and continuous control tasks, and was even applied to navigation of Wikipedia links.
InfoGAN: Interpretable Representation Learning by Information Maximizing Gen...Shuhei Yoshida
Unsupervised learning of disentangled representations was the goal. The approach was to use GANs and maximize the mutual information between generated images and input codes. This led to the benefit of obtaining interpretable representations without supervision and at substantial additional costs.
Dual Learning for Machine Translation (NIPS 2016)Toru Fujino
The paper introduces a dual learning algorithm that utilizes monolingual data to improve neural machine translation. The algorithm trains two translation models in both directions simultaneously. Experimental results show that when trained with only 10% of parallel data, the dual learning model achieves comparable results to baseline models trained on 100% of data. The dual learning mechanism also outperforms baselines when trained on full data and can help address the lack of large parallel corpora.
Fast and Probvably Seedings for k-MeansKimikazu Kato
The document proposes a new MCMC-based algorithm for initializing centroids in k-means clustering that does not assume a specific distribution of the input data, unlike previous work. It uses rejection sampling to emulate the distribution and select initial centroids that are widely scattered. The algorithm is proven mathematically to converge. Experimental results on synthetic and real-world datasets show it performs well with a good trade-off of accuracy and speed compared to existing techniques.
Interaction Networks for Learning about Objects, Relations and PhysicsKen Kuroki
For my presentation for a reading group. I have not in any way contributed this study, which is done by the researchers named on the first slide.
https://papers.nips.cc/paper/6418-interaction-networks-for-learning-about-objects-relations-and-physics
Introduction of "TrailBlazer" algorithmKatsuki Ohto
論文「Blazing the trails before beating the path: Sample-efficient Monte-Carlo planning」紹介スライドです。NIPS2016読み会@PFN(2017/1/19) https://connpass.com/event/47580/ にて。
Conditional Image Generation with PixelCNN Decoderssuga93
The document summarizes research on conditional image generation using PixelCNN decoders. It discusses how PixelCNNs sequentially predict pixel values rather than the whole image at once. Previous work used PixelRNNs, but these were slow to train. The proposed approach uses a Gated PixelCNN that removes blind spots in the receptive field by combining horizontal and vertical feature maps. It also conditions PixelCNN layers on class labels or embeddings to generate conditional images. Experimental results show the Gated PixelCNN outperforms PixelCNN and achieves performance close to PixelRNN on CIFAR-10 and ImageNet, while training faster. It can also generate portraits conditioned on embeddings of people.
Safe and Efficient Off-Policy Reinforcement Learningmooopan
This document summarizes the Retrace(λ) reinforcement learning algorithm presented by Remi Munos, Thomas Stepleton, Anna Harutyunyan and Marc G. Bellemare. Retrace(λ) is an off-policy multi-step reinforcement learning algorithm that is safe (converges for any policy), efficient (makes best use of samples when policies are close), and has lower variance than importance sampling. Empirical results on Atari 2600 games show Retrace(λ) outperforms one-step Q-learning and existing multi-step methods.
Introduction of “Fairness in Learning: Classic and Contextual Bandits”Kazuto Fukuchi
1. The document discusses fairness constraints in contextual bandit problems and classic bandit problems.
2. It shows that for classic bandits, Θ(k^3) rounds are necessary and sufficient to achieve a non-trivial regret under fairness constraints.
3. For contextual bandits, it establishes a tight relationship between achieving fairness and Knows What it Knows (KWIK) learning, where KWIK learnability implies the existence of fair learning algorithms.
Improving Variational Inference with Inverse Autoregressive FlowTatsuya Shirakawa
This slide was created for NIPS 2016 study meetup.
IAF and other related researches are briefly explained.
paper:
Diederik P. Kingma et al., "Improving Variational Inference with Inverse Autoregressive Flow", 2016
https://papers.nips.cc/paper/6581-improving-variational-autoencoders-with-inverse-autoregressive-flow
The document summarizes the paper "Matching Networks for One Shot Learning". It discusses one-shot learning, where a classifier can learn new concepts from only one or a few examples. It introduces matching networks, a new approach that trains an end-to-end nearest neighbor classifier for one-shot learning tasks. The matching networks architecture uses an attention mechanism to compare a test example to a small support set and achieve state-of-the-art one-shot accuracy on Omniglot and other datasets. The document provides background on one-shot learning challenges and related work on siamese networks, memory augmented neural networks, and attention mechanisms.
