Deep Learning is a branch of Machine Learning, it is a family of techniques you can use to solve different problems. It is currently the state-of-the-art in computer vision, speech recognition, and understanding text documents for example.
Deep Learning Techniques shines when there is a lot of data to perform the learning. You may see Deep Learning as a rebrand of Neuronal Networks (NN exists from decades ago) but Deep Learning is taking off now, basically because we have big data to train and better/affordable computational power, in that case you should have better performance in your AI models with Deep Learning algorithms than other Machine Learning techniques.
I think that in traditional classification approaches we start by extracting significant features from the images (for example), then we classify the datasets based on the extracted features. But for the deep learning the whole signal (image for example) is directly used for the separation between the dataset categories.
Deep learning (DL) represents a new promising trend in machine learning. Recently, DL algorithms have achieved the state-of-the-art in many domains, particularly in computer vision, by giving spectacular results comparing to classic machine learning algorithms. These DL algorithms are different from classic machine learning algorithms in the following points:
1) - Data Consumption: The supervised training of DL classifiers requires a large number of labeled examples, for this reason, data availability in the last decade contributes to DL success. DL classifiers require a huge training set because these classifiers contain a large number of parameters to tune. This constraint of labeled data represents a limiting factor when the labeling is expensive. For example, the biological labeled examples are expensive and difficult to collect in most of the cases.
2)- Dedicated Hardware: The training phase of DL classifiers requires a dedicated hardware like the Graphics Processing Units (GPUs) to reduce execution time. These GPUs represent an essential component in DL approaches and training without GPUs leads to further many days of training.
3)- Feature Extraction: Machine Learning algorithms contain feature engineering phase. In this phase, experts propose the hand-crafted features to facilitate the learning from examples. This phase is very important and affects the overall performance of the learning system. Unfortunately, feature engineering is a manual component in machine learning pipeline and it is time-consuming. On the other hand, in a DL pipeline, feature extraction is embedded in the learning algorithm where features are extracted in a fully automated way and without any intervention of a human expert. CNN represents a good example of automatic feature extraction in computer vision.
There is no clear definition in literature about what a DL is. Generally speaking DL are artificial Neural Networks with many layers. Some authors tried to use the capital assignment path as measure to distinguish shallow from DL. Others say that having more than 3 layers qualifies as DL.
Like Andreas Theissler mentioned, Deep Learning is a sub class of Machine Learning. Deep Learning is basically "Deep" Artificial Neural Networks (with many hidden layers) as opposed to "Shallow" neural networks with just one hidden layer. One major difference between DL and traditional ML is that it does not require "Feature Engineering". That is the raw data or signal is fed directly to the Deep Neural Network. Whereas in traditional ML, we first extract useful features, and then feed them to the ML system. Given enough data, and enough layers, deep neural networks can learn the useful features by themselves. As deep neural networks have many hidden layers (i.e. many parameters or weights), they need a lot more data and lot more computational power for training or learning the optimal parameters (e.g. weights). However, given enough data for training (millions of labeled examples), deep neural networks usually outperform other ML systems like SVM, Random Forest, KNN etc. Given enough data, adding more hidden layers usually results into better performance. This is one of the major reasons why deep learning has been so popular lately.
Deep Neural Learning (DDLs) Methodology is one logical architecture of ML. It is based on the concept of perusing thousands of images for the ML to learn. Unfortunately, based on Uber self-autonomous vehicle accidents, such concept has many shortcomings, per Dr. Higgins of University of Toronto University and also top Google advisor. This is why he and other top ML experts are entertaining the concept of the Convolutional Neural Networks (CNNs) which is based on the old-fashion sampling approach. I just completed an article which considers Sabotage by Hallucination as the common impairment of elder folks and Uber/Tesla vehicles.