SLIDE 1 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 1
Administrative
- Poster Session on Wednesday, worth 3% of final grade,
+2% for top few posters. There will be food
- CS224D (Deep Learning for NLP) was announced for next
quarter taught by Richard Socher, natural followup for more DL.
SLIDE 2 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 2
CS224D Syllabus and Schedule
Event Type Date Description Lecture Week 1 Intro to NLP Lecture Week 1 Simple Word Vector representations: word2vec, GloVe Lecture Week 2 Optimization (SGD, mini-batches), Visualization (PCA, t-sne) Lecture Week 2 Advanced word vector representations: language models, softmax, clustering (k-means) Lecture Week 3 Neural Networks and backpropagation Lecture Week 3 Practical tips: gradient checks, overfitting, regularization, activation functions, details Lecture Week 4 Recurrent neural networks Lecture Week 4 GRUs and LSTMs Lecture Week 5 Recursive neural networks Lecture Week 6 Convolutional neural networks Lecture Week 6 Novel Memory Models Lecture Week 7 Additional applications not covered as motivating examples yet Lecture Week 7 Efficient implementations and GPUs Lecture Week 8 Invited Speaker: TBD Lecture Week 8 Future applications and open problems
SLIDE 3
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 3
Tiny ImageNet Spotlights
SLIDE 4
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 4
SLIDE 5
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 5
Together, we’ve defined Score Functions...
SLIDE 6
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 6
And Loss Functions...
SLIDE 7
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 7
We’ve learned how to optimize them...
Chain rule:
SLIDE 8
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 8
NEURAL NETWORKS
We learned to express more powerful Score Functions...
SLIDE 9
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 9
For an extra wiggle...
SLIDE 10
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 10
Together we tamed the learning process...
SLIDE 11 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 11
Together we explored image-specific Neural Nets...
CONV ReLU CONV ReLU POOLCONV ReLU CONV ReLU POOL CONV ReLU CONV ReLU POOL FC (Fully-connected)
SLIDE 12
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 12
We explored how they work...
SLIDE 13
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 13
And how they don’t… (but really they still do)
SLIDE 14
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 14
We looked at what makes ConvNets “tick”...
SLIDE 15
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 15
And studied their mysterious generalization powers…
SLIDE 16
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 16
We learned tips/tricks for making ConvNets work well in practice
SLIDE 17
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 17
And explored their practical bottlenecks...
Moving parts lol
SLIDE 18
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 18
And we bravely ventured beyond Image Classification...
SLIDE 19 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 19
And developed an understanding of cutting-edge research
Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
[Kaiming He et al., 2015] (MSR)
4.94% error Top 5 ImageNet error
We saw many 2015 citations... e.g.
SLIDE 20
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 20
You are now ready.
SLIDE 21
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 21
You are now ready.
SLIDE 22
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 22
You are now ready.
SLIDE 23
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 23
Hints of beyond
SLIDE 24 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 24
Reinforcement Learning meets Computer Vision
Human-level control through deep reinforcement learning [Mnih et al.], Nature 2015 http://www.nature.com/nature/journal/v518/n7540/full/nature14236.html
SLIDE 25 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 25
(play videos)
http://www.nature. com/nature/journal/v518/n7 540/full/nature14236.html
SLIDE 26
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 26
Reinforcement Learning meets Computer Vision
SLIDE 27
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 27
Reinforcement Learning meets Computer Vision
SLIDE 28 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 28
action values Q(s,a)
(Approximate idea of the model)
(screen pixels from few time steps) ConvNet
SLIDE 29 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 29
action values Q(s,a)
- Assume finite number of actions
- Each number here is a real-valued quantity
that represents the “Q function” in RL (Approximate idea of the model)
(screen pixels from few time steps) ConvNet
- Collect experience dataset:
set of tuples {(s,a,s’,r), … } (State, Action taken, New state, Reward received)
SLIDE 30 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 30
action values Q(s,a)
- Assume finite number of actions
- Each number here is a real-valued quantity
that represents the “Q function” in RL (Approximate idea of the model)
(screen pixels from few time steps) ConvNet
- Collect experience dataset:
set of tuples {(s,a,s’,r), … } (State, Action taken, New state, Reward received) L2 Regression loss:
target value predicted value
SLIDE 31 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 31
action values Q(s,a)
- Assume finite number of actions
- Each number here is a real-valued quantity
that represents the “Q function” (Approximate idea of the model)
(screen pixels from few time steps) ConvNet
- Collect experience dataset:
set of tuples {(s,a,s’,r), … } (State, Action taken, New state, Reward received)
target value predicted value
reward estimate of future reward (discounted by \gamma)
SLIDE 32 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 32
Recurrent Attention Models
Multiple Object Recognition with Visual Attention [Jimmy Lei Ba, Volodymyr Mnih, Koray Kavukcuoglu], 2014 web demo http://www.psi.toronto.edu/~jimmy/dram/ also DRAW: https://www.youtube.com/watch?v=Zt- 7MI9eKEo
SLIDE 33 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 33
Show, Attend and Tell: Neural Image Caption Generation with Visual Attention [Kiros et al.] 2015 Neural machine translation by jointly learning to align and translate. [Bahdanau et al.], 2014
SLIDE 34 Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 34
Show, Attend and Tell: Neural Image Caption Generation with Visual Attention [Kiros et al.] 2015 Neural machine translation by jointly learning to align and translate. [Bahdanau et al.], 2014
SLIDE 35
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 35
SLIDE 36
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 36
SLIDE 37
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 37
END
SLIDE 38
Fei-Fei Li & Andrej Karpathy Lecture 8 - 2 Feb 2015 Fei-Fei Li & Andrej Karpathy Lecture 12 - 4 Mar 2015 38
