Icml2018 naver review

ICML 2018 at a glance
4 / 30
Stockholmsmässan, Stockholm SWEDEN
Tuesday July 10 -- Sunday July 15, 2018

• ICML 2018
• Participants: ~5000
• Program Committee
• 160 Area Chairs
• 1776 Reviewers
• Submissions
• 2473 Submissions
• 621 Accepted Papers
• 9 Tutorials
• 67 Workshops! (with IJCAI & AAMAS)

• Federated AI 2018
• 다른 AI 학회와 공동 개최
• AAMAS
(International Conference on
Autonomous Agents and
Multiagent Systems)
• IJCAI
(International Joint Conference on
Artificial Intelligence)
• ICML 2019는 CVPR과 함께 개최
• ICML 2019: 6/10 ~ 6/15
• CVPR 2019: 6/15 ~ 6/21
• Long Beach, CA

ICLR 2018
stats
4 / 30
We need your collaboration with lots of labs~!

Recall: Taxonomy of Deep Generative Models
GAN tutorial @NIPS 2016, Ian Goodfellow

Generative Models in ICML 2018
Main session: Generative Models 1 ~ 5 (22 papers)
Workshop: Theoretical Foundations and Applications
of Deep Generative Models (2days, 66 papers)

Generative Models in ICML 2018
Main session: Generative Models 1 ~ 5 (22 papers)
Workshop: Theoretical Foundations and Applications
of Deep Generative Models (2days, 66 papers)
More related papers in main sessions:
Representation Learning / Deep Learning (Bayesian) /
Deep Learning (Adversarial) / Deep Learning (Neural
Network Architectures), more than 24 papers

Deep Learning (Neural Network Architectures)
RadialGAN: Leveraging multiple datasets to improve target-specific predictive models using GAN. Jinsung Yoon et al.
A Hierarchical Latent Vector Model for Learning Long-Term Structure in Music, Adam Roberts et al.
PixelSNAIL: An Improved Autoregressive Generative Model. Xi Chen et al.
Image Transformer. Niki Parmar et al.
Efficient Neural Audio Synthesis, Nal Kalchbrenner et al.
Deep Learning (Bayesian)
Variational Inference and Model Selection with Generalized Evidence Bounds. Liqun Chen, et al
Fixing a Broken ELBO. Alex Alemi, et al.
Tighter Variational Bounds are Not Necessarily Better, Tom Rainforth et al.
Neural Autoregressive Flows, Chin-Wei Huang et al.
Deep Learning (Adversarial)
Composite Functional Gradient Learning of Generative Adversarial Models. Rie Johnson et al.
Tempered Adversarial Networks. Mehdi S. M. Sajjadi et al.
Improved Training of Generative Adversarial Networks Using Representative Features. Duhyeon Bang et al.
(Techtalk: https://yobi.navercorp.com/techtalk/posts/545)
A Two-Step Computation of the Exact GAN Wasserstein Distance. Huidong Liu et al.
Is Generator Conditioning Causally Related to GAN Performance? Augustus Odena et al.
The Mechanics of n-Player Differentiable Games. David Balduzzi et al.
K-Beam Minimax: Efficient Optimization for Deep Adversarial Learning, Jihun Hamm et al.
First Order Generative Adversarial Networks. Calvin Seward et al.
Augmented CycleGAN: Learning Many-to-Many Mappings from Unpaired Data. Amjad Almahairi et al.
Mixed batches and symmetric discriminators for GAN training. Thomas LUCAS et al.
Mutual Information Neural Estimation. Mohamed Ishmael Belghazi et al.
Adversarially Regularized Autoencoders. Jake Zhao et al.
JointGAN: Multi-Domain Joint Distribution Learning with Generative Adversarial Nets. Yunchen Pu et al.

