Enroll Course: https://www.udemy.com/course/modeles-avances-des-series-temporelles-avec-keras-partie-2/
For those looking to push the boundaries of time series forecasting beyond the foundational DA-RNN model, Udemy’s ‘Prédiction des séries temporelles en deep learning – Partie 2’ offers a comprehensive and advanced curriculum. Building directly on its predecessor, this course delves into cutting-edge models like DSTP-RNN, HRHN, STAM, and WaveNet, all derived from recent scientific research and specifically designed for deep learning-based time series prediction using Python.
The course excels in its practical approach, leveraging popular libraries such as Tensorflow, Keras, Pandas, Numpy, and Scikit-learn. A key highlight is the integration of Raytune for hyperparameter optimization, showcasing advanced techniques like Bayesian optimization and schedulers. The entire learning experience is designed for accessibility, with all work conducted online via Jupyter notebooks on Google Colab, eliminating the need for any local software installations.
Each module introduces new concepts with clear explanations, covering nine distinct study themes:
1. **DSTP-RNN:** An evolution of DA-RNN, this multivariate Seq2Seq model with spatio-temporal attention addresses the limitations of using current exogenous series information for predictions, significantly improving performance.
2. **HRHN:** This model utilizes Recurrent Highway Networks (RHN) and 1D convolutional layers, addressing deep network convergence issues and enhancing spatial attention.
3. **PID Controller for Error Compensation:** The course explores the application of Proportional-Integral-Derivative (PID) controllers, widely used in industry, to enhance the performance of DSTP-RNN and HRHN models.
4. **Raytune for Hyperparameter Optimization:** Moving beyond manual or grid search, Raytune is introduced for efficient hyperparameter tuning using algorithms like Hyperband and Bayesian optimization, with Tensorboard for monitoring and Google Drive synchronization.
5. **Random Forest for Prediction:** The power of Random Forest is explored for prediction tasks, especially beneficial when dealing with a large number of exogenous variables.
6. **Random Forest for Variable Selection:** The course demonstrates how Random Forest can be used to identify the most important variables in multivariate series, reducing computational time and improving model focus, including techniques like Recursive Feature Selection (RFE).
7. **VSURF with Google Cloud:** Leveraging the VSURF library in R on a dedicated Google Cloud server for parallel processing, this section focuses on variable importance identification for both regression and classification tasks.
8. **STAM Model:** This recent (2020) spatio-temporal attention model is causal and evolving, offering reliable insights for variable selection.
9. **WaveNet Model:** Originally designed for audio generation, Google’s WaveNet is explored for its successful applications in time series prediction, including coding its architecture for forecasting.
With over 9 hours of content, the course empowers learners to confidently code custom layers and models using Keras and Tensorflow through class inheritance. The curriculum is grounded in scientific research, introducing state-of-the-art concepts.
**Prerequisites:** While the course is a direct continuation of ‘Part 1’, individuals with existing Deep Learning experience will find new valuable insights. A foundational understanding of Deep Learning is recommended.
**Support:** The instructor provides excellent online support, acknowledging that learners may have varying backgrounds in mathematics and programming. Questions are encouraged and answered promptly, fostering a supportive learning environment.
**Recommendation:** This course is highly recommended for anyone serious about mastering advanced time series prediction techniques with deep learning. Its blend of theoretical depth, practical implementation with cutting-edge models, and accessible online environment makes it an invaluable resource for data scientists and machine learning engineers looking to elevate their forecasting capabilities.
Enroll Course: https://www.udemy.com/course/modeles-avances-des-series-temporelles-avec-keras-partie-2/