Deep Learning - The Past, Present and Future of Artificial IntelligenceLukas Masuch
The document provides an overview of deep learning, including its history, key concepts, applications, and recent advances. It discusses the evolution of deep learning techniques like convolutional neural networks, recurrent neural networks, generative adversarial networks, and their applications in computer vision, natural language processing, and games. Examples include deep learning for image recognition, generation, segmentation, captioning, and more.
This document provides a tutorial on Generative Adversarial Networks (GANs). It introduces GANs as having two networks, a generator and discriminator, where the generator generates fake data from random noise to fool the discriminator, and the discriminator learns to distinguish real from fake data. Several types of GANs are then briefly outlined, including standard GANs, Wasserstein GANs, and applications of GANs.
[DL輪読会]Neural Radiance Flow for 4D View Synthesis and Video Processing (NeRF...Deep Learning JP
Neural Radiance Flow (NeRFlow) is a method that extends Neural Radiance Fields (NeRF) to model dynamic scenes from video data. NeRFlow simultaneously learns two fields - a radiance field to reconstruct images like NeRF, and a flow field to model how points in space move over time using optical flow. This allows it to generate novel views from a new time point. The model is trained end-to-end by minimizing losses for color reconstruction from volume rendering and optical flow reconstruction. However, the method requires training separate models for each scene and does not generalize to unknown scenes.
Value Iteration Networks is a machine learning method for robot path planning that can operate in new environments not seen during training. It works by predicting optimal actions through learning reward values for each state and propagating rewards to determine the sum of future rewards. The method was shown to be effective for planning in grid maps and continuous control tasks, and was even applied to navigation of Wikipedia links.
InfoGAN: Interpretable Representation Learning by Information Maximizing Gen...Shuhei Yoshida
Unsupervised learning of disentangled representations was the goal. The approach was to use GANs and maximize the mutual information between generated images and input codes. This led to the benefit of obtaining interpretable representations without supervision and at substantial additional costs.
Dual Learning for Machine Translation (NIPS 2016)Toru Fujino
The paper introduces a dual learning algorithm that utilizes monolingual data to improve neural machine translation. The algorithm trains two translation models in both directions simultaneously. Experimental results show that when trained with only 10% of parallel data, the dual learning model achieves comparable results to baseline models trained on 100% of data. The dual learning mechanism also outperforms baselines when trained on full data and can help address the lack of large parallel corpora.
Fast and Probvably Seedings for k-MeansKimikazu Kato
The document proposes a new MCMC-based algorithm for initializing centroids in k-means clustering that does not assume a specific distribution of the input data, unlike previous work. It uses rejection sampling to emulate the distribution and select initial centroids that are widely scattered. The algorithm is proven mathematically to converge. Experimental results on synthetic and real-world datasets show it performs well with a good trade-off of accuracy and speed compared to existing techniques.
Interaction Networks for Learning about Objects, Relations and PhysicsKen Kuroki
For my presentation for a reading group. I have not in any way contributed this study, which is done by the researchers named on the first slide.
https://papers.nips.cc/paper/6418-interaction-networks-for-learning-about-objects-relations-and-physics
Introduction of "TrailBlazer" algorithmKatsuki Ohto
論文「Blazing the trails before beating the path: Sample-efficient Monte-Carlo planning」紹介スライドです。NIPS2016読み会@PFN(2017/1/19) https://connpass.com/event/47580/ にて。
Conditional Image Generation with PixelCNN Decoderssuga93
The document summarizes research on conditional image generation using PixelCNN decoders. It discusses how PixelCNNs sequentially predict pixel values rather than the whole image at once. Previous work used PixelRNNs, but these were slow to train. The proposed approach uses a Gated PixelCNN that removes blind spots in the receptive field by combining horizontal and vertical feature maps. It also conditions PixelCNN layers on class labels or embeddings to generate conditional images. Experimental results show the Gated PixelCNN outperforms PixelCNN and achieves performance close to PixelRNN on CIFAR-10 and ImageNet, while training faster. It can also generate portraits conditioned on embeddings of people.
