Uncertainty estimation for Neural Network — Dropout as Bayesian ApproximationNokBlockedUnblockFollowFollowingJan 28The key theme of this article is, you can use dropout to create prediction confidence.
This article is mainly about how I start with the Uber’s paper Deep and Confident Prediction for Time Series at Uber.
Model interpretation with neural networks has not been an easy task, knowing the confidence of a neural network could be very important for business.
Despite a lot of complicated proof in these series of papers, they all trying to answer a simple questionHow confident is my model about a particular prediction?BackgroundsWe are dealing with some forecasting problem for business, so we are researching new methods for forecasting especially new approach to neural networks.
(I am aware of the LSTM/encoder-decoder/seq2seq but I have not heard a lot of exciting performance from it).
There are not many public discussions about Time Series from the big company that I can find, the closest thing I can find is Prophet from Facebook, which is a purely statistical base forecasting method.
Uber is relatively outstanding as the paper Deep and Confident Prediction for Time Series at Uber draw our attention as it offers an easy way to estimate the model uncertainty, which could be as important as the accuracy itself.
After some research, I find that Uber has also won the M4 competition (a well-known time series competition) The M4 competition does not only requires an accurate prediction, but also the confidence interval of it.
They have also open source Pyro, a probabilistic programming library, so combining these facts, I have reasonable faith that they are doing some interesting(and practically useful) works.
I do not have a deep statistic background and the word “Bayesian” is almost meaningless to me, all I know is that it relates to conditional probability, that’s it.
I try very hard to recall my statistic knowledge about Bayesian vs frequentist (which I think the majority of machine learning/deep learning is about), this fantastic thread explains it quite well.
The PaperUber has already written a blog post about this work, I start with the 1st paper but end up reading a few more papers (some are earlier groundworks, some are follow up works).
I will not spend too many words to explain the papers, as I could not understand every derives steps.
Instead, I will only highlight the important parts of my research journey and I encourage you to go through the paper.
We show that the use of dropout (and its variants) in NNs can be interpreted as a Bayesian approximation of a well known probabilistic model: the Gaussian process (GP)¹I personally do not 100% convince by the work, but empirically it works well and that is the most important part.
(Like BatchNorm people have been thinking why it works for a wrong reason for a long time, but that does not stop anyone using it as it did improve model convergence)Studied Papers:Deep and Confident Prediction for Time Series at UberTime-series Extreme Event Forecasting with Neural Networks at UberDropout as a Bayesian Approximation: Representing Model Uncertainty in Deep LearningVariational Bayesian dropout: pitfalls and fixesVariational Gaussian Dropout is not BayesianThe M4 Competition: Results, findings, conclusion and way forwardUncertaintyOne of the key distinction about Bayesian is that parameters are distributions instead of fixed weights.
Error = Model Uncertainty + Model misspecification + inherent noiseThe Bayesian neural network decomposes uncertainty into model uncertainty, model misspecification, and inherent noise.
MCDropoutMCDropoutOne of the key here in Bayesian is that everything is a probabilistic distribution but not a point estimate.
This means there are uncertainties about your weight as well.
They use MCDropout to deal with model uncertainty and misspecification.
Basically, they have claimed that using Dropout at inference time is equivalent to doing Bayesian approximation.
The key idea here is letting dropout doing the same thing in both training and testing time.
At test time, you will repeat B times (Few hundreds of times as the paper said), i.
passing the same input to the network with random dropout.
You then take means of your prediction and you can generate a prediction interval with these # of B predictions.
MC is referring to Monte Carlo as the dropout process is similar to sampling the neurons.
Inherent NoiseInherent NoiseThey also introduce the term Inherent Noise which refer to noise that is irreducible.
In short, they use a very common technique to model this error — held-out validation.
They call this an adaptive approach and talk about smoothness and prior, but I don’t see any difference from the standard train/validation practice that the ML community is familiar with.
In the end, you will combine the two error terms and get the final uncertainty term.
DiscussionYou can find an interesting discussion on Reddit, which provide some counterargument from a theoretical standpoint.
In fact, I am not 100% convinced by Uber’s paper.
However, they have shown good result in both internally and M4 competition.
Like a lot of advance in deep learning, theory come later than practical results.
If you are interested, feel free to try it out, the implementation should be relatively easy as you just need to keep dropout at inference time.
ConclusionThe takeaway is, uncertainty exists not only in your model, but your weight as well.
Bayesian Neural Network tries to model the weights as distributions.
MCDropout offer a new and handy way to estimate uncertainty with minimal changes in most existing networks.
In the simplest case, you just need to keep your dropout on at test time, then pass the data multiple times and store all the predictions.
The downside is, this could be computationally expensive, although Uber claims this adds less than ten milliseconds.
They didn’t discuss how they achieve this, but my guess is they do heavily parallel computation as you can imagine the multiple passes of data does not come in sequential order, so this process can be parallel easily.
I am very eager to hear more discussion about the latest time series forecast and method estimating uncertainty with NN, let me know if you know some better way!Appendixhttps://tensorchiefs.
pdf (Yarin Gal).. More details