deep recurrent neural networks for sequence learning in spark by yves mabiala
TRANSCRIPT
Deep recurrent neural network for sequence learning in Spark
Yves MABIALATHALES
Outline• Thales & Big Data• On the difficulty of Sequence Learning• Deep Learning for Sequence Learning• Spark implementation of Deep Learning• Use cases
– Predictive maintenance– NLP
Thales & Big Data
Thales systems produce a huge quantity of dataTransportation systems (ticketing, supervision, …)Security (radar traces, network logs, …)Satellite (photos, videos, …)
which is oftenMassiveHeterogeneousExtremely dynamic
and where understanding the dynamics of the monitored phenomena is mandatory Sequence Learning
What is sequence learning ?Sequence learning refers to a set of ML tasks where a model has to either deal with sequences as input, produce sequences as output or both
Goal : Understand the dynamic of a sequence to– Classify– Predict– Model
Typical applications– Text
• Classify texts (sentiment analysis)• Generate textual description of images (image captioning)
– Video• Video classification
– Speech• Speech to text
How is it typically handled ?Taking into account the dynamic is difficult
– Often people do not bother• E.g. text analysis using bag of word (one hot encoding)
– Problem for certain tasks such as sentiment classification (order of the words is important)
– Or use popular statistical approaches • (Hidden) Markov model for prediction (and classification)
– Short term dependency (order 1) : 𝑃(𝑋$ = 𝑥 (𝑋$'( = 𝑥$'(,… , 𝑋$', = 𝑥$',)⁄ ) = 𝑃(𝑋$ = 𝑥$ 𝑋$'( = 𝑥$'()⁄
• Autoregressive approaches for time series forecasting
The chair is red 1 0 1 1 0 0 0 0The cat is on a chair
The cat is young 1 1 0 0 1 1 0 0
1 1 1 0 0 1 1 1The is chair red young cat on a
Link with artificial neural network ?Artificial neural network is a set of statistical models inspired from the brain
– Transforms the input by applying at each layer (non linear) functions– More layers equals more capabilities (≥ 2hidden layers : Deep Learning)
• From manual features building to feature learning
Set of transformation and activation operations– Affine : 𝒀 = 𝑾𝒕𝑿+𝒃, sigmoid activation : 𝟏
𝟏8𝐞𝐱𝐩('𝑿), tanh activation : 𝒀 = 𝐭𝐚𝐧𝐡(𝑿)
• Only affine + activation layers = multi layer perceptron (available in Spark ML since 1.5.0)
– Convolutional : Apply a spatial convolution on the 1D/2D input (signal, image, …) : 𝐘 = 𝒄𝒐𝒏𝒗 𝑿,𝑾 +𝒃• Learns spatial features used for classification (images) , prediction
– Recurrent : Introduces a recurrent part to learn dependencies between observations (features related to the dynamic)
Objective– Find the best weights W to minimize the difference between the predicted output and the desired one
(using back-propagation algorithm)
inputhidden layers
output
Able to cope with varying size sequences either at the input or at the output
Recurrent Neural Network basics
One to many (fixed size input, sequence output)
e.g. Image captioning
Many to many(sequence input to sequence
output)
e.g. Speech to text
Many to one(sequence input to fixed size
output)e.g. Text classification
Artificial neural networks with one or more recurrent layers
Classical neural network Recurrent neural network
𝒀𝒌'𝟑 𝒀𝒌'𝟐 𝒀𝒌'𝟏 𝒀𝒌𝒀𝒌
𝑿𝒌'𝟑 𝑿𝒌'𝟐 𝑿𝒌'𝟏 𝑿𝒌𝒀𝒌 = 𝒇(𝑾𝒕𝑿𝒌 +𝑯𝒀𝒌'𝟏)
