최근 이수가 되고 있는 Bayesian Deep Learning 관련 이론과 최근 어플리케이션들을 소개합니다. Bayesian Inference 의 이론에 관해서 간단히 설명하고 Yarin Gal 의 Monte Carlo Dropout 의 이론과 어플리케이션들을 소개합니다.
3. Classic Deep Learning
∙ A classification model is expressed as f(x) = p(y ∈ c|x, θ)
”The probability that y belongs to the class c predicted from the
observation x”
∙ Training a model is defined as θ∗
= arg minθ
1
N
∑N
i L(xi, yi, θ)
”Finding the parameter θ∗
that minimizes the loss metric L”
2
4. Likelihood
A dataset is denoted as {(x, y)} = D
L(D, θ) = − log p(D|θ)
∙ How likely is the distribution p to fit the data.
∙ minimizing L is maximum likelihood estimation (MLE)
∙ The log negative probability density function (PDF) of p is often
used as MLE
∙ binary cross entropy (BCE) loss
∙ Ordinary Least Squares (OLS) loss
3
8. Regularized Log Likelihood
L(D, θ) = −(log p(D|θ) + logp(θ))
∙ The use of Bayes’ rule to incorporate ’prior knowledge’ into the
problem
∙ Also called maximum a posteriori estimation (MAP)
p(θ|D) =
p(D|θ)p(θ)
p(D)
∝ p(D|θ)p(θ)
L(x, y, θ) = − log p(θ|D)
∝ − log (p(D|θ)p(θ))
= −(log p(D|θ) + logp(θ))
7
9. MAP and MLE Estimation
θ∗
MAP = arg min
θ
[− log p(D|θ) − logp(θ)]
θ∗
MLE = arg min
θ
[− log p(D|θ)]
∙ MLE and MAP estimation only estimate a fixed θ
∙ The resulting predictions are a fixed probability value
∙ In reality, θ might be better expressed as a ’distribution’
f(x) = p(y|xθ∗
MAP) ∈ R
8
10. Bayesian Inference
Eθ[ p(y|x, D) ] =
∫
p(y|x, D, θ)p(θ|D)dθ
∙ Integrating across all probable values of θ (Marginalization)
∙ Solving the integral treats θ as a distribution
∙ For a typical modern deep learning network, θ ∈ R1000000...
∙ Integrating for all possible values of θ is intractable (impossible)
9
11. Bayesian Methods
Instead of directly solving the integral,
p(y|x, D) =
∫
p(y|x, D, θ)p(θ|D)dθ
we approximate the integral and compute
∙ The expectation E[ p(y|x, D) ]
∙ The variance V[ p(y|x, D) ]
using...
∙ Monte Carlo Sampling
∙ Variational Inference (VI)
10
12. Output Distribution
Predicted distribution of p(y|x, D) can be visualized as
∙ Grey region is the confidence interval computed from V[ p(y|x, D) ]
∙ Blue line is the mean of the prediction E[ p(y|x, D) ]
11
13. Why Bayesian Inference?
Modelling uncertainty is becoming important in failure critical
domains
∙ Autonomous driving
∙ Medical diagnostics
∙ Algorithmic stock trading
∙ Public security
12
14. Decision Boundary and Misprediction
∙ MLE and MAP estimations lead to a fixed decision boundary
∙ ’Distant samples’ are often mispredicted with very high confidence
∙ Learning a ’distribution’ can fix this problem
13
15. Adversarial Attacks
∙ Changing even a single pixel can lead to misprediction
∙ These mispredictions have a very high confidence
2
2Su, Jiawei, Danilo Vasconcellos Vargas, and Sakurai Kouichi. ”One pixel attack for
fooling deep neural networks.” arXiv preprint arXiv:1710.08864 (2017).
14
16. Autonomous Driving
3
3Kendall, Alex, and Yarin Gal. ”What uncertainties do we need in bayesian deep
learning for computer vision?.” Advances in neural information processing systems.
