Question

Explain what is meant when people refer to ‘the Deep Learning’ revolution in Machine Learning and Artificial Intelligence. Do you think it is just ‘Hype’ (unjustified or exaggerated ‘grandstanding’), or do you feel there is something significant happening with it? Justify your answer using terms such as ‘feature map’ and ‘supervised’ and ‘unsupervised learning’. How do people overcome challenges of the scale of the numerical optimization and the large-parameter-related generalization issues which arise?

Answer 1

The Deep learning revolution in Machine Learning and Artificial Intelligence is not hyped . Deep learning is the use of deep networks to make intelligent systems, or to predict something which might be useful to us. While machine learning algorithms may be supervised ( classification algorithms like Random forest, SVM, Naive Bayes etc.) or unsupervised (K means clustering etc. i.e in which labels of classes are not required to be known), deep learning makes use of neutral networks with hidden layers. These layers extract and create feature maps from the given data. Based on the features extracted by each hidden layer, we can use the final layers for tasks like classification, regression and reinforcement learning. Thus deep networks definitely perform better as compared to traditional machine learning algorithms. Consider an example of object recognition. We can train a type of neural network called Convolution Neural Network with the help of labeled images. Using convolution, feature maps are extracted followed by max pooling and fully connected layers. Finally a classifier can be used to identify an unknown object. Thus supervised learning is done here using a deep network. And the performance achieved is superior to many state of the art methods.

Challenges of the scale of numerical optimization and large parameter related issues can be overcome with the help of fast processors and fine libraries available these days. Talking about generalization issues, there are various ways by which we can prevent issues like overfitting. We can tune the parameters of a deep network, for example the optimization algorithm to be used, reduction in the number of features, loss function to be used etc. With good hyperparameter tuning, the network can adapt well and generalize to unseen data . By mapping vectorized data to various processors, it is also possible to get over numerical optimization challenges using deep learning.