From machine learning to digitalization

From GEST-S482 Digital Business
Jump to navigation Jump to search

Machine learning (ML) is a subset of artificial intelligence which has undergone significant development in recent years, particularly with the appearance and evolution of Big Data, automation, and new techniques such as deep learning and neural networks. It provides systems the ability to automatically learn and improve from previous experience, and is for example used in problems for which it is difficult or unfeasible to develop conventional algorithms to perform the needed tasks. And it is the case in many problems of the contemporary societies where the learning machine has enabled great advances to be made. It is for example the case in image recognition, product recommandations, self-driving cars and many others. [1] [2]

As one will see, the learning machine is intimately linked to data, hence its evolution with the arrival of Big Data. Indeed, this exploration of data to be able to learn, predict, anticipate and even replace humans in certain tasks is one of the essential components of machine learning. To do so, the learning machine is based on mathematics, and more precisely, the (computational) statistics.


Machine learning as subfield of AI ([3])

Overview

Purpose and framework

The essence of the learning machine is based on a very simple principle made up of 3 fundamental points: a pattern exists, one cannot pin it down mathematically, and one have data on it. Targeting these 3 points enables us to understand for which problem the machine learning algorithms can be applied. The purpose of machine learning is therefore to use the previous available data - the set of observations - to uncover the underlying pattern, understand the structure of data, or be able to predict the result for new data point that are not being seen so far. This is precisly why machine learning is so used today. Indeed, a lot of tasks with a set of rules or a with a data-defined pattern can nowadays be completed by machine learning. It is a very useful way to transform processes that were previously only possible for humans to perform faster. It is mainly due to the computational speed and the actual computational power. [4]

Learning process

As previously explained, the goal of machine learning algorithms is to learn from and make prediction on data. To do so, a large amount of data is usually required. For information, it is moreover due to this inability to manage a large amount of data that the learning machine took a long time to develop when it was first created, experiencing a long period of stand-by, before knowing the evolution we can observe today.

This large amount of data forms a dataset, which can be defined as a collection of related sets of information that is composed of separate elements but can be manipulated as a unit by a computer. And this dataset, which contains data, is typically split into 3 different parts : training data, validation data and test data sets. Each of this set has a goal that is different from the others, and that is important for the algorithm to learn and validate a model.

Different datasets ([5])

Training Dataset

The training dataset is, as its name indicates, the set of data on which the machine learning algorithm will be trained. To be more precise, it is the sample of data used to fit the model and the parameters. The goal is to produce a trained model that will generalize well the unkwon data.

However, it is important to be careful about the use of the training set. In fact, by using too large a part of the training set, we can overfit the data. It means that the algorithm identify and exploit some patterns of the training data, which cannot be hold in general. It is due to the fact that it is too much data from the dataset that are used in the training set - and so, not enough in the other two. [6]

Validation Dataset

The second stage of the learning process is to evaluate the predictions of the model and to learn from mistakes before validating the model and the dataset. To do so, an estimation of the mistakes or the lossed between the model and the real data is done through the evolution process. yields on the validation set at any given point of time. This part of the learning process is important because it allows to check how accurate is the model output, but it is also helping to turn the the parameters of the model.

To resume, the validation is mainly used in order to evaluate a given model for a frequent evaluation, but also to fine-tune and update the (hyper)parameters of the model. [7]

Test Dataset

The test dataset is the final evaluation that a model need to be validated. This step is important to test the final testing of model that helps to generalizability and finds out the working accuracy of the model. In addition, it is used to evaluated competing models for example.

Generally, if the model is good, the training dataset also fits the test dataset well. If it is not the case, it can often be due to the phenomena of overfitting previously explained. [8]

Machine learning approaches

As it exists several approaches in many scientific fields, it is also the case in machine learning. Indeed, it exists different methods in machine learning that can be applied. This last differ by the type of problem that they are intended to solve but also by the type of data that they are using for input/output. Depending on these, machine learning can be divided in three categories depending on which approach is used, and on the nature of the feedback and data as previously explained: supervised learning, unsupervised learning and reinforcement learning.