Representation Learning
Generative Temporal Models with Spatial Memory for Partially Observed Environments. Marco Fraccaro et al.
Disentangling by Factorising, Hyunjik Kim et al.
Discovering Interpretable Representations for Both Deep Generative and Discriminative Models. Tameem Adel et al.
Generative Models
Which Training Methods for GANs do actually Converge? Lars Mescheder, et al.
Chi-square Generative Adversarial Network. Chenyang Tao, et al.
Learning Implicit Generative Models with the Method of Learned Moments. Suman Ravuri, et al.
A Classification-Based Study of Covariate Shift in GAN Distributions. Shibani Santurkar et al.
Adversarial Learning with Local Coordinate Coding Jiezhang Cao et al.
Geometry Score: A Method For Comparing Generative Adversarial Networks. Valentin Khrulkov, et al.
Optimizing the Latent Space of Generative Networks. Piotr Bojanowski, et al.
Learning Representations and Generative Models for 3D Point Clouds. Panos Achlioptas, et al.
Theoretical Analysis of Image-to-Image Translation with Adversarial Learning. Morino Pan, et al.
Junction Tree Variational Autoencoder for Molecular Graph Generation. Wengong Jin, et al.
Semi-Amortized Variational Autoencoders. Yoon Kim, et al.
Iterative Amortized Inference. Joe Marino, et al.
DVAE++: Discrete Variational Autoencoders with Overlapping Transformations. Arash Vahdat, et al.
Parallel WaveNet: Fast High-Fidelity Speech Synthesis. Aäron van den Oord, et al.
Autoregressive Quantile Networks for Generative Modeling. Georg Ostrovski, et al.
Stochastic Video Generation with a Learned Prior, Emily Denton et al.
Disentangled Sequential Autoencoder. Yingzhen Li, et al.
----
Selected paper
Korean author (1st author only)

Augmented CycleGAN: Learning Many-to-Many Mappings from Unpaired Data.
TL; DR: proposed a new model which learns MANY-TO-MANY mappings between domains by
augmenting auxiliary latent spaces
Motivation: one-to-one mapping issue on Cycle GAN Recall: Overview of CycleGAN
Limitation of CycleGAN:
- CycleGAN only learns deterministic mappings, i.e., one-to-one mapping
- To resolve the problem, one can add ‘stochastic noise’ to CycleGAN, but
because of some theoretical reason, it also fails to map many-to-many

CycleGAN + augmenting auxiliary latent spaces

Experimental Results (‘many-to-many’)

JointGAN: Multi-Domain Joint Distribution Learning with Generative Adversarial Nets
TL; DR: GAN for joint distribution matching by learning marginals and conditionals simultaneously
https://zhegan27.github.io/Papers/JointGAN_poster.pdf
GANs and Adversarially Learned Inference (ALI)

JointGAN: Multi-Domain Joint Distribution Learning with Generative Adversarial Nets
Experimental Results

Semi-Amortized Variational Autoencoders
TL;DR Reduce amortization gap in VAEs by combining AVI / SVI
Motivation: Posterior Collapse and Inference Gap
http://www.people.fas.harvard.edu/~yoonkim/data/sa-vae-slides.pdf

Combining AVI / SVI (Amortized Variational Inference / Stochastic Variational Inference)

Experimental Results

Experimental Results (Visualization using output saliency of each token)

Paper List
● Speech Synthesis
○ Towards End-to-End Prosody Transfer for Expressive Speech Synthesis
with Tacotron
○ Style Tokens: Unsupervised Style Modeling, Control and Transfer in En
d-to-End Speech Synthesis
● Music
○ A Hierarchical Latent Vector Model for Learning Long-Term Structure i
n Music
○ 2018 Joint Workshop on Machine Learning for Music

● InterSpeech(2017), Google
● End-To-End Speech Synthesis model
TacoTron: Towards End-To-End Speech Synthesis

Towards End-to-End Prosody Transfer for Expressive Speech
Synthesis with Tacotron
● Tacotron 기반의 음성 합성 네트워크
● Text뿐 아니라 원하는 prosody(운율, 억양)을 가진 오디오 샘플을 넣어서 생성
Audio Samples
Text : I’ve Swallowed a Pollywog.
https://ai.googleblog.com/2018/03/expressive-speech-synthesis-
with.html
Reference Prosody :
Without Prosody :
With Prosody :

