Welcome to the third chapter of our five-part professional article series! In the past, we have looked at the historical arc and basic concepts of AI, but now we have reached the true core of technology. It is important to clarify: a\u00a0artificial intelligence<\/em>\u00a0The term is often misleading from a scientific point of view. These systems do not \"think\" in the sense of human cognition, but are complex.\u00a0pattern recognition<\/em>\u00a0and\u00a0Probability Based Supplement<\/em>\u00a0finish.<\/p>\n\n\n\n Let us take the example of \u2018apple\u2019: if an algorithm has to finish the sentence, it relies on the frequency distribution of the trained dataset. If \u2018apple red\u2019 is reported 7 times, \u2018apple green\u2019 is reported 2 times and \u2018apple yellow\u2019 is reported 1 time, the system is 70%with a probability of red<\/em> You will choose the word. This \"statistical prediction\" is the basis for the functioning of today's intelligent systems.<\/p>\n\n\n\n Main points of the entry:<\/strong><\/p>\n\n\n\n According to the literary definition, the learning<\/strong> a process of improving the effectiveness of our future decisions and actions based on observations (data) of the world. In the case of machine learning, this means that the algorithm does not implement pre-recorded rules, but draws conclusions from the data.<\/p>\n\n\n\n Already in 1950, Alan Turing outlined the \"child mind\" metaphor<\/strong>: Instead of trying to program the entire complexity of an adult human brain, create a simpler, learning-capable system that is raised to a higher level of intelligence by teaching (data). Modern ML systems do just that: They learn from the examples.<\/p>\n\n\n\n Based on the methodology of the learning process, three main types are distinguished. It is worth noting that the literature often defines unsupervised learning as practicing along a set of rules.<\/p>\n\n\n\n Managed learning tasks are divided into two large groups according to the type of output value:<\/p>\n\n\n\n The decision tree is one of the most transparent tools of machine learning, which breaks down complex relationships into a series of elementary tests.<\/p>\n\n\n\n Decision trees can handle Noisy data<\/strong>, because they automatically filter out variables that are irrelevant to the decision. Their special form is anytime decision tree<\/strong>: In this model, we can also assign a decision to internal nodes. This allows the system to provide a preliminary (not necessarily optimal) response immediately if there is little time available, and to refine it by traversing the deeper levels of the tree if there is time.<\/p>\n\n\n\n In machine learning, a hypothesis ( Expert comment:<\/strong> Principle of model selection Ockham's Razor<\/strong>: If two hypotheses match the data in the same way, always choose the simplest hypothesis. A straight line (linear model) tends to be much more generalized than an all-point curve (consistent hypothesis) that waves pointlessly. The aim is to compromise between the accuracy shown on the training data and the simplicity of the model.<\/p>\n\n\n\n In 1943, McCullogh and Pitts created the Perceptron<\/strong>, The first scientific model of AI. It is a mathematical unit that uses input values ( Perceptron calculates the following function: If the result Neural networks model the network of brain neurons and break through the limitations of Perceptron. They consist of several layers:<\/p>\n\n\n\n Learning here is the continuous modification of all connections (weights) in the network until the output error is minimized.<\/p>\n\n\n\n A Deep Learning<\/strong> (deep learning) the evolution of multilayer neural networks, which became dominant in the mid-2000s with the explosion of computational capacity. Milestone of 2016 AlphaGo<\/strong> It was a victory that demonstrated the extraordinary power of technology.<\/p>\n\n\n\n In the field of visual vision Convolutional Neural Networks (CNN)<\/strong> They brought a breakthrough. In these special convolution filters<\/strong> (so-called filters) that scan pixels like digital magnifiers. Filters look for specific patterns \u2013 edges, curvatures, shapes \u2013 so the machine can not only see a set of dots, but also recognise objects. This technology extends the machine vision capabilities expected in the Turing test from modern diagnostics to self-driving cars.<\/p>\n\n\n\n The standard environment for modern AI development is the Python language and its ecosystem (based on BME thematic sources):<\/p>\n\n\n\n 11. Summary and outlook<\/p>\n\n\n\n We have seen that machine learning is not black magic, but a series of precise mathematical and logical models that derive their power from data. Whether it's a credit assessment based on income or number of children, or convolution filters analyzing pixels, the goal is common: maximising the effectiveness of future decisions.