The Advantages of Deep Learning with Recurrent Neural Networks
summary
Realtime anomaly detection using LSTM RNNs frames forecasting as predict-and-analyze residuals where each metric’s next-step forecast is compared to the observed value and residuals are scored (univariate or multivariate) with GMM or smoothing layers applied to reduce daily spikes. The architecture examples include stacked LSTMs, encoder/decoder reconstruction for outlier scoring, and LSTM variants with a phase gate to handle irregular telemetry. Practical fixes: dropout only on vertical connections, multiple predictors to avoid quick “new normal” drift, and retrain after redacting confirmed anomalies.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:27Agenda 1. Problem Landscape II. Traditional Strategies
- 1:08Anomaly Detection: Problem Landscape
- 2:11Supervised or Unsupervised? Anomaly Detection is an exercise in Finding subsets of the data that deviate from the pattern exhibited by the majority of the data.
- 2:45Multivariate Target or Univariate Target
- 4:00Static Data or Time Series
- 5:41Traditional Tactic: Gaussian Mixture Models
- 11:17Recurrent Neural Networks (RNNS)
- 19:46Common Issues: LSTM is "too good" Solutions 1. Change error metric
- 21:24Encoder-Decoder Architecture
- 23:35Common Issues: Single step residual spikes. Solution: Smooth the residuals before analysis: • Moving Average, Exponential smoothing, Autoregression
related talks
34mDeep NLP - Redes neurais revolucionando o processamento de linguagem natural
27mMore Then Words : Business Applications of Recurrent Networks
30mRealtime Time Series Anomaly Detection in Production
25mLooking beyond LSTMs: Alternatives to Time Series Modelling using Neural Nets
43mPersonalization and Scalable Deep Learning with
53mTensorflow, deep learning and modern RNN architectures, without a PhD
38mUsing Recurrent Neuronal Networks for Customer Predictions
26mProactive Failure Prediction Through Anomaly Detection