Machine learning approaches ([9])


Supervised learning

Supervised learning is an interesting method because it works in a similar way to how humans learn. Indeed, supervised machine learning algorithms apply what has been learned by using previous data with the aim of using it to predict future events for new data.

It is done by example starting from the analysis of a known traning set, from which the learning algorithm will produce an inferred function to make predictions about the output values. With this last, the system will be able after a sufficient training to provide targets for any new input. [10]

Unsupervised learning

The main difference between supervised and unsupervised machine learning algorithms is that while the first learns from data that are labeled, the second one - unsupervised learning - learns from data that are neither classified or labeled. So the unsupervised algorithm tries to learn some inherent structure to the data with only unlabeled examples, and tries to find a hidden structure from these data. So it is more difficult for the system to figure out the right output, that is why it is not its goal, which one is just to explore the data and draw inferences from the dataset.

To give some examples of application, this type of machine learning algorithms are mainly used for clustering or dimensionality reduction.

Reinforcement learning

The reinforcement learning is a machine learning method a little bit different from the other two. Indeed, it is a learning method that is mainly characterized by two points : its interaction with its environment by producing actions and and the fact that a reward (or "errors") is resulting from this action.

To explain the methodology used by reinforcement learning, it is in a first time important to understand that its purpose is to allow an agent to automatically determine the ideal behavior in a specific context. To do so, the agent is using a trial and error search by testing several possibilities by means of actions. Each of these actions have an impact on the environment and these impacts can be characterized by a reward. By this reward feedback, it is then possible to train the agent, and to determine the ideal behavior which is the one that maximizes the performance, and so, for which the actions are the best - meaning have the best rewards. [11]

Applications in digitalization

Digital transformation has been one ofthe hottest technology trend in these last decade. But it is also the case of some artificial intelligence technique, as for exemple machine learning. The idea of making them work together to obtain even better results thus became obvious, and the machine learning became one of the components of many digitization processes.

This power of AI and machine learning can be seen with concret examples and fact. For example, according to a recent survey, 98% of respondents of this last said that the use of AI to support their activities to power digital transformation was a very good choice, and had a improving impact on the revenue of their organization. And in the different AI techniques, 75% of respondents said that it was machine learning techniques that has had the largets impact and has player a important role in the digital transformation of their companies. It is for example confirmed by Andrew Hallman, head of the CIA Digital Innovation Directorate, who said that AI and machine learning "can fundamentally change the way we do business". [12]

To do so, the question is how it is possible to exploit machine learning in dizitalization. A.Hallman explains it with good words by saying that machine learning allow "to make sense of our sensory environment and the signals they’re getting from that threat environment" and "to develop models for how the world behaves, to be able to provide more anticipation for unfolding events and changing conditions". [13]

These last are the main points for which machine learning is a very useful tool in digitalization of many companies. Indeed, it can be use to detect some changes in the different regions of the world, and discover some clusters in these regions, what can be used for an adaptation and customization for each regions, but it can be alos be used to predict some patterns of customers behaviours - what can be use to adapt the product deliverated by a company for example -, and many others.

To be more concrete, the machine learning is an indispensable tool for companies and their digitizations because it has allowed them to use the mass of data they had, to benefit from it, and to apply it in a concrete way. All this information has therefore been able to make sense, and entering into an automated system, saving time and efficiency, but also opening up new possibilities. This has brought the company closer to the customers, using their data to establish suggestions and individualized profiles for each of them according to their preferences. It has also facilitated resource management, in particular by anticipating and being able to predict the quantity of products needed each day based on several parameters. Many other applications of the learning machine exist, but in short, it is this ability to learn, predict and establish profiles that has made the learning machine an exceptional tool and a data catalyst that for the digitalization of each company.

Reference