Hello! And welcome to this video on feature engineering. In this video, we will talk about what feature engineering is, we will talk about problems with traditional approaches to feature engineering, and we will talk about the advantage of using deep learning for feature engineering. As we've seen, unstructured data does not look like tabular data, and the traditional solution to connect the two is feature engineering. However, feature engineering comes with some limitation. In feature engineering, we need an expert that uses his or her domain knowledge to create features that correctly encapsulate the relevant information from the unstructured data. 

Feature engineering is fundamental to the application of machine learning, but it is both difficult and expensive. Let's see a few examples. If we're training a machine learning model on face recognition task, we could encode the information on the face by using a well-tested method that consists in identifying key points in the face, and measure the distance between such points. In this way, we went from a 2-D image of a face to features that represent key distances of elements of that face. These distances are then the features passed to the machine learning model that will recognize the faces. In a similar way, in the domain of speech recognition, we can build features based on wavelet transforms, or Fourier transform, or short-time Fourier transform, and these were the standard until not very long ago. Deep learning has disrupted feature engineering by introducing a way to learn the best features directly from the raw, unstructured data. 

This approach is not only very powerful, but also much, much faster. It's kind of a paradigm shift. It's a more versatile technique, that takes the role of the domain expert. So, in this video, we've discussed traditional feature engineering, and how it takes time and expertise, and we've highlighted the fact that deep learning is automated feature engineering. So, thank you for watching, and see you in the next video.