In this chapter, you have seen the building blocks of computer vision models. We've learned about convolutional layers, and both the ReLU activation and regularization methods. You have also seen a number of ways to use neural networks creatively, such as with Siamese networks and bounding box predictors.

You have also successfully implemented and tested all these approaches on a simple benchmark task, the MNIST dataset. We scaled up our training and used a pretrained VGG model to classify thousands of plant images, before then using a Keras generator to load images from disk on the fly and customizing the VGG model to fit our new task.

We also learned about the importance of image augmentation and the modularity tradeoff in building computer vision models. Many of these building blocks, such as convolutions, batchnorm, and dropout, are used in other areas beyond computer vision. They are fundamental tools that you will see outside of computer vision applications as well. By learning about them here, you have set yourself up to discover a wide range of possibilities in, for example, time series or generative models.

Computer vision has many applications in the financial industry, especially in back-office functions as well as alternative alpha generation. It is one application of modern machine learning that can translate into real value for many corporations today. An increasing number of firms incorporate image-based data sources in their decision making; you are now prepared to tackle such problems head-on.

Over the course of this chapter, we've seen that an entire pipeline is involved in a successful computer vision project, and working on the pipeline often has a similar or greater benefit as compared to working on the model.

In the next chapter, we will look at the most iconic and common form of financial data: time series. We will tackle the task of forecasting web traffic using more traditional statistical methods, such as ARIMA (short for AutoRegressive Integrated Moving Average), as well as modern neural network-based approaches. You will also learn about feature engineering with autocorrelation and Fourier transformations. Finally, you will learn how to compare and contrast different forecasting methods and build a high-quality forecasting system.