# What are the characteristics of a deep learning algorithm, apart the fact of being a multilayer network?

I have read a lot about Deep learning, and I'm little bit confused. Most of multilayer Machine Learning algorithms are considered as Deep Learning algorithm. I don't get it, does being a multilayer architecture is what characterizes a Deep learning algorithm or there's something else that I'm missing?

The major advantage of using a multi-layered perceptron is that it is a non linear classifier.

Every neuron in the current layer can be computed using the following formula:

$$h_i(x_t)=f_i(\mathbf{x}_{t-1})$$

where $\mathbf{x}_{t-1}$ is the subset of the input neurons that will participate in the current neuron computation, and $f_i$ is a non linear function.

Generally, the transformation is $g_i(\mathbf{wx_{t-1}+b})$ where $\mathbf{w}$ is the weight that is multiplied to the input neurons to the current layer, and $\mathbf{b}$ is the optional bias. This is however, a linear transormation. The main difference that makes a neural network (or for that matter any multi-layered perceptron) is the non linear function $g_i$ which essentially determines if the neuron under consideration will "$\textit{fire}$" or not.

Additionally, in the application of computer vision, early layers determine the low level features such as edges in the input image, and later layers determine the higher level features such as objects, presence of a particular color etc.

Yes, the multilayer architecture is the thing which makes it "deep". Even in a single layer you can have a non linear output. But with a multi layer you can use this hidden layers output as in input to the next layer, which gives it the tremendous ability to develop feature representations in its internal layers. Hence , it might develop representations like "edges" in an image in its internal layers.

• isn't deep learning : the application of the gradient calculation method based on backpropagation on a deep neural network!? – Yasmina Jan 7 '18 at 1:46