IA 2021-12-15

A submissão "Processamento de linguagem natural aplicado à inteligência artificial" é um escrito interessante e extremamemnte oportuno, sobre técnicas estatísticas aplicadas à análise de textos, para problemas como a atribuição de autor.

O documento é o resultado do esforço de alunos de Inteligência Artificial, que merecem ser elogiados por isso e pela qualidade relativa de um primeiro escrito científico.

A submissão pode ser aceite para publicação, desde que os autores aceitem submeter uma nova versão com as seguintes modificações:
- todos os lapsos de escrita em Português e Inglês, corrigidos;
- uma atenção redobrada quanto às descrições dos conceitos matemáticos, optando por simplificá-las - cautela especial com as generalizações, considerando mesmo a eliminição do que não estiver aplicado nos algoritmos concretos descritos;
- são aceitáveis excertos de código, desde que sejam images de fundo branco, mas pede-se moderação quanto à extensão.

A submissão "Classification of Handwriting Digits" é um artigo muito interessante, oportuno, e criador de oportunidades diversas de primeiros contactos com conceitos de Machine Learning.

O documento é o resultado do esforço de alunos de Inteligência Artificial, que merecem ser elogiados por isso e pela qualidade relativa de um primeiro escrito científico.

A submissão pode ser aceite para publicação, desde que os autores aceitem submeter uma nova versão com as seguintes modificações:
- todos os lapsos de escrita, em Inglês, corrigidos;
- total consistência nas expressões - por exemplo Machine Learning sempre, ou sempre "machine learning";
- uma atenção redobrada quanto às descrições dos conceitos, como overfitting e underfitting. Se não houve certezas absolutas quanto às descrições, re-estudá-las ou mesmo omiti-las, mas não deixar passar imprecisões;
- são aceitáveis excertos de código, desde que sejam images de fundo branco, mas pede-se moderação quanto à extensão.

*****

Getting started with Artificial Intelligence
(to be published)
Artur Marques

IA 2021-12-15
arturmarques.com/edu/ia
AmTriangles + PDF

10:20
ler a1 => "reescrever" => antecipação dos comentários dos revisores
ler a2 => "reescrever" =>  antecipação dos comentários dos revisores

10:30-11:20 => apresentações


Sessões da última semana de aulas:
APRESENTAÇÕES
IA + 2022-01-12

TESTE:
IA = 2022-01-19

arturmarques.com/edu/ia/files/w11/nlp.pdf

arturmarques.com/edu/ia/files/w11/digits.pdf


******


ABSTRACT
In this article, we will present our first contact with Artificial Intelligence, more specifically, Machine Learning. We aim to introduce some concepts and to present an example of a project related to the classification of handwritten digits in Python. To do so, this article follows a project studied in one of our classes of «AI», complemented by literature on the subject. Finally, the article presents results obtained by testing the algorithms in question, a corresponding interpretation, along with a discussion on the applicability and viability of this approach in real-life conditions.
Keywords: Machine Learning, Neural Network, Supervised Learning, Training set, Test set.
1	INTRODUCTION
In this era of digitalization, Artificial Intelligence (AI) is progressing rapidly into diverse areas. Society gradually sees the need for machines able to learn and mimic our actions, as humans. The use of machine learning has grown exponentially as it is used in our everyday life. We see it, for example, in image and speech recognition, medical diagnosis, advertising and many other fields of the industry.
This article intends to present one example of Machine Learning (ML). It introduces some related concepts. Our investigation was mainly guided by classes of Artificial Intelligence, with our Professor Artur Marques, from the Polytechnic Institute of Santarém, and some online research of our own. Accordingly, throughout this article, to help understand the theories, we quote segments from relevant books and papers. We illustrate some of the results and include images of the project's source code.
The project in this article - “Classification of Handwriting Digits" - was created and discussed during one of our classes. It is a generic example of how machine learning can happen. It is essentially about the machine learning to correctly classify handwritten digits. In other words, the machine receives a series of pictures of handwritten digits and is expected to return their correct corresponding value or classification. Before mentioning, in detail, how the project works, we need to establish some concepts, essential for the understanding of the overall problem.
This is a paper on supervised ML that bridges to related concepts, also mentioned in the source code itself, for an overall understanding of the whole project.
2	METHODOLOGY
2.1	Machine Learning
Machine Learning is precisely what you may think it is - it is about the machine learning something. How does the machine learn? It is a complex process but, briefly, in supervised learning, it learns by analysing training data on a specific subject and having access to the correct classification for each sample in the data. Thus, the dataset is a set of examples for the machine to “learn” from. In the end, it is expected that the machine will be able to make correct classifications upon new data - data that was not included in these examples. Usually, the more data we give the machine, the more chances we create for the machine to learn, making it abler to correctly classify samples, predicting their corresponding classes.
2.2	Supervised Learning VS Unsupervised Learning
For many projects, including the one we are discussing, Supervised Learning is used. This term is a subset of machine learning and is based on training. This concept revolves around using algorithms that learn from supervised pairs (input and desired output). It uses labelled datasets for training that tell the system what output should correspond to each input value. After giving the machine such labelled datasets, another dataset, different from training one, is used for testing, to measure the accuracy of the model the machine built, while learning.

Data for testing is also labelled, but we only feed the algorithm the inputs, not the classifications. Later, after running the algorithm, we can check if the computed outputs match the correct labels, or not. If the accuracy is too low, one solution might be to increase the amount of data available for training, or change its quality, or change the entire model complexity.

Data for testing is also labelled, but we only feed the algorithm the inputs, not the classifications. Later, after running the algorithm, we can check if the computed outputs match the correct labels, or not. If the accuracy is too low, one solution might be to increase the amount of data available for training, or change its quality, or change the entire model complexity.
What if the entire data in a dataset is used only for training, and nothing is reserved for testing? This is where the terms overfitting and underfitting enter. If we give all the data to training, then we risk "overfitting", where the model learns the detail and noise of the training data to the extent that it might negatively affect its performance on new data. If most of the data was used for testing, at the cost of not enough examples being used, that might cause "underfitting", where the model is lacking in capturing the relationship between the input and output variables accurately, generating a low accuracy software.