Sequences, Time Series and Prediction

This is the last course in the TensorFlow in Practice specialisation leading to a professional certificate.

For me this was the most rewarding and best of the 4 courses in this specialization. There are plenty of courses on neural networks and computer vision but very few on time series. If you want to predict time series don't look any further, this is a course for you.

You start with learning the basic components of time series: trends, seasonality, autocorrelation and noise. You will build a time series with synthetic data and as a first exercise you use classic statistics (differencing, moving averages, smoothing, trailing and centered windows) to extract the components out of the time series.

You will see how to apply fixed and roll-forward partitioning to build your training, validation and test sets. The different error types are handled (mse, rmse, mae, mape).

In week 2 and 3 you will apply neural network techniques with TensorFlow to analyse and predict the same time series: preparing features and labels, batching the data, shuffling to avoid sequence bias. You start with a simple model of a single dense layer, visualize the results and evaluate using mean absolute errors. You further improve the model by adding more layers (additional dense layer, RNN, LSTM, lambda layers), different loss functions like Huber, dynamically adapting the learning rate.

In the final week you apply what you have learned on real-world data and experiment on how to improve by gradually adapting the model and its hyperparameters, for example the impact of batch size on the stability of the training loss.

Completing this professional certification made me confident to start creating TesorFlow models and prediction on my own. Another great set of courses by deeplearning.ai !