Safe and Efficient Off-Policy Reinforcement Learningmooopan
This document summarizes the Retrace(λ) reinforcement learning algorithm presented by Remi Munos, Thomas Stepleton, Anna Harutyunyan and Marc G. Bellemare. Retrace(λ) is an off-policy multi-step reinforcement learning algorithm that is safe (converges for any policy), efficient (makes best use of samples when policies are close), and has lower variance than importance sampling. Empirical results on Atari 2600 games show Retrace(λ) outperforms one-step Q-learning and existing multi-step methods.
Introduction of “Fairness in Learning: Classic and Contextual Bandits”Kazuto Fukuchi
1. The document discusses fairness constraints in contextual bandit problems and classic bandit problems.
2. It shows that for classic bandits, Θ(k^3) rounds are necessary and sufficient to achieve a non-trivial regret under fairness constraints.
3. For contextual bandits, it establishes a tight relationship between achieving fairness and Knows What it Knows (KWIK) learning, where KWIK learnability implies the existence of fair learning algorithms.
Improving Variational Inference with Inverse Autoregressive FlowTatsuya Shirakawa
This slide was created for NIPS 2016 study meetup.
IAF and other related researches are briefly explained.
paper:
Diederik P. Kingma et al., "Improving Variational Inference with Inverse Autoregressive Flow", 2016
https://papers.nips.cc/paper/6581-improving-variational-autoencoders-with-inverse-autoregressive-flow
The document summarizes the paper "Matching Networks for One Shot Learning". It discusses one-shot learning, where a classifier can learn new concepts from only one or a few examples. It introduces matching networks, a new approach that trains an end-to-end nearest neighbor classifier for one-shot learning tasks. The matching networks architecture uses an attention mechanism to compare a test example to a small support set and achieve state-of-the-art one-shot accuracy on Omniglot and other datasets. The document provides background on one-shot learning challenges and related work on siamese networks, memory augmented neural networks, and attention mechanisms.
Deep Learning - The Past, Present and Future of Artificial IntelligenceLukas Masuch
The document provides an overview of deep learning, including its history, key concepts, applications, and recent advances. It discusses the evolution of deep learning techniques like convolutional neural networks, recurrent neural networks, generative adversarial networks, and their applications in computer vision, natural language processing, and games. Examples include deep learning for image recognition, generation, segmentation, captioning, and more.
This document provides a tutorial on Generative Adversarial Networks (GANs). It introduces GANs as having two networks, a generator and discriminator, where the generator generates fake data from random noise to fool the discriminator, and the discriminator learns to distinguish real from fake data. Several types of GANs are then briefly outlined, including standard GANs, Wasserstein GANs, and applications of GANs.
[DL輪読会]Neural Radiance Flow for 4D View Synthesis and Video Processing (NeRF...Deep Learning JP
Neural Radiance Flow (NeRFlow) is a method that extends Neural Radiance Fields (NeRF) to model dynamic scenes from video data. NeRFlow simultaneously learns two fields - a radiance field to reconstruct images like NeRF, and a flow field to model how points in space move over time using optical flow. This allows it to generate novel views from a new time point. The model is trained end-to-end by minimizing losses for color reconstruction from volume rendering and optical flow reconstruction. However, the method requires training separate models for each scene and does not generalize to unknown scenes.
The document discusses generative adversarial networks (GANs) and several variants of GANs including:
- Deep convolutional GANs (DCGANs) which use convolutional neural networks for the generator and discriminator.
- Wasserstein GANs which address the problem of mode collapse.
- Conditional GANs which generate images conditioned on additional information like text.
- CycleGANs which are used for tasks like image-to-image translation without paired training examples.