Style Tokens: Unsupervised Style Modeling, Control and Tra
nsfer in End-to-End Speech Synthesis
● 원하는 스타일에 해당하는 audio를 직접 넣어서 하지 말고 내가 특정한 toke
n들을 가지고 스타일을 조절할 수 있게 만들어 보자.
● Random initialized된 style embedding을 attention을 이용해서 상대적인 wei
ght를 학습하게 하고, inference 때는 이 weight를 사용자가 조절

● 음악 구조를 반영할 수 있는 latent space를 Recurrent VAE기반으로
Hierarchical Decoder를 추가한 MusicVAE 학습
● RVAE의 posterier collapse 문제 해결
● Latent space interpolation도 가능
A Hierarchical Latent Vector Model for Learning Long-Term
Structure in Music
https://magenta.tensorflow.org/music-vae
Interpolation example: https://youtu.be/G5JT16flZwM

Joint Workshop on Machine Learning for Music
● 매년 열리던 뮤직 워크샵을 이번에 IJCAI등과 같이 Joint로 진행
● Clova에서는 Poster, Oral 하나씩 진행
○ Onset Detection, Representation Learning
Modeling Musical Onset Probabilities via
Neural Distribution Learning (Huh et. al)

Joint Workshop on Machine Learning for Music
● 매년 열리던 뮤직 워크샵을 이번에 IJCAI등과 같이 Joint로 진행
● Clova에서는 Poster, Oral 하나씩 진행
○ Onset Detection, Representation Learning
A Hybrid of Deep Audio Feature and i-ve
ctor for Artist Recognition (Park et. al)

Trends in Music Research
● Representation Learning
● Siamese Networks
● Deep network analysis

Trends in Music Research
● Style Transfer
● Source Separation
● Using traditional DSP features

AI and Secure on ICML 2018
• One Invited Talk
• AI and Security: Lessons,
Challenges and Future
Directions (Dawn Song)
• Deep Learning (Adversarial)
• 28 Papers on 6 Sessions
• Privacy, Anonymity, and Security
• 10 Papers on 2 Sessions

Adversarial Example
Akhtar, Naveed, and Ajmal Mian. "Threat of adversarial attacks on deep learning in
computer vision: A survey." arXiv preprint arXiv:1801.00553 (2018).

Adversarial Attack on All Area
http://ai.berkeley.edu/reinforcement.html
blog.malwarebytes.com
• Reinforcement Learning
• Generative Modeling
• Face Recognition
• Object Detection
• Semantic Segmentation
• Reading Comprehension
• Malware Detection
Cisse, Moustapha, et al. "Houdini: Fooling deep structured prediction
models." arXiv preprint arXiv:1707.05373 (2017).

Threat of Adversarial Attacks: Autonomous Driving
Sitawarin, Chawin, et al. "DARTS: Deceiving Autonomous Cars with Toxic Signs."
arXiv preprint arXiv:1802.06430 (2018).

Adversarial Attacks and Defenses
Goodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. "Explaining and
harnessing adversarial examples (2014)." arXiv preprint arXiv:1412.6572.
Adversarial Attack
Adversarial Example

Recent Studies about Adversary
Yuan, Xiaoyong, et al. "Adversarial examples: Attacks and defenses for deep
learning." arXiv preprint arXiv:1712.07107 (2017).

White-box and Black-box attacks
• White-box attack
• Attackers know the complete knowledge of the target model
• model parameters, architecture, training method
Obfuscated gradients give a false sense of security: circumventing
defenses to adversarial examples (Anish et al.)
• Black-box attack
• Out of knowledge (or limited knowledge) of the target model
• Number of features, algorithm
Black-box Adversarial Attacks with Limited Queries and Information
(Andrew et al.)

Obfuscated gradients give a false sense of security: circumventing
defenses to adversarial examples
• One-liner: Most of current defenses fall into a category of gradient
masking which is a failed defense
Athalye, Anish, Nicholas Carlini, and David Wagner. "Obfuscated gradients give a false
sense of security: Circumventing defenses to adversarial examples." arXiv preprint
arXiv:1802.00420 (2018).

Gradient Masking
• A term introduced in Practical Black-Box Attacks against Deep Learning
System using Adversarial Examples (Nicolas et al., 2017)
• An entire category of failed defense methods that work by trying to deny
the attacker access to a useful gradient
* Papernot, Nicolas, et al. "Practical black-box attacks against machine learning."
Proceedings of the 2017 ACM on Asia Conference on Computer and Communications
Security. ACM, 2017.