<\/p>\n\n\n\n In Part 4 of the series, we move from the technical foundations to the social impact of AI: We will examine the ethical issues<\/strong>, Algorithm bias, and Strong AI<\/strong> and a Superintelligence<\/strong> prospects for the future.<\/p>","protected":false},"excerpt":{"rendered":" Imagine machine learning as a process where software continuously improves the effectiveness of your decisions through experience (data).<\/p>","protected":false},"author":2,"featured_media":1557,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","format":"standard","meta":{"slim_seo":{"title":"3. R\u00e9sz: A motorh\u00e1ztet\u0151 alatt \u2013 Hogyan tanul a g\u00e9p? - Tesz\u00e1ry P\u00e9ter","description":"K\u00e9pzelj\u00fck el a g\u00e9pi tanul\u00e1st \u00fagy, mint egy folyamatot, ahol a szoftver a tapasztalatok (adatok) r\u00e9v\u00e9n folyamatosan jav\u00edtja d\u00f6nt\u00e9sei hat\u00e9konys\u00e1g\u00e1t."},"footnotes":""},"categories":[93],"tags":[96],"class_list":["post-2212","publikacio","type-publikacio","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","tag-mesterseges-intelligencia"],"desktop_mode_lock":null,"desktop_mode_contributors":[],"desktop_mode_attached_media":[1557],"meta_box":[],"_links":{"self":[{"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/publikacio\/2212","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/publikacio"}],"about":[{"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/types\/publikacio"}],"author":[{"embeddable":true,"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/comments?post=2212"}],"version-history":[{"count":3,"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/publikacio\/2212\/revisions"}],"predecessor-version":[{"id":2255,"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/publikacio\/2212\/revisions\/2255"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/media\/1557"}],"wp:attachment":[{"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/media?parent=2212"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/categories?post=2212"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/teszarypeter.hu\/en\/wp-json\/wp\/v2\/tags?post=2212"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}\n
2. What is Machine Learning?<\/h3>\n\n\n\n
3. The Three Major Paradigms of Machine Learning<\/h3>\n\n\n\n
\n
Point of view<\/th> Supervised learning<\/th> Unsupervised learning<\/th><\/tr> Purpose<\/strong><\/td> Predicting the output of new data<\/td> Grouping of data, searching for samples<\/td><\/tr> Data requirements<\/strong><\/td> Tagged learning set<\/td> Unlabelled data<\/td><\/tr> Example<\/strong><\/td> Letter recognition, credit assessment<\/td> Customer segmentation, clustering<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n 4. Classification and regression: The Two Faces of Supervised Learning<\/h3>\n\n\n\n
Type<\/th> Description<\/th> Example<\/th><\/tr> Classification<\/strong><\/td> The output is an element of a finite set (discrete value).<\/td> Handwriting recognition (what letter?), credit assessment (yes\/no).<\/td><\/tr> Regression (Regression)<\/strong><\/td> The output is a real number moving on a continuous scale.<\/td> Estimation of the distance of the object, the expected price of the property.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n 5. Decision Trees - Logic of Algorithms<\/h3>\n\n\n\n
\n
6. The Modeling Trap: Overstudy (Overfitting) and Ockham's Razor<\/h3>\n\n\n\n
h<\/code>) search, which generalizes well on new data. If your model is too complex, e.g. a higher polynomial of type 1 (b) is inserted at some points, the Overstudy (overfitting)<\/strong>. In this case, the model also \u2018magnetises\u2019 the noise and individual errors of the training data, but fails on the unknown test data.<\/p>\n\n\n\n7. Perceptron and linear separability<\/h3>\n\n\n\n
x1<\/code>, x2<\/code>\u2026) with weights (w1<\/code>, w2<\/code>\u2026) and adds a constant offset (bias<\/strong>, b<\/code>).<\/p>\n\n\n\nf(x) = w1*x1 + w2*x2 + ... + wn*xn + b<\/code><\/p>\n\n\n\n> 0<\/code>, classifies the data into one class, otherwise into the other. The learning phase is a w<\/code> Weights and a b<\/code> It is a fine-tuning of value. However, the limit of Perceptron is linear separability<\/strong>: It can only isolate groups that can be separated by a straight line or plane. Because of this, a single Perceptron is unable to learn XOR (exclude or) function<\/strong>, since its points cannot be separated by a single line.<\/p>\n\n\n\n8. Artificial Neural Networks (ANN)<\/h3>\n\n\n\n
\n
9. Deep Learning and Convolutional Networks (CNN)<\/h3>\n\n\n\n
10. Practical implementation: Tools and Libraries<\/h3>\n\n\n\n
\n