This document discusses generative adversarial networks (GANs) and their applications. It first introduces GANs, describing the generator, discriminator, and objective function. It then discusses various improvements to GANs such as Wasserstein GANs and mode regularization. The document also summarizes applications of GANs to image-to-image translation using CycleGAN and StarGAN, as well as applications to video, speech synthesis, and voice conversion.
A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)Thomas da Silva Paula
A basic introduction to Generative Adversarial Networks, what they are, how they work, and why study them. This presentation shows what is their contribution to Machine Learning field and for which reason they have been considered one of the major breakthroughts in Machine Learning field.
【DL輪読会】Toward Fast and Stabilized GAN Training for Highfidelity Few-shot Imag...Deep Learning JP
1) The document proposes a lightweight GAN architecture with a skip-layer channel-wise excitation module and self-supervised discriminator for fast and stabilized GAN training for few-shot high-fidelity image synthesis.
2) Experimental results show the proposed method can generate high quality 1024x1024 images using a single GPU within 1 day of training, much faster than baseline StyleGAN2 which requires 5 days of training using 20 GPUs.
3) The skip-layer channel-wise excitation module and self-supervised discriminator help improve the model stability and reduce mode collapse during GAN training.
This document discusses some of the challenges in developing AI systems that utilize machine learning. It notes that machine learning systems rely on probabilities and statistics based on training data, making quality assurance difficult. It is also difficult to fully understand and interpret models from deep neural networks. The document suggests that new approaches are needed for developing machine learning-based systems, as traditional software engineering approaches do not work well. Establishing the field of "machine learning engineering" is important for building AI systems that can reliably ensure quality.
【DL輪読会】Physion: Evaluating Physical Prediction from Vision in Humans and Mach...Deep Learning JP
This document summarizes a research paper that proposes a new dataset called Physion for evaluating how well machine learning models can predict physical interactions from vision, similar to humans. The dataset contains videos of common physical phenomena. Several state-of-the-art models were evaluated on the dataset, including particle-based simulators and vision-based models. Particle-based simulators achieved performance on par with humans, while vision-based models performed poorly. The document provides background on the motivation for the dataset and describes the different models and their approaches.
TensorFlow London: Progressive Growing of GANs for increased stability, quali...Seldon
Speaker: Jakub Langr, R&D Data Scientist at Mudano
Title: Progressing with GANs: Progressive growing for increasing stability and variation
Abstract:
Generative Adversarial Networks (GANs) have recently reached few tremendous milestones: generating full-HD synthetic faces, to image compression better than the state of the art to cryptography. In this talk we will start with the basics of generative models, but eventually, explore the state of the art in generating full HD images as presented in https://arxiv.org/abs/1710.10196
Bio: Jakub Langr graduated from the University of Oxford where he also taught at OU Computing Services. He has worked in data science since 2013, most recently as a Data Science Tech Lead at Filtered.com and as an R&D Data Scientist at Mudano. Jakub is a co-author of GANs in Action by Manning Publications. Jakub also designed and teaches Data Science courses at the University of Birmingham and is a Guest Lecturer at the University of Oxford's course "Data Science for IoT".
Twitter: @langrjakub; jakublangr.com
Jakub Langr (University of Oxford) - Overview of Generative Adversarial Netwo...Codiax
This document provides an overview of Generative Adversarial Networks (GANs) in 3 sections. It begins by briefly discussing supervised and unsupervised machine learning. It then explains that GANs use two neural networks, a generator and discriminator, that compete against each other in a game theoretic setup. The generator learns to produce more realistic samples while the discriminator learns to better distinguish real and fake samples. Popular GAN architectures like CycleGAN and BigGAN are also summarized.
This document summarizes research papers on domain transfer techniques from NIPS 2017 to 2018. It discusses papers that use generative adversarial networks (GANs) for unsupervised image-to-image translation tasks between different domains. Key papers mentioned include Pix2Pix, CycleGAN, DiscoGAN, BicycleGAN, MUNIT, StarGAN, and Fader Networks. The document provides brief descriptions of each paper's proposed method and contributions to domain transfer and image translation.