Obfuscated gradients
• Gradient-based attacks cannot succeed without a gradient signal
• Three types of obfuscated gradients:
• Shattered gradients
• Stochastic gradients
• Vanishing/exploding gradients
• Defenses that cause obfuscated gradients appear to defeat iterative
optimization-based attacks, but defenses relying on this effect can be
circumvented using new techniques we develop
arXiv:1802.00420 (2018).

The Proposed Attacks
• Backward Pass Differentiable Approximation (BPDA)
• Expectation Over Transformation (EOT)
• Reparameterization
arXiv:1802.00420 (2018).

Case study: ICML 2018 Defenses
• 9 white-box defenses proposed
at ICLR 2018
• 7 defenses rely on obfuscated
gradients
• 6 defenses fall to the new
attacks (1 partially succumbs)
arXiv:1802.00420 (2018).

Contributions
• Attacked and beat most of recently developed defenses
• Illuminated the strengths and weaknesses of current adversarial defense
technology
• Won a best paper award at ICML 2018
arXiv:1802.00420 (2018).
blog.acolyer.org

Black-box Adversarial Attacks with Limited Queries and Information
• One-liner: Define three realistic black-box settings and propose new
adversarial attacks
https://www.labsix.org/partial-information-adversarial-examples/

Real-world AI
• “White-box” assumption is rarely
applicable to machine learning
systems deployed in the wild
• Real-world AI is black box
• Three types of black-box settings
• Query-limited setting
• Partial-information setting
• Label-only setting
Google cloud Vision API (GCV)
www.clarifai.com

Query-limited setting example: clarifai NSFW
The Clarifai NSFW (Not Safe for Work) detection API is a binary classifier that outputs P(NSFW|x) for any
image x and can be queried through an API. However, after the first 2500 predictions, the Clarifai API
costs upwards of $2.40 per 1000 queries. This makes a 1-million query attack, for example, cost $2400.
Ilyas, Andrew, et al. "Black-box Adversarial Attacks with Limited Queries and
Information." arXiv preprint arXiv:1804.08598 (2018).
sfw: 0.650 nsfw: 0.759 sfw: 0.641
nsfw: 0.359
excerpted from www.clarifai.com

Partial information setting example: Google Cloud Vision
The Google Cloud Vision API (GCV) only outputs scores for a number of the top classes (the number
varies between queries). The score is not a probability but a “confidence score” (that does not sum to
one).

Label-only setting example: Google Photos
Photo tagging apps such as Google Photos add labels to user-uploaded images. However, no
“scores” are assigned to the labels, and so an attacker can only see whether or not the classifier
has inferred a given label for the image (and where that label appears in the ordered list)

Approach: Natural evolution strategies (NES)
• Introduced by Wierstra, Daan, et al. "Natural evolution strategies."
Evolutionary Computation, 2008.
• More query-efficient than previous approaches
• substitute networks
• coordinate-wise gradient estimation

Results
Figure 2. The distribution of the number of queries required for the query-limited (top) and
partial-information with k = 1 (bottom) attacks.

Case study: Google Cloud Vision Demo

Conclusion
• Current status
• Attack, Attack, Attack
• The robustness of the model should be considered
• Attack and defense improve each other
• More human-like, higher level semantic understanding,
is needed
source: Jo Seok
source: i,Robot

ICML 2018
Reinforcement Learning

ICML’18: Reinforcement Learning
• 논문이 가장 많았던 분야
• 총 17개의 session, 72개의 paper
• 모든 oral 시간대에 RL session 존재
• “International Conference on RL”
• 튜토리얼
• Imitation Learning
• Optimization Perspectives on Learning to
Control
• 워크샵
• Credit Assignment in DL and DRL
• Prediction and Generative Modeling in RL
• Lifelong Learning: A RL Approach
• Goal Specifications for RL
• Exploration in RL

