Deep learning is a major driving force in the field of machine learning.The term ‘Deep Learning’ refers to building layered and complex hierarchical representations from simple building blocks. These building blocks are called perceptrons which can have ‘n’ inputs and one output where we require (n + 1) parameters. These perceptrons are layered on top of each other to get non-linearities. Deep neural networks usually have six to eight layers and the lower layers most often identify base features and edges and the subsequent layers continue to detect more complex features. A deep network can have thousands of outputs, one for each label that is built on top of a common deep network underneath it, which hopefully identifies useful high level representation needed for analysis. It has similarities to the human visual cortex that has a similar type of hierarchical structure with neurons on lower layers recognising lower level features.
It is important to note that perceptrons have linear and non-linear parts. The non-linearity property allows deep learning to be functionally expressive and enables layering. Modern deep learning networks have millions of parameters that provide it a large search space. The best way to understand the intricacies of the artificial neural network is to start with the workings of a neuron and how data enters it. An output of each neuron depends on several factors like the input, the random weight, the random bias and the activation function. An artificial neural network is a sequential network of interlinked neurons, which can be understood as the output of one neuron will be the input of another neuron. These networks learn to understand patterns out of the training data and train itself to become a prediction model. They continue to evolve to provide lesser and lesser error rates in prediction models and have achieved almost human-level accurate results in several cases.
- Tensorflow
- Keras
- SKLearn
- PyTorch
- DL4J
- Caffe
- Microsoft Cognitive Toolkit