Generative Adversarial Networks (GANs) use two neural networks, a generator and discriminator, that compete against each other. The generator learns to generate fake images that look real, while the discriminator learns to tell real images apart from fakes. This document discusses various GAN architectures and applications, including conditional GANs, image-to-image translation, style transfer, semantic image editing, and data augmentation using GAN-generated images. It also covers evaluation metrics for GANs and societal impacts such as bias and deepfakes.
The document discusses variational divergence minimization for training generative neural networks using f-GAN. It introduces f-divergence as a generalization of divergence measures used in GANs like KL divergence. F-divergence allows the training of generative models by minimizing the divergence between the generated distribution and real data distribution. The paper presents an algorithm for minimizing f-divergence in generative neural samplers.
Generative Adversarial Networks And Deep Learning Roshani Rautlogerhoyal3q
Generative Adversarial Networks And Deep Learning Roshani Raut
Generative Adversarial Networks And Deep Learning Roshani Raut
Generative Adversarial Networks And Deep Learning Roshani Raut
Generative Adversarial Networks (GANs) are a class of deep learning models that are trained using an adversarial process. GANs consist of two neural networks, a generator and a discriminator, that compete against each other. The generator tries to generate new samples from a latent space to fool the discriminator, while the discriminator tries to distinguish real samples from fake ones. GANs can learn complex high-dimensional distributions and have been applied to image generation, video generation, and other domains. However, training GANs is challenging due to issues like non-convergence and mode collapse. Recent work has explored techniques like minibatch discrimination, conditional GANs, and unrolled GANs to help address these training issues.
Apache Mahout - Random Forests - #TokyoWebmining #8 Koichi Hamada
The document discusses social media, social graphs, personality modeling, data mining, machine learning, and random forests. It references social media, how individuals connect through social graphs, modeling personality objectively, extracting patterns from data through data mining and machine learning techniques, and the random forests algorithm developed by Leo Breiman in 2001.
「樹木モデルとランダムフォレスト(Tree-based Models and Random Forest) -機械学習による分類・予測-」。 Tree-based Model, Random Forest の入門的な内容です。機械学習・データマイニングセミナー 2010/10/07 。 hamadakoichi 濱田晃一
Towards Scientific Foundation Models (Invited Talk)Steffen Staab
Foundation models are machine-learned models that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. Foundation models have been used successfully for question answering and text generation (ChatGPT), image understanding (Clip, VIT), or image generation. Recently, the basic idea underlying foundation models been considered for learning scientific foundation models that capture expectations about partial differential equations. Existing scientific foundation models have still been very much limited wrt. the type of PDEs or differential operators . In this talk, I present some of our recent work on paving the way towards scientific foundation models that aims at making them more robust and better generalisable.
Biological application of spectroscopy.pptxRahulRajai
Spectroscopy in biological studies involves using light or other forms of electromagnetic radiation to analyze the structure, function, and interactions of biological molecules. It helps researchers understand how molecules like proteins, nucleic acids, and lipids behave and interact within cells.
biological applications of spectroscopy:
1. Studying Biological Molecules:
Proteins:
Spectroscopy can reveal protein structure, including folding patterns and interactions with other molecules.
Nucleic Acids:
It helps analyze the structure of DNA and RNA, including their base sequences and interactions.
Lipids:
Spectroscopy can be used to study lipid interactions within cell membranes and their role in cellular processes.
Metabolic Pathways:
Spectroscopy can monitor changes in metabolic processes and cellular signaling pathways, providing insights into how cells function.
Glymphatic system dysfunction and neurodegenerationKanakChaudhary10
Glymphatic system dysfunction and neurodegeneration
and its other factors
Slide 1: Clearing the Brain’s Waste
Slide 2: Anatomy of the Glymphatic System
Slide 3: Role in Neurotoxic Protein Clearance
Slide 4: Glymphatic Dysfunction in Alzheimer’s
Slide 5: Link to Traumatic Brain Injury
Slide 6: Mechanisms of Glymphatic Impairment
Slide 7: Imaging Glymphatic Function
Slide 8: Therapeutic Modulation Strategies
Slide 9: Challenges in Glymphatic Research
Slide 10: Future Directions for Glymphatic Therapies
This presentation provides a concise overview of the human immune system's fundamental response to viral infections. It covers both innate and adaptive immune mechanisms, detailing the roles of physical barriers, interferons, natural killer (NK) cells, antigen-presenting cells (APCs), B cells, and T cells in combating viruses. Designed for students, educators, and anyone interested in immunology, this slide deck simplifies complex biological processes and highlights key steps in viral detection, immune activation, and memory formation. Ideal for classroom use or self-learning.