Trend#1. Imitation Learning
• 전통적인 RL에서는 agent가 직접 행동하면서 sample을 수집
• Reward function 디자인 문제
• 학습이 오래 걸리며 많은 sample 필요
• Imitation learning: 전문가의 행동 기록을 이용하여 agent 학습
• 전문가의 과거 기록을 사용하거나, 혹은 실시간으로 전문가에게 시범을 요청
4 / 30 Tutorial: Imitation Learning

• Trajectory 기반
• Behavioral cloning: supervised learning으로 환원
• Inverse RL: trajectory가 목표로 하는 reward function 예측
• Interactive demonstrator 기반
• Agent 학습 중간에 expert에게 특정 demonstration을 요청
• DAgger, SEARN, Learning to Search
4 / 30
Tutorial: Imitation Learning

• Some ICML’18 papers:
• Hierarchical Imitation and Reinforcement Learning
• Self-Imitation Learning
• Programmatically Interpretable Reinforcement Learning
• Other recent papers:
• One-Shot Imitation Learning (NIPS’17)
• Playing hard exploration games by watching YouTube (arxiv 1805)
• OpenAI Blog: Learning Montezuma’s Revenge from a Single Demonstration (1806)
• BAIR Blog: One-Shot Imitation from Watching Videos (CoRL’17, RSS’18)

Hierarchical Imitation and Reinforcement Learning
• Hierarchical RL: high-level policy와 low-level policy로 분리하여 모델
• High-level policy는 “열쇠를 얻는다”처럼 subtask 생성을 담당
• Low-level policy는 세부적인 컨트롤을 담당
• Hierarchically guided DAgger
• Interactive demonstrator 알고리즘 DAgger 기반
• Expert는 주로 high-level planning을 demonstrate
• Low-level control은 DQN 등의 기존 RL 사용 가능
• Expert의 demonstration 비용을 줄이면서도 빠르게 학습 가능
4 / 30

Trend#2. Model-based Learning
• 기존의 많은 (deep) RL 알고리즘은 model-free 기반
• 세계가 어떻게 동작하는지를 배우는 대신, 행동의 결과만을 학습
• Q-learning (DQN), policy gradient
• Reward나 dynamic가 조금만 바뀌어도 처음부터 학습 필요
• Model-based learning:
• 세계의 dynamics를 배우면서 이에 기반하여 policy 학습
• Goal이나 reward가 바뀌더라도 model을 재활용 가능
• Multi-task/transfer/meta learning
• Representation learning
Learning World Models With Self-Supervised Learning

• LeCun: Learning World Models With Self-Supervised Learning
4 / 30

• Recht: Tutorial: Optimization Perspectives on Learning to Control
• Control theory: 산업공학/기계공학에서 로봇 등의 컨트롤 문제를 다루는 분야
• 문제는 RL과 유사하지만, 방법론에서 차이 존재

• Recht: Tutorial: Optimization Perspectives on Learning to Control
• ”REINFORCE is NOT Magic”
• 학습이 느리고, 학습 결과의 편차가 큼
• Random search (ES)가 더 좋은 결과 리포트

• Some ICML’18 papers:
• Machine Theory of Mind
• Deep Variational Reinforcement Learning for POMDPs
• Recurrent Predictive State Policy Networks
• Gated Path Planning Networks
• Universal Planning Networks: Learning Generalizable Representations for Visuomotor Control
• Workshop: Prediction and Generative Modeling in RL
• Other recent papers:
• The Predictron: End-To-End Learning and Planning (ICML’17)
• Value Prediction Network (NIPS’17) (PGMRL’18 talk slides)
• TreeQN and ATreeC: Differentiable Tree-Structured Models for Deep Reinforcement Learning (ICLR’18)
• World Models (web), Recurrent World Models Facilitate Policy Evolution (NIPS’18)

Machine Theory of Mind
• Theory of Mind: 다른 사람의 생각이나 느낌을 예측하는 능력
• Sally-Anne test
• “Sally가 구슬을 바구니에 넣었어요.”
• “Sally가 없을 때 Anne가 구슬을 상자로 옮겼어요.”
• “Sally는 구슬이 어디에 있다고 생각할까요?”
• 실제 상황을 예측하는 것이 아니라,
상대방의 관찰 능력에 기준한 상황을 예측하는 능력