2017. 15
17. Monte Carlo Intergration
p(y|x, D) =
∫
p(y|x, D, θ)p(θ|D)dθ
≈
1
S
S∑
s=0
p(y|x, D, θs)
where θs are samples from p(θ|D)
∙ Samples are directly pulled from p(θ|D)
∙ In case sampling from p is not possible, use MCMC
16
19. Variational Inference
∙ Variational Inference converts an inference problem into an
optimization problem.
∙ instead of using a complicated distribution such as p(θ | D) we
find a tractable approximation q(θ, λ) parameterized with λ
∙ This is equivalent to minimizing the KL divergence of p and q
∙ Using a distribution q very different to p leads to bad solutions
minimize
λ
KL(q(x; λ) || p(x))
18
21. Evidence Lower Bound (ELBO)
Because of the evidence term p(D) is intractable, optimizing the KL
divergence directly is hard.
However By reformulating the problem,
KL(q(θ; λ)||p(θ|D)) = Eq[− log p(θ, D) + log q(θ; p)] + log p(D)
log p(D) = KL(q(θ; λ)||p(θ|D)) − Eq[− log p(θ, D) + log q(θ; λ)]
log p(D) ≥ Eq[log p(θ, D) − log q(θ; λ)]
∵ KL(q(θ, λ)||p(θ|D)) ≥ 0
20
22. Evidence Lower Bound (ELBO)
maximizeλ L[q(θ; λ)] = Eq[log p(θ, D) − log q(θ; λ)]
∙ Maximizing the evidence lower bound is equivalent of minimizing
the KL divergence
∙ ELBO and KL divergence become equal at the optimum
21
24. Dropout Regularization
∙ Very popular deep learning regularization method before batch
normalization (9000 citations!)
∙ Make weight Wij = 0 following a Bernoulli(p) distribution
4
4Srivastava, Nitish, et al. ”Dropout: a simple way to prevent neural networks from
overfitting.” The Journal of Machine Learning Research 15.1 (2014): 1929-1958. 23
26. Dropout As Variational Approximation
Solving MLE or MAP using dropout is variational inference.
Yarin Gal, PhD Thesis, 2016
The distribution of the weights p(W|D) is approximated using q(p, W)
q(p) is the distribution of the weight W with dropout applied
yi = (Wiyi−1 + bi) ri where ri ∼ Bern(p)
Since L2 loss and L2 regularization assumes W ∼ N(µ, σ2
), the
resulting distribution q is,
q(Wij; p) ∼ p N(µij, σ2
ij) + (1 − p) N(0, σ2
ij)
25
27. Dropout As Variational Approximation
Since the ELBO is given as,
maximizeW,p L[q(W; p)]
= Eq[ log p(W, D) − log q(W; p) ]
∝ Eq[ log p(W|D) −
p
2
|| W ||2
2 ]
=
1
N
N∑
i∈D
log p(W|xi, yi) −
p
2σ2
|| W ||2
2
is the optimization objective.
∙ if p approaches 1 or 0, q(W; p) becomes a constant distribution.
26
28. Monte Carlo Inference
Eθ[ p(y|x, D)] =
∫
p(y|x, D, θ)p(θ)dθ
≈
∫
p(y|x, D, θ)q(θ; p)dθ
= Eq[p(y|x, D)]
≈
1
T
T∑
t
p(y|x, D, θt) θt ∼ q(θ; p)
∙ Prediction is done with dropout turned on and averaging multiple
evaluations.
∙ This is equivalent to monte carlo integration by sampling from the
variational distribution.
27
29. Monte Carlo Inference
Vθ[ p(y|x, D)] ≈
1
S
S∑
s
( p(y|x, D, θs) − Eθ[p(y|x, D)] )2
Uncertainty is the variance of the samples taken from the variational
distribution.
28
30. Monte Carlo Dropout
Examples from the mauna loa CO2 dataset 6
6Gal, Yarin, and Zoubin Ghahramani. ”Dropout as a Bayesian approximation:
Representing model uncertainty in deep learning.” ICML 2016.