A review on simple heterocyclics involved in chemical ,biochemical and metabo...DrAparnaYeddala
Heterocyclics play crucial role in the drug discovery process and exhibit various
biological activities. Among aromatic heterocycles, the prevalent moieties are five membered
rings.The role and utility of heterocycles in organic synthesis paved the way to develop
precursors for aminoacids, medicinaldrugs and other chemical componetnts.For an organic
molecule the potency is measured based on its non toxic nature, lower dosage and inhibition
of microbial cellwall growth.
Also for evaluating their potential to be used as drugs, pharmaceuticals, special
chemicals and agrochemicals.
Heterocyclic chemistry credits for nearly thirty percent of contemporary
publications. In fact seventy five percent of organic compounds are heterocyclic compounds.
The alkaloids with nitrogen atoms like ergotamine show antimigraine activity, cinchonine,
and display antimalarial activity. The loaded activity of these compounds was explored by
many researchers in medicinal, insecticidal, pesticidal and naturally occurring aminoacids.
Nucleic acid strands contain heterocylic compounds as major components. Also they display
their major role as central nervous system activators, insecticidal, pesticidal and physiological
processes like antiinflammation activity and antitumor activity.
A review on simple heterocyclics involved in chemical ,biochemical and metabo...DrAparnaYeddala
NIPS 2016 Overview and Deep Learning Topics
1. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
NIPS 2016 読み会
@Preferred Networks
2017/1/19
NIPS 2016
Overview and Deep Learning Topics
@hamadakoichi
濱田晃一
Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
2. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
2
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
4. 4
Copyright (C) 2014 DeNA Co.,Ltd. All Rights Reserved.
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
5. 5
Copyright (C) 2014 DeNA Co.,Ltd. All Rights Reserved.
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
6. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
6
NIPS 2016
・第30回の開催
・期間: 2016年12月5-10日
・ICML 33回に続き長い伝統
・チュートリアル: 5(1日)
・本会議: 5-8(4日)
・ワークショップ: 9-10(2日)
・開催地: バルセロナ(スペイン)
貼る:会場雰囲気
7. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
7
NIPS 2016
参加者が 6000人に増加 (2015年の1.5倍)
※Terrence Sejnowskiは NIPS foundationの President
8. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
8
NIPS Features
・採択の92%はポスター
・採択率: 23%
・投稿数: 2500+、採択数: 568
・Oral(45) : 20分の口頭発表 + ポスター
・Poster(523) : ポスターのみ
・少数トラックでの進行(最大3)
(昨年までシングルトラックだったがパラレルに)
9. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
9
NIPS Features
・ポスター発表による活発な議論
(昨年までの19-24時の5時間ポスターからは時間縮小したが、最後まで活発な議論)
・210 min(3.5 hour)/ day
・130 Poster x 4 days
10. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
10
NIPS2016 Hot Topics
引用元:
The review process for NIPS 2016
http://www.tml.cs.uni-tuebingen.de/team/
luxburg/misc/nips2016/index.php
Deep Learning Computer Vision Large Scale Learning Learning Theory Optimization Sparsity
11. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
11
NIPS2016 Hot Topics
Tutorial 3/9、Symposium 2/3 が Deep Learning
Reinforcement Learning, Generative Adversarial Net, Recurrent Net
Tutorial Symposium
12. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
12
NIPS2016 Hot Topics
Tutorial Symposium
Tutorial 3/9、Symposium 2/3 が Deep Learning
Reinforcement Learning, Generative Adversarial Net, Recurrent Net
上記2トピックに関し、本会議論文をピックアップし概要紹介します
(Reinforcement Learningは、このNIPS読み会での個別論文の発表も多いため)
13. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
13
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
14. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
14
