Basics for Beginners — Paradigms of Learning — Part 3
Broadly there are 3 — types of learning. A small description of all the methods are mentioned in this blog.
In the previous blogs, we understood about Machine Learning and Neural networks. But how exactly do we train the neural networks and what are the types of learnings that are available ? Why are these learnings categorised in the way they are ? The answers to these can be found in this blog. Let us dive into the topics already. The blog is short and sweet and pretty direct to the point.
Ability to learn has been one of the most impressive features of Artificial Neural Networks (ANN).The word learning, in context of neural networks refers to improvement of performance of given task or range of tasks with experience. The neural networks are influenced by the biological nervous system particularly human brain. Ability to learn from experiences has been major feature of human brain. Neural network resembles the same. A simplified model of the neural network would be able to mimic this by adjusting the weighted connections found between neurons in the network. The network learns by stronger connections of synaptic informations whereas the less relevant information will gradually have its synaptic connections weaken, making it harder to recall (this is with respect to brain in analogy with weights as neurons). There are unique advantages and disadvantages with the different learning techniques. With the different learning techniques the weights are altered in different ways. Thus it depends on algorithm technique employed to make the network learn. The objective is to get a correct output when input is passed by having a set of weight matrices applied to the network. A short description about learning is explained below and same is shown pictorially in Fig.
Supervised Learning:
If the desired output for the network is also provided with the input during learning process, then the algorithm falls under Supervised Learning. With the input and its corresponding output is paired and weights are calculated during learning. Error calculation based on it is target output and actual output is done in order to make corrections to the network by updating it is weights. Here in this student project, supervised learning technique is used.
Unsupervised Learning:
In this algorithm no supervision is done with respect to output, that is, the neural network will be responsible to and some kind of pattern within the inputs provided without any external aid. This algorithm can be found using in data mining applications, where in neural network is expected to predict user’s preferences based on the preferences of other similar users it has grouped together.
Reinforcement Learning:
This learning is similar to supervised learning. In supervised learning some feedback is given, however rewards are provided to the network about its performance rather than target output. The goal of this learning process is to maximise the reward that system receives through trial-and-error. This paradigm process can be related to learning from nature. For example an animal might remember the actions it’s previously taken which helped it to and food (the reward).