29
31. Monte Carlo Dropout Example
Prediction using only 10 samples 7
7Gal, Yarin, and Zoubin Ghahramani. ”Dropout as a Bayesian approximation:
Representing model uncertainty in deep learning.” ICML 2016.
30
32. Monte Carlo Dropout Example
Semantic class segmentation 8
8Kendall, Alex, and Yarin Gal. ”What uncertainties do we need in bayesian deep
learning for computer vision?.” NIPS 2017.
31
33. Monte Carlo Dropout Example
Spatial depth regression 9
9Kendall, Alex, and Yarin Gal. ”What uncertainties do we need in bayesian deep
learning for computer vision?.” NIPS 2017.
32
34. Medical Diagnostics Example
∙ Green: True positive, Red: False Positive
10
10DeVries, Terrance, and Graham W. Taylor. ”Leveraging Uncertainty Estimates for
Predicting Segmentation Quality.” arXiv preprint arXiv:1807.00502 (2018).
33
35. Medical Diagnostics Example
11
∙ Green: True positive, Blue: False Negative
11DeVries, Terrance, and Graham W. Taylor. ”Leveraging Uncertainty Estimates for
Predicting Segmentation Quality.” arXiv:1807.00502 (2018).
34
36. Possible Medical Applications
∙ Statistically correct uncertainty quantification
∙ Bandit setting clinical treatment planning (reinforcement learning)
35
37. Possible Applications: Bandit Setting
Maximizing outcome from multiple slot machines
with estimated distribution.
36
38. Possible Applications: Bandit Setting
Highest predicted outcome? or Lowest prediction uncertainty?
Choose highest predicted outcome? or explore more samples?
(Exploitation-exploration tradeoff)
37
39. Mice Skin Tumor Treatment
Mice with induced cancer tumors.
Treatment options:
∙ No threatment
∙ 5-FU (100mg/kg)
∙ imiquimod (8mg/kg)
∙ combination of imiquimod and 5-FU 38
40. Upper Confidence Bound
Treatment selection policy
at = arg max
a∈A
[µa(xt) + βσ2
a(xt)]
Quality measure
R(T) =
T∑
t
[max
a∈A
µa(xt) − µa(xt)]
where A is the set of possible treatments
µ(x), σ2
(x) is the predicted mean, variance at x
39
41. Upper Confidence Bound
Treatment based on a Bayesian method (Gaussian Process) lead to
longest life expectancy.
12
12Contextual Bandits for Adapting Treatment in a Mouse Model of de Novo
Carcinogenesis, A. Durand, C. Achilleos, D. Iacovides, K. Strati, G. D. Mitsis, and J.
Pineau, MLHC 2018
40
42. References
∙ Murphy, Kevin P. ”Machine learning: a probabilistic perspective.”
(2012).
∙ Yarin Gal, ”Uncertainty in Deep Learning”, Ph.D Thesis (2016)
∙ Blundell, Charles, et al. ”Weight uncertainty in neural networks.”
arXiv preprint arXiv:1505.05424 (2015).
∙ Gal, Yarin, and Zoubin Ghahramani. ”Dropout as a Bayesian
approximation: Representing model uncertainty in deep learning.”
international conference on machine learning. 2016.
∙ Kendall, Alex, and Yarin Gal. ”What uncertainties do we need in
bayesian deep learning for computer vision?.” Advances in neural
information processing systems. 2017.
41
43. References
∙ Leibig, Christian, et al. ”Leveraging uncertainty information from
deep neural networks for disease detection.” Scientific reports 7.1
(2017): 17816.
∙ Contextual Bandits for Adapting Treatment in a Mouse Model of de
Novo Carcinogenesis A. Durand, C. Achilleos, D. Iacovides, K. Strati,
G. D. Mitsis, and J. Pineau Machine Learning for Healthcare
Conference (MLHC)
∙ Su, Jiawei, Danilo Vasconcellos Vargas, and Sakurai Kouichi. ”One
pixel attack forfooling deep neural networks.” arXiv preprint
arXiv:1710.08864 (2017).
42