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
15. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
15
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
16. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
16
Generative Adversarial Network (GAN)
Generative Adversarial Nets(GAN)
Goodfellow+, NIPS2014
17. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
17
Generative Adversarial Network (GAN)
Generator(生成器)と Discriminator(識別器)を戦わせ
生成精度を向上させる
識別器: “本物画像”と “生成器が作った偽画像”を識別する
生成器: 生成画像を識別器に“本物画像”と誤識別させようとする
(Goodfellow+, NIPS2014, Deep Learning Workshop, Presentation)
18. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
18
Generative Adversarial Network (GAN)
Minimax Objective function
Discriminator が
「本物画像」を「本物」と識別
(Goodfellow+, NIPS2014, Deep Learning Workshop, Presentation)
Discriminator が
「生成画像」を「偽物」と識別する
Discriminatorは
正しく識別しようとする
(最大化)
Generatorは Discriminator に誤識別させようとする(最小化)
Generator(生成器)と Discriminator(識別器)を戦わせ
生成精度を向上させる
19. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
19
自然画像の表現ベクトル空間学習・演算・画像生成
ICLR16: Deep Convolutional GAN : DCGAN (Radford+)
自然画像のクリアな画像生成 画像演算
Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks.
Alec Radford, Luke Metz, Soumith Chintala.
arXiv:1511.06434. In ICLR 2016.
20. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
20
ICML16: Autoencoding beyond pixels (Larsen+)
Autoencoding beyond pixels using a learned similarity metric.
Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle,
Ole Winther.
arXiv:1512.09300. In ICML 2016.
自然画像の表現ベクトル空間学習・演算・画像生成
21. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
21
ICML16: Generative Adversarial Text to Image Synthesis(Reed+)
Generative Adversarial Text to Image Synthesis.
Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen
Logeswaran, Bernt Schiele, Honglak Lee.
arXiv:1605.05396. In ICML 2016.
文章からの画像生成
文章で条件付したGAN
22. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
22
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
23. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
23
Generative Adversarial Text to Image Synthesis(Reed+)
Learning What and Where to Draw.
Scott Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, Honglak Lee.
arXiv:1610.02454. In NIPS 2016.
文章からの画像生成
表示位置情報も条件付したGAN
24. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
24
InfoGAN (Chen+)
InfoGAN: Interpretable Representation
Learning by Information Maximizing
Generative Adversarial Nets.
Xi Chen, Yan Duan, Rein Houthooft, John
Schulman, Ilya Sutskever, Pieter Abbeel.
arXiv:1606.03657. In NIPS 2016
Latent code c、Generator 出力との Mutual Information を加え
GANで狙って表現ベクトル空間を学習
25. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
25
3Dモデルの表現ベクトル空間学習・演算・生成
3D GAN (Wu+)
3Dモデルの生成 3Dモデル演算
写真からの3Dモデル生成
3D VAE-GAN
3D GAN
Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling.
Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, Joshua B. Tenenbaum.
arXiv:1610.07584. In NIPS 2016.
26. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
26
Generating Videos with Scene Dynamics(Vondrick+)
動画の表現ベクトル空間学習・動画生成
Generating Videos with Scene Dynamics.
Carl Vondrick, Hamed Pirsiavash, Antonio Torralba. In NIPS 2016.
http://web.mit.edu/vondrick/tinyvideo/
動画生成 1画像からその後の動画生成
27. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
27
f-GAN (Nowozin+)
GAN目的関数を Symmetric JS-divergence から
f-divergence に一般化。各Divergence を用い学習・評価
f-GAN: Training Generative
Neural samplers using
variational Divergence
Minimization.
Sebastian Nowozin, Botond
Cseke, Ryota Tomioka.
arXiv:1606.00709.
In NIPS 2016.
Kernel Density Estimation on the MNIST
f-divergence
LSUN
28. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
28
Improved Techniques for Training GANs (Salimans+)
Improved Techniques for Training GANs.
