Hello and welcome to this video on going beyond images. In this video, we will discuss other uses for convolutional neural networks and we'll also explain the limits of convolutional neural networks. CNNs can be used on sound using a technique called the spectrogram. With a spectrogram we represent the sound on two axes. The vertical axis corresponds to the frequency bands while the horizontal axis indicates the time steps along the sound. At this point we can feed the spectrogram to a convolutional layer and treat it like an image. Some of the most famous speech recognition engines use the technique to obtain great results.

Similarly, we can map a sentence of texts onto an image where the vertical axis indicates the word index in a dictionary and the horizontal axis is for the position of the word in the sentence. Although they're very powerful, CNNs are not useful in some cases. Since they are good at capturing spatial patterns, they are of no use when such local patterns don't exist. Take for example a two-dimensional table with demographic user data from your web application. Each row corresponds to a user and each column to a feature such as the age, the email or how long the person has spent on the website. Since there is no special order in the columns or the rows, we can swap the order of rows and columns and still make perfect sense of the data. In a case like this, a CNN is completely useless because there are no local patterns to exploit with the convolutional filter. In conclusion, in this video we talked about how to use convolutional neural networks on data that is different from images, for example sound and text. We've also talked about types of data where the convolutional neural networks are completely useless, for example database tables of user data. So, thank you for watching and see you in the next video.