Looking beyond LSTMs: Alternatives to Time Series Modelling using Neural Nets
summary
Claims that convolutional architectures beat LSTMs on noisy physiological vitals: experiments on the MIMIC dataset use two hours of inputs (150 samples) to predict deterioration in the next ~20–25 minutes, with a 1D CNN hitting ~87% AUC versus LSTM ~80% AUC and training time about one quarter as long. A Temporal Convolutional Network reaches ~91% for highest-risk cases; a simple discretized bag-of-words + gradient boosting baseline scores ~74% F1. Results drive a beta deployment in Bangalore hospitals.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:07Outline
- 1:36Time Series Data
- 2:27Machine Learning in Time Series
- 4:17Disadvantages of LSTM
- 5:09Convolutional Networks
- 8:15Temporal Cognitive Networks
- 11:05Results
- 12:18Data Analysis
- 13:03Bagofwords
- 13:55Converting Time Series into Images
- 15:05Summary
- 16:45QuestionsComments
- 17:45Piecewise Aggregation
related talks
36m1D Convolutional Neural Networks for Time Series Modeling
31mAdvanced machine learning time series....
31mThe Advantages of Deep Learning with Recurrent Neural Networks
06mExtraction and Interpretation of Deep Autoencoder-based Temporal Features from Wearables for ...
31mML time series forecasting methods
37mRecent Solutions Using Deep Learning Models/ The How and Why
32mA word is worth a thousand pictures: Convolutional methods for text
43mThe Brain’s Guide to Dealing with Context in Language Understanding