{"id":1451,"date":"2022-02-23T20:34:21","date_gmt":"2022-02-23T20:34:21","guid":{"rendered":"https:\/\/nilg.ai\/?p=1451"},"modified":"2025-03-17T12:39:39","modified_gmt":"2025-03-17T12:39:39","slug":"modelos-de-ensino-com-dados-gratuitos","status":"publish","type":"post","link":"https:\/\/nilg.ai\/pt\/202202\/teaching-models-with-free-data\/","title":{"rendered":"Modelos de Ensino com Dados Gratuitos"},"content":{"rendered":"<p><img decoding=\"async\" class=\"aligncenter size-large wp-image-1468\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/02\/pexels-erik-mclean-4582541-1024x683.jpg\" alt=\"\" width=\"1024\" height=\"683\" \/><\/p>\n<p>\u201cQuanto mais vejo, menos sei\u201d pode ser um ditado, mas n\u00e3o se aplica a modelos de IA. \u00c9 bem sabido que o desempenho de uma rede neuronal artificial depende muito do volume e da diversidade dos dados que foram apresentados ao modelo. Isto acontece porque expor os modelos \u00e0 diversidade ajuda-os a selecionar caracter\u00edsticas relevantes e a mitigar potenciais preconceitos, ou seja, a compreender os objetos de estudo e a executar melhor as suas tarefas.<\/p>\n<p>A obten\u00e7\u00e3o de dados para alimentar tais modelos pode parecer trivial, pois os dados est\u00e3o por todo o lado. No entanto, o acesso a dados bem estruturados, rotulados, livres de direitos de autor e n\u00e3o privados ainda \u00e9 um problema para a maioria dos cientistas de dados.<\/p>\n<p>Para superar isto, peritos em IA desenvolveram v\u00e1rias abordagens para otimizar o processo de aprendizagem, em particular, arquiteturas inteligentes para modelos de Aprendizagem Auto-Supervisionada.<\/p>\n<p>Neste artigo do blogue, (re)apresentamos-lhe o Self-Supervised Learning, juntamente com uma das nossas estrat\u00e9gias favoritas \u2013\u00a0 <b>Modelos siameses <\/b>\u2013 e poss\u00edveis aplica\u00e7\u00f5es. Se quiser aprender ainda mais sobre Self Supervised Learning e outras t\u00e9cnicas de Machine Learning, inscreva-se no nosso curso online:<\/p>\n<div class=\"thinkific-product-card\" data-btn-txt=\"Learn more!\" data-btn-txt-color=\"#ffffff\" data-btn-bg-color=\"#1b9eea\" data-card-type=\"card\" data-link-type=\"landing_page\" data-product=\"1787642\" data-embed-version=\"0.0.2\" data-card-txt-color=\"#7d7d7d\" data-card-bg-color=\"#ffffff\" data-store-url=\"https:\/\/learn.nilg.ai\/embeds\/products\/show\">\n<p><noscript><a href=\"https:\/\/learn.nilg.ai\/courses\/the-machine-learning-spectrum\" target=\"_blank\" rel=\"noopener\">Inscreva-se j\u00e1!<\/a><\/noscript><\/p>\n<\/div>\n<p>&nbsp;<\/p>\n<h3><\/h3>\n<h3>O que \u00e9 Aprendizagem Auto-supervisionada?<\/h3>\n<p>A Aprendizagem Supervisionada \u00e9 uma abordagem de aprendizagem autom\u00e1tica que recebe dados de entrada juntamente com um objetivo espec\u00edfico e visa aprender os padr\u00f5es dos dados e a fun\u00e7\u00e3o de transforma\u00e7\u00e3o que converte a entrada na sa\u00edda. Do outro lado da moeda, temos a Aprendizagem N\u00e3o Supervisionada, que \u00e9 um m\u00e9todo que n\u00e3o necessita de um objetivo para cumprir a sua miss\u00e3o, uma vez que visa principalmente encontrar padr\u00f5es nas distribui\u00e7\u00f5es dos dados.<\/p>\n<p>De seguida, temos o Self Supervised Learning, que \u00e9 um m\u00e9todo de Unsupervised Learning visto que utiliza dados n\u00e3o rotulados e tem a particularidade de criar r\u00f3tulos sint\u00e9ticos para se comportar como um modelo de Supervised Learning.