Tim Salimans, Ian Goodfellow, Wojciech
Zaremba, Vicki Cheung, Alec Radford, Xi Chen.
arXiv:1606.03498. In NIPS 2016.
収束が難しいGANの学習方法論
GAN半教師あり学習
1. Feature Matching
2. Minibatch discrimination
3. Historical averaging
4. One-sided label smoothing
5. Virtual batch normalization
Techniques Semi-supervised learning
MNIST
Semi-supervised training
with feature matching
Semi-supervised training
with feature matching and
minibatch discrimination
CIFAR-10
Generated samples
29. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
29
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
30. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
30
Extended Architectures for Generative Adversarial Nets 2016
Extended Architectures for GANs
Figure by Chris Olah (2016) : https://twitter.com/ch402/status/793535193835417601
Ex)
Conditional Image Synthesis With
Auxiliary Classifier GANs.
Augustus Odena, Christopher Olah,
Jonathon Shlens.
arXiv:1610.09585.
Generative Adversarial Net の各種拡張
31. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
31
Stack GAN: Text to PhotoRealistic Image Synthesis(Zhang+2016)
1段目で文章から低解像度画像を生成
2段目で低解像度画像から高解像度画像を生成
StackGAN: Text to Photo-realistic Image
Synthesis with Stacked Generative Adversarial
Networks.
Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang,
Xiaolei Huang, Xiaogang Wang, Dimitris Metaxas.
arXiv:1612.03242.
32. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
32
Plug & Play Generative Networks (Nguyen+2016)
高解像度な画像生成
227 x 227 ImageNet
Plug & Play Generative Networks: Conditional
Iterative Generation of Images in Latent Space.
Anh Nguyen, Jason Yosinski, Yoshua Bengio,
Alexey Dosovitskiy, Jeff Clune.
arXiv:1612.00005.
33. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
33
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016
34. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
34
Phased LSTM (Neil+)
時間で開閉するGateを導入した LSTM
Sensor Data 等、Event 駆動の長期系列特徴を学習
Phased LSTM: Accelerating Recurrent Network Training for Long or Event-based Sequences.
Daniel Neil, Michael Pfeiffer, Shih-Chii Liu.
arXiv:1610.09513. In NIPS 2016.
LSTM Phased LSTM
Phased LSTM Behavior
Frequency Discrimination Task
35. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
35
Using Fast Weights to Attend to the Recent Past (Ba+)
早く学習・減衰する Fast Weight 追加で、系列固有の情報を扱う
Slow Weight での長期特徴とあわせ、双方の系列特徴を学習
Using Fast Weights to Attend to the Recent Past.
Jimmy Ba, Geoffrey Hinton, Volodymyr Mnih, Joel Z. Leibo, Catalin Ionescu.
arXiv:1610.06258. In NIPS 2016.
Associative Retrieval Task
Classification Error Test Log Likelihood
36. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
36
Learning to learn by GD by GD (Andrychowicz+)
LSTMを用いたOptimizer
Parameterごとに 勾配系列から適切な次の更新量を算出
Learning to learn by gradient descent by gradient descent.
Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W. Hoffman, David Pfau, Tom Schaul, Brendan Shillingford,
Nando de Freitas.
arXiv:1606.04474. In NIPS 2016.
37. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
37
Matching Network for One Shot Learning (Vinyals+)
Attention Mechanism を用いた One Shot Learning
参照構造を学習しておき、新規小規模データセットでも高精度で動作
Matching Networks for One Shot Learning.
Oriol Vinyals, Charles Blundell, Timothy Lillicrap,
Koray Kavukcuoglu, Daan Wierstra.
arXiv:1606.04080. In NIPS 2016.
Omniglot
miniImageNet
38. Copyright (C) 2016 DeNA Co.,Ltd. All Rights Reserved.
38
AGENDA
◆Deep Learning Topics
◆NIPS 2016 Overview
◆Generative Adversarial Networks(GANs)
◆Recurrent Neural Networks(RNNs)
◆GANs
◆GANs in NIPS2016
◆Recent GANs
◆RNNs in NIPS2016