Machine Theory of Mind
• Theory of Mind와 유사하게, 다른 agent의 능력이나 행동을 예측하는 학습 framework를 제안
• 3개의 네트워크로 구성
• Characteristic network: agent의 agent의 trajectory에서 특성 embedding 생성
• Mental state network: agent의 현재 trajectory에서 목적 embedding 생성
• Prediction network: 현재 state에서 다음 행동 예측
• Sally-Anne test
• POMDP belief state tracking 문제로 formulate
• Agent가 false belief를 갖고 있다면 그에 맞는 예측을 생성
4 / 30

The Mirage of Action-Dependent Baselines in Reinforcement Learning
• Policy gradient
• Policy parameter를 gradient estimator를 이용하여 학습
• 구현하기 쉽지만, estimator의 variance가 커서 비효율적
• low-variance 튜닝 필요: “baseline”, “control variate”, “actor-critic”
• Recent papers on better baselines
• Qprop (ICLR’17), IPG (NIPS’17), REBAR (NIPS’17), RELAX (ICLR’18), Stein control variate (ICLR’18), …
• 과거 논문의 코드를 분석해보니 버그 존재
• 버그가 estimator bias를 일으켜서 성능 향상, 버그를 수정하면 성능 향상 X
4 / 30

Investigating Human Priors for Playing Video Games
• RL 알고리즘이 비디오 게임을 클리어하기까지 상당한 시간 필요
• 반면 사람은 빠르게 클리어 가능 – 사람이 가진 prior knowledge의 정체는?
4 / 30

• 게임의 그래픽을 변형하여 사람의 학습 속도 변화 측정
4 / 30

4 / 30
• 사다리나 열쇠 등의 semantic을 숨겨도 빠르게 학습 가능

4 / 30
• 반면, 물체와 배경을 구분하지 못하면 학습에 상당한 시간 필요
• 유아 발달 과정 순서가 물체 범위 구별 – 시맨틱 학습인 것과 유사

Contents
1
2
3
GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep
Multitask Networks
Deep Asymmetric Multi-task Feature Learning
Pseudo-task Augmentation: From Deep Multitask Learning to Intratask
Sharing—and Back

# Introduction of Multi-task Learning
- 정의:
§ Single task: Input X à output Y
§ Multi task : Input X à output Y1 ,Y2 ,Y3 …
- 장점:
§ Knowledge Transfer : Y1 을 학습하면서 얻은 유용한 정보가 다른 Y2, Y3 에 좋은 영향을 줌
§ Overfitting 감소: 여러 task를 동시에 맞추기 위해 보다 generalized된 shared representation을 학습
§ Computational Efficiency : Input X à output Y1 ,Y2 ,Y3 …
§ Real-world Application에서 매우 필요함 : 실세계 문제에서는 매우 다양한 task가 요구 됨. ex) 무인차
- 어려움:
§ Negative Transfer: 다른 task에 악영향을 미치는 Y가 존재할 수 있음.
§ Task Balancing의 어려움 : task 마다 학습 난이도가 크게 차이 나면 수렴하지 않거나, robust하지 않음.

1. GradNorm: Gradient Normalization for Adaptive Loss Balancing in
Deep Multitask Networks
Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, Andrew Rabinovich
4 / 30

4 / 30

2. Deep Asymmetric Multi-task Feature Learning
Hae Beom Lee, Eunho Yang, Sung Ju Hwang

2. Deep Asymmetric Multi-task Feature Learning

3. Pseudo-task Augmentation: From Deep Multitask Learning to
Intratask Sharing—and Back
Elliot Meyerson, Risto Miikkulainen

3. Pseudo-task Augmentation: From Deep Multitask Learning to
Intratask Sharing—and Back

Contents
- Introduction
- AutoML in ICML 2018
- AutoML in NSML
- Conclusion

Introduction
- Perspectives of AutoML
- Hyperparameter Optimization
- Neural Architecture Search
- Meta-learning
- Optimizing ML Pipeline
- Solutions for AutoML
- RL / other game-playing algorithm (MCTS)
- Bayesian Optimization
- Systematic method (parallel computing, resource controlling, …)

Growth of AutoML
- workshops (https://www.ml4aad.org/workshops/)
- ICML AutoML Workshop (since 2014 ~)
- Since 2017, there are multiple workshops related to AutoML
- ICML AutoML workshop accepted papers
- 2014: 12, 2015: 14, 2016: 10, 2017: 17, 2018: 36
- challenges
- 3rd AutoML Challenge in NIPS 2018

AutoML in ICML 2018
- In sessions (Bayesian Optimization)
- 2 Bayesian Optimization sessions, 8 papers
- 5 out of 8 papers mainly use hyperparameter optimization as an appli
cation of BO
- AutoML Workshop (https://sites.google.com/site/automl2018icm
l/)
- 36 accepted papers
- 3 invited talks by Chelsea Finn, Zoubin Ghahramani, and Luc de Raedt
- 1 panel discussion session

Papers
Bayesian Optimization (BO)
- BOCK: Bayesian Optimization with Cylindrical Kernel

Papers
Neural Architecture Search (NAS)
- Dynamic Architecture Search

Papers
Systematic AutoML
- Tune

Papers
AutoML as one-player game
- AutoML with Monte Carlo Tree Search

Summary
- So far, systematic AutoML works better than sequential learning
like Bayesian Optimization or Reinforcement learning
- Hybrid of BO/RL with systematic AutoML (hyperband, PBT, and s
o on) is trendy
- People are trying to use RL for AutoML, but not so good so far

AutoML in NSML
Introduction of NSML AutoML
- NSML AutoML is python framework based on nsml
Systematic AutoML
- Parallel Random search with early stopping
- Population Based Training
- Hyperband

Conclusion
AutoML is growing
- many other interesting algorithms on especially NAS
- many global companies (google, facebook, deepmind, etc) and
top-tier universities (UC Berkeley) are focusing on AutoML
nsml AutoML
- on alpha test (hopely start beta test soon)
- trying to add more algorithms and features

Paper list
- BO
- BOHB (https://arxiv.org/abs/1807.01774)
- BOCK (https://arxiv.org/abs/1806.01619)
- NAS
- Evolutionary vs. RL (https://docs.google.com/viewer?a=v&pid=sites&sr
cid=ZGVmYXVsdGRvbWFpbnxhdXRvbWwyMDE4aWNtbHxneDo0NDFm
ZWM5ZTljMTZkZDg1)
- Dynamic Architecture Search (https://docs.google.com/viewer?a=v&pid
=sites&srcid=ZGVmYXVsdGRvbWFpbnxhdXRvbWwyMDE4aWNtbHxne
Do2ZDc4N2Q1NGZjMDM2Y2Vm)

Paper list (cont’d)
- Systematic AutoML
- Tune (https://docs.google.com/viewer?a=v&pid=sites&srcid=ZGVmYXV
sdGRvbWFpbnxhdXRvbWwyMDE4aWNtbHxneDo0MDhmMjEyYjgzY2Y5
YmEx)
- Orion (https://openreview.net/pdf?id=r1xkNLPixX)
- AutoML as one player game
- with MCTS (https://docs.google.com/viewer?a=v&pid=sites&srcid=ZG
VmYXVsdGRvbWFpbnxhdXRvbWwyMDE4aWNtbHxneDozNWU4NWE3
OWRhM2IyZTk4)
- AlphaD3M(https://docs.google.com/viewer?a=v&pid=sites&srcid=ZGV
mYXVsdGRvbWFpbnxhdXRvbWwyMDE4aWNtbHxneDo0OTQzOThjNmZ
mYjYxYjc3)

Other papers (not in ICML)
Population Based Training (https://arxiv.org/abs/1711.09846)
Differentiable Architecture Search (https://arxiv.org/abs/1806.09055)
Hyperband (https://arxiv.org/abs/1603.06560)
NAS with RL (https://arxiv.org/abs/1611.01578)

Icml2018 naver review

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Icml2018 naver review

Similar to Icml2018 naver review (20)

More from NAVER Engineering

More from NAVER Engineering (20)

Recently uploaded

Recently uploaded (20)

Icml2018 naver review