Machine Learning, let’s divide the words. Machine means a device or instrument that easy our works, and Learning is to learn something. So machine learning is teaching your machine about something. You collect data, clean up data, construct an algorithm, train the algorithm basic data patterns, and then ask the algorithm to give you a reasonable answer.
If the algorithm is according to your standards, you’ve taught your algorithm correctly. If not, scrap it all and start from scratch. That’s how it happened here. Yeah, however, if you’re looking for a formal description, Machine Learning is an optimization algorithm that can perform a specific task without the need for a human.
SUPERVISED LEARNING | Machine Learning:
Creating an algorithm to learn to map an input to a given output is supervise Learning. therefore, Supervised Learning is done using the datasets you have obtained that are numbered. If the mapping is right, you have successfully learned the algorithm. Else, you make the algorithm the requisite improvements so that it can understand correctly. Supervised learning algorithms can help to forecast new unknown knowledge that we will gain later in the future. (Caruana et al., 2006)
Types of Supervised Learning
Supervised Learning has been broadly classified into two types.
- Regression
- Classification
regression
Regression is the method of Supervised Learning that is learned from the Defined Datasets and can predict the new data presented to the algorithm for a continuous value output. Whenever an appropriate production is a number such as money or height, comes in use. Some standard algorithms for Supervised Learning are described below: (Zhu et al.,2005)
- Linear Regression: This algorithm implies that the data derived from has a linear relationship between the two variables, Input (X) and Output (Y). The Input variable and the Output variable are thus the Independent Variable and Dependent Variable, respectively.
As unseen information is transmitted to the algorithm, the input is used, measured, and mapped to a continuous value for the output function.
- Logistic Regression: For the set of independent variables that have been transfer to it, this algorithm estimates discrete values. By mapping the new data to the logit feature that has been built into it, it makes the prediction. The algorithm calculates the original data’s probability, so its output lies between 0 and 1.
classification
Classification is the kind of Learning where the algorithm wants to map the new data obtained for each of the two groups in our dataset. The groups must be mapped to either 1 or 0, which is translated into ‘Yes’ or No,” Rains’ or Does Not Rain’ in real-life, and so on. Either one of the groups would be the output and not a sum as it was in Regression. Below, some of the best-known algorithms are discuss:
Decision Trees are classified based on attribute values. They use the Knowledge Gain approach to figure out which dataset function offers the best information making it the root node, and so on, before each instance of the dataset can be categorized. A branch of the Decision Tree defines the process of the dataset. thus, They are one of the most commonly in use classification algorithms.
Algorithms from Naive Bayes presume that the data set properties are all independent of each other. On vast databases, they work well. For classification, Directed Acyclic Graphs (DAG) are use.
therefore, SVM algorithms are based on Vap Nik’s theory of mathematical Learning. For most of the learning exercises, they use Kernal functions that are a core principle. These algorithms construct a hyperplane that is use to differentiate between the two groups. Unsupervised learning-There are no marks in the data gather here, and you are uncertain about the outputs. Then you model the algorithm so that it can recognize the data patterns and produce the response you need. When the algorithm learns, you should not intervene.
UNSUPERVISED LEARNING | Machine Learning:
Let me tell you all about it. Unsupervised learning algorithms operate on unlabel datasets and discover correlations that we were not previously inform about. If we need to categorize the elements or find a relation between them, these acquired habits are beneficial. They can also help spot inconsistencies and errors in the details that we can take care of us. (Barlow, H.B., 1989)
Types:
Unsupervised learning has two types:
- Clustering
- Association
Clustering
Clustering is the method of unsupervised Learning in which the material you are focusing on seeks trends. It may be the form, scale, color, etc., that can be use to group or construct data object clusters.
- Hierarchical Clustering:- Based on the similarities of various data points in the dataset, this algorithm generates clusters. so, It looks at the different aspects of the data points and looks for correlations between them. They are pair together if the data points are consider to be identical. This persists until the dataset that establishes a hierarchy for each of these clusters is cluster together.
- K-Means Clustering:- This algorithm operates step-by-step to achieve clusters with labels to mark them, the critical target. Through calculating the centroid of the cluster, thus the algorithm generates clusters of separate data points that are as homogeneous as possible and guarantees that the difference between this centroid and the current data point is as minimal as possible.
While maintaining that the clusters do not interfere with each other, the smallest distance between the data point and the centroid decides which set it belongs to. The centroid behaves like the cluster’s nucleus. This essentially provides us with a set that can be classified as needed.