𝑿𝒌𝑿
𝒀𝒌 = 𝒇(𝑾𝒕𝑿𝒌)
𝒀
Unrolled through time
𝒀𝒌'𝟑 𝒀𝒌'𝟐 𝒀𝒌'𝟏 𝒀𝒌
𝑿
𝒀𝒌'𝟑 𝒀𝒌'𝟐 𝒀𝒌'𝟏 𝒀𝒌
𝑿𝒌'𝟑 𝑿𝒌'𝟐 𝑿𝒌'𝟏 𝑿𝒌𝑿𝒌'𝟑 𝑿𝒌'𝟐 𝑿𝒌'𝟏 𝑿𝒌
𝒀
On the difficulty of training recurrent networksRNNs are (were) known to be difficult to learn
– More weights and more computational steps • More computationally expensive (accelerator needed for matrix ops : Blas or GPU)• More data needed to converge (scalability over Big Data architectures : Spark)
– Theano, Tensor Flow, Caffe do not have distributed versions
– Unable to learn long range dependencies (Graves & Al 2014)• At a given time t, RNN does not remember the observations before 𝑋J',
⇒ New RNN architectures with memory preservation (more context)𝑍$ = 𝑓 𝑊NO𝑋$ + 𝐻N𝑌$'(𝑅$ = 𝑓(𝑊SO𝑋$ +𝐻S𝑌$'()
𝐻T$ = tanh(𝑊YJZ[\O𝑋$ +𝑈 𝑌$'(o𝑅$ )
𝑌$ = 1 − 𝑍$ 𝑌$'(+ 𝑍$𝐻T$LSTM GRU
Recurrent neural networks in SparkSpark implementation of DL algorithms (data parallel)
– All the needed blocks• Affine, convolutional, recurrent layers (Simple and GRU)• Sigmoid, tanh, reLU activations• SGD, rmsprop, adadelta optimizers
– CPU (and GPU backend)– Fully compatible with existing DL library in Spark ML
Performance– On 6 nodes cluster (CPU)
• 5.46 average speedup (some communication overhead)– About the same speedup as MLP in Spark ML
Driver
Worker 1
Worker 2
Worker 3
Resulting gradients (2)
Model broadcast (1)
Use case 1 : predictive maintenance (1)Context
– Thales and its clients build systems in different domains• Transportation (ticketing, controlling)• Defense (radar)• Satellites
– Need better and more accurate maintenance services• From planned maintenance (every x days) to an alert maintenance• From expert detection to automatic failure prediction• From whole subsystem changes to more localized reparations
Goal– Detect early signs of a (sub)system failure using data coming
from sensors monitoring the health of a system (HUMS)
Use case 1 : predictive maintenance (2)Example on a real system
– 20 sensors (20 values every 5 minutes), label (failure or not)
– Take 3 hours of data and predict the probability of failure in the next hour (fully customizable)
Learning using MLLIB
Use case 1 : predictive maintenance (3)Recurrent net learning
Impact of recurrent nets– Logistic regression
• 70% detection with 70% accuracy– Recurrent Neural Network
• 85% detection with 75% accuracy
Use case 2 : Sentiment analysis (1)
Context– Social network analysis application developed at Thales (Twitter, Facebook,
blogs, forums)• Analyze both the content of the texts and the relations (texts, actors)
– Multiple (big data) analysis • Actor community detection• Text clustering (themes)• …
Focus on– Sentiment analysis on the collected texts
• Classify texts based on their sentiment
Use case 2 : Sentiment analysis (2)Learning dataset
– Sentiment140 + Kaggle challenge (1.5M labeled tweets)– 50% positives, 50% negatives
Compare Bag of words + classifier approaches (Naïve Bayes, SVM, logistic regression) versus RNN
Use case 2 : Sentiment analysis (3)
NB SVM LogReg
Neural Net (perceptron) RNN (GRU)
100 61.4 58.4 58.4 55.6 NA
1 000 70.6 70.6 70.6 70.8 68.1
10 000 75.4 75.1 75.4 76.1 72.3
100 000 78.1 76.6 76.9 78.5 79.2
700 000 80 78.3 78.3 80 84.1
Results
4045505560657075808590 NB
SVM
LogReg
NeuralNet
RNN (GRU)
The end…
THANK YOU !