컴퓨터
16/10/24 01:05(년/월/일 시:분)
2016년 올해 벌써 학회 시즌이 돌아왔다. 매년 10월 말에서 11월 초에 열리는 사내 임직원 대상 학회로, 작년에 기계학습에 관심이 많았는데 마침 요슈아 벤지오 교수님 등 저명한 분들이 오셔서 잘 듣고 왔다.
하여튼 그때 간단히 노트했던 것들을 키워드 위주로 여기에 적어두고자 한다.
Deep Learning Beyond Classification
- deep learning (DL)
Trends and Challenges of Deep Learning Prof. Yoshua Bengio
(University of Montreal)
pattern recognition
speech recognition
memory networks - for natural language
caption generation
batch normalization
mini batch
importance sampling
{un,semi}supervised learning
neuro-biotics
training procedure w/ HW implementation
visual attention to internal attention (sequentialing)
ex) input(French) -> output(English)
time sequencing (AND gate)
attention mechanism
1. soft-attention : train by back-prop (fast)
2. stochastic hard-attiontion : noisy gradient (slower) reinforce base-line
end-to-end machine translation
3-inputs
1. higher-level RNN state
2. lower-level RNN state
3. previous attention pattern
image-to-text
caption generation
memory access
memorynet
batch normalization
must-have
small size
mini-batch importance sampling tricks
reduces (GPU) computational cost
unsupervised learning
VAEs - PRW
GANs - LAPGAN
ladder networks
100 -> topdown signal
challenges:
natural language understanding
reasoning & question answering (QA)
long-term dependencies
-Multimodal
Multi-modal Deep Learning Prof. Ruslan Salakhutdinov
(University of Toronto)
unlabeled data
drug discoveries: Merck, Novartis
zero-shot learning
caption generation
input -> CNN-LSTM encoder -> multimodal -> SC-NVM decoder
image generation from caption
Skip-thought
SemEval
MS Research Paraphrase Corpus
problems
one-shot problems
- DL & reinforcement learning
Deep Learning with Reinforcement Learning Prof. Joelle Pineau
(McGill University)
Dynamic System
Learning Agent
stall, reward : DS->LA
action : LA -> DS
supervised: input -> (desired) -> output
reinforce: input -> ( ) ->output -> environment -(reward)-> input, ( )
markov decision process
a a
St-1 -> St -> St+1
state-action value function
Q(s,a) = R(s,a) (immediate reward)
+ sum ( ) (future expected sum of reward)
Q function -> act -> new transition -> (feedback to Q function)
regression
random forest -> deep learning
Atari 2600 (DeepMind)
Q function
state -> action -> reward
stochastic gradient training
Deep QN
Deep Q learning
- experience replay
- error clipping
- periodic updates to target value
double DQN
learning ∩ planning = RL(reinforce learning)
dual encode model
RNN -> ->
AND gate ->
RNN -> ->
DL(deep learning)
dropout - full dropout policy, block dropout policy
- Scene understanding w/ DL
Scene Understanding with Deep Learning Prof. Antonio Torralba
(MIT)
computer vision
IMAGENET (2009)
based on WordNet
ontology
http://www.image-net.org
image detection
scene recognition
http://places.csail.mit.edu/demo.html
gaze following
saliency modeling
frames / scenario recognition
- Panel discussion
Panel Discussion : Deep Learning - Future Directions Moderator :
Prof. Ruslan Salakhutdinov
(Univ. of Montreal)
invest? maybe to drone, vision, health
pre-training?
-> After CNN(Alexnet), no more need?
-> still useful