- K-NN Clustering:– This is perhaps the easiest of the algorithms for machine learning. The algorithm does not actually learn but instead classifies the new data point depending on the datasets it has stored. This algorithm is often refer to as a lazy learner since it only learns when a unique data point is present to the algorithm. For smaller databases, it fits well, as large datasets take time to understand. (Ghahramani, Z., 2003,)
REINFORCEMENT LEARNING | Machine Learning:
Reinforcement Learning is characterize as a machine learning method that discusses how software agents should perform actions in an environment. ultimately, It is an aspect of the process of deep Learning that lets you optimize some portion of the total reward.
so, This learning approach for the neural network allows you to learn how to accomplish a complex goal or optimize a single dimension over several steps.
How does Reinforcement Learning work?
Let’s look at a quick illustration that lets you explain the process of reinforcement learning. Remember the case of showing the cat new tricks. We can’t tell her precisely what to do because the cat doesn’t understand English or any other human language. Instead, we’re pursuing that approach. We imitate a situation, and in several different ways, the cat tries to respond. If the cat’s reaction is the way we want it to be, we’ll give her the treat.
Still, if the cat is introduce to the same scenario, the cat performs a similar behavior to have more incentive (food) with much more excitement. “It’s like learning that a cat receives through good experiences from “what to do. Around the same moment, when met with traumatic encounters, the cat still knows what to do. (Jaakkola, T et al., 1995.)
Types of Reinforcement Learning:
Two kinds of reinforcement learning methods are:
Positive:
It is defined as an incident that happens because of a particular behavior. This strengthens the strength and duration of the conduct and positively affects the action taken by the agent. however, For a prolonged period, this sort of motivation lets you improve efficiency and manage improvement. Too much reinforcement can lead to over-optimization of the state, which can impact the outcome.
Negative:
Negative reinforcement is characterized as behavioral reinforcement due to a negative situation that should have been prevent or avoid upon. so, This encourages you to establish a minimum level of achievement. The downside to this strategy, though, is that it contains enough to fulfill the minimal behavior.
Learning Models of Reinforcement
so, The two essential learning models in reinforcement learning are:
- Markov Decision Process
- Q learning
Markov Decision Process
The following parameters are used to get a solution:
- Set of Actions-A
- Set of states- S
- Reward- R
- Policy-n
- Value-V
thus, In reinforcement learning, the mathematical approach to mapping a solution is a Markov Decision Process or (MDP).
Q-Learning
Q learning is a method of presenting information dependent on meaning to notify which action an agent should take. Let’s understand this method by the following example:
- In a home, there are five rooms which are connected by doors.
- then, Any room is number from 0 to 4.
- One wide outdoor area could be the outside of the building (5)
- Doors 1 and 4 lead from room five into the house.
First, you need to equate each door with an incentive value:
- Doors leading straight to the target have a reward of 100.
- Doors that are not explicitly linked to the goal space provide zero rewards.
- As doors are two-way, for each room, two arrows are allocate upon. (Gullapalli, V., 1990.)
SUPERVISED LEARNING, UN-SUPERVISED LEARNING, AND REINFORCEMENT LEARNING:
Finally, let’s look at the distinction between supervised unattended and reinforcement learning now that you are well aware of Supervised, Unattended, and Reinforcement learning algorithms.
Supervised Learning, in a nutshell, is when a model learns with instruction from a labeled dataset. And unsupervised Learning is when, without any supervision, the computer is given training based on un-labeled results. Whereas reinforcement learning is when a program or an entity engages with its surroundings, performs actions, and learns through trial and error.
Advantages and disadvantages of supervised Learning
Advantages:
1. Linear Regression is an example that is easy to explain and reasonably clear. thus, To discourage overfitting, it is normalized. Besides, linear models can be modified with new data by merely using stochastic gradient descent.
2. Using well-known and labeled input data makes it much more effective and consistent to deliver supervised Learning than unsupervised Learning. It may be used to boost its output on a particular assignment with access to labels.
3. Effective in discovering solutions to various linear and non-linear problems, such as classification, robotics, prediction, and factory regulation. By having a hidden neuron layer, you can solve complex problems.
Disadvantages:
1. Since supervised Learning will increase in sophistication, the algorithm takes a long time to compute during preparation. Therefore, because much of the world’s knowledge is unlabelled, the output is minimal. so, It does not occur in real-time.