<\/p>\n<p>Estes r\u00f3tulos podem ser criados aplicando transforma\u00e7\u00f5es triviais aos dados. Aqui ficam alguns exemplos:<\/p>\n<table>\n<tbody>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Objetivo<\/b><\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\"><b>Transforma\u00e7\u00e3o<\/b><\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\"><b>Etiqueta Sint\u00e9tica<\/b><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\">Quantificar a rota\u00e7\u00e3o de imagem<\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\">Rota\u00e7\u00e3o aleat\u00f3ria<\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\">\u00c2ngulo de rota\u00e7\u00e3o<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\">Avaliar a Qualidade dos Dados<\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\">Inser\u00e7\u00e3o de valores aleat\u00f3rios num dataframe<\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\">Bandeira bin\u00e1ria para dados corrompidos<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\">Coloriza\u00e7\u00e3o de Imagem<\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\">RGB para preto e branco<\/p>\n<\/td>\n<td>\n<p style=\"text-align: center;\">Imagem RGB<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>&nbsp;<\/p>\n<p>Os modelos de autoaprendizagem podem ser utilizados de duas maneiras diferentes:<\/p>\n<ul>\n<li aria-level=\"1\">Para <b>gerar previs\u00f5es<\/b>, caso a tarefa sint\u00e9tica corresponda ao prop\u00f3sito principal do modelo. Por exemplo, treinar um modelo de coloriza\u00e7\u00e3o de imagens e utiliz\u00e1-lo para colorir imagens a preto e branco.<\/li>\n<li aria-level=\"1\">Para <b>partilhar o seu conhecimento <\/b>(ao partilhar os seus pesos) com um modelo de Aprendizagem Supervisionada com uma tarefa diferente. Por exemplo, treinar um modelo de Autoaprendizagem para detetar se uma imagem est\u00e1 invertida e usar os pesos para inicializar um modelo de dete\u00e7\u00e3o de objetos.<\/li>\n<\/ul>\n<div class=\"course-cta\">\n\t\t<div class=\"course-cta-img\"><img decoding=\"async\" width=\"582\" height=\"903\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/08\/1-194x301@3x.png\" class=\"attachment-full size-full\" alt=\"\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/08\/1-194x301@3x.png 582w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/08\/1-194x301@3x-193x300.png 193w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/08\/1-194x301@3x-300x465.png 300w\" sizes=\"(max-width: 582px) 100vw, 582px\" \/><\/div>\n\t\t<div class=\"course-cta-content\"><h6>Curso<\/h6><h3>O Espectro do Machine Learning<\/h3>\n\t\t\t<p>Se quiser saber mais sobre aprendizagem auto-supervisionada e outros m\u00e9todos de aprendizagem, consulte este curso.<\/p>\n\t\t\t<a href=\"https:\/\/nilg.ai\/pt\/product\/the-machine-learning-spectrum\/\" class=\"cta_btn\">Saber mais<\/a>\n\t\t<\/div>\n\t<\/div>\n<h3><\/h3>\n<h3>Como \u00e9 que aplicamos isto?<\/h3>\n<p>Na NILG.AI, uma das nossas arquiteturas de modelo favoritas \u00e9 a <b>Redes Siamese<\/b>. Os modelos siameses s\u00e3o compostos por uma arquitetura de modelo com dois (ou mais) ramos, onde cada um deles recebe uma entrada diferente. Os pesos dos ramos podem ser partilhados ou n\u00e3o e, na camada final do modelo, as sa\u00eddas dos ramos s\u00e3o comparadas.<\/p>\n<p>Abaixo pode ver um exemplo de arquitetura para uma rede siamesa aplicada a imagens, que implement\u00e1mos para lhe demonstrar a utilidade destes modelos:<\/p>\n<p><a href=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/02\/Free-Data-Siamese.svg\"><img decoding=\"async\" class=\"aligncenter size-full wp-image-1462 attachment-svg\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/02\/Free-Data-Siamese.svg\" alt=\"\" \/><\/a><\/p>\n<p>Este modelo tem dois ramos que partilham pesos. Cada ramo \u00e9 composto por um codificador (modelo CNN para extra\u00e7\u00e3o de caracter\u00edsticas) e um bloco de camadas totalmente ligadas. A sa\u00edda dos dois ramos \u00e9 ent\u00e3o combinada para calcular a <b>fun\u00e7\u00e3o de perda.\u00a0<\/b><\/p>\n<p>Para alimentar o modelo, implement\