2. It performs poorly where non-linear interactions occur. Most of the supervised models of Learning, such as linear Regression, is not scalable to consider a more dynamic system. therefore, It takes a lot of computational time, and it is often difficult to add proper polynomials or terms of interaction.
3. If the knowledge keeps increasing, it is not cost-effective, contributing to data labelling’s complexity to predefine outputs. hence, It is costly to label an image dataset manually, and the most high-quality image dataset has only one thousand labels. (Zhu et al., 2009.)
Advantages and disadvantages of unsupervised Learning
Advantages:
1. Let’s refer back to the algorithm for patterns that have not been previously accounted for, thereby resulting in the independence of the learning path for the algorithm in unsupervised Learning.
2. Excels where there is an inadequately labeled dataset or unclear or continuously changing pattern recognition—learning the secret pattern of the data on which it was educated. therefore, In discovering latent structures in past data and future projections, it makes previously unmanageable challenges more solvable and more agile.
3. Simplified the human task of labeling by grouping similar objects and differentiating the rest. The group of a dataset is then labeled instead of labeling it one by one.
Disadvantages:
In some instances, an unsupervised learner cannot learn what action to take because the data has not been presented. The goal of unsupervised Learning is only to identify patterns in the available data feed, inaccessible to any output.
1. It is very sluggish and consumes enormous memory resources. so it is more challenging to scale to more massive datasets. Besides, it just implies that the dataset’s underlying clusters are glob-shaped.
2. Since it is mostly about estimation, the findings are not that reliable. Furthermore, we do not know the number of classes, so the conclusions are not absolute. Unsupervised Learning is less adept at addressing problems that are narrowly defined. (Figueiredo et al., 2002.)
Advantages and disadvantages of reinforcement learning
Advantages:
1. It is one of the closest kinds of Machine Learning that humans and mammals do. In truth, much of the simple RL algorithm comes from the human brain and the neurological system.
2. RL is one of the most involved AI, ML, and neural network science fields. Reinforcement teaching on strong mathematical pillars has seen remarkable implementation and growth. however, RL’s computational work is now a significant study field with successful researchers’ cooperation in multiple disciplines.
3. The learning agent or device itself, by communicating with the environment, crafts the data independently. Building a generalized formula such as supervised Learning does not need a large volume of data to train itself.
Disadvantages:
1. To be more precise and robust relative to other learning algorithms. so, you need a lot of training data and need some time to prepare. For example, to reach 100 percent median distributional DQN output, 70 million frames are required.
2. Domain-specific reinforcement learning implementation is not recommended. however, RL is good at overall problem-solving. Still, much of the time, those problems will work easier on domain-specific strategies than Reinforcement Learning.
3. It is hard to describe the incentive. The algorithm builder typically offers it or hand-tunes it. The incentive feature must conform with the particular target or face overfitting and be stranded at local optima. (Silver. D et al., 2018)
References:
Caruana, R. and Niculescu-Mizil, A., 2006, June. An empirical comparison of supervised learning algorithms. In Proceedings of the 23rd international conference on Machine learning (pp. 161-168).
Zhu, X.J., 2005. Semi-supervised learning literature survey. University of Wisconsin-Madison Department of Computer Sciences.
Zhu, X. and Goldberg, A.B., 2009. Introduction to semi-supervised learning. Synthesis lectures on artificial intelligence and machine learning, 3(1), pp.1-130.
Figueiredo, M.A.T. and Jain, A.K., 2002. Unsupervised learning of finite mixture models. IEEE Transactions on pattern analysis and machine intelligence, 24(3), pp.381-396.
Barlow, H.B., 1989. Unsupervised learning. Neural computation, 1(3), pp.295-311.
Ghahramani, Z., 2003, February. Unsupervised learning. In Summer School on Machine Learning (pp. 72-112). Springer, Berlin, Heidelberg.
Gullapalli, V., 1990. A stochastic reinforcement learning algorithm for learning real-valued functions. Neural networks, 3(6), pp.671-692.
Jaakkola, T., Singh, S.P. and Jordan, M.I., 1995. Reinforcement learning algorithm for partially observable Markov decision problems. In Advances in neural information processing systems (pp. 345-352).Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T. and Lillicrap, T., 2018. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419), pp.1140-1144.
Written By: Saarika R Nair
Reviewed By: Rushikesh Lavate
If you are Interested In Machine Learning You Can Check Machine Learning Internship Program
Also Check Other Technical And Non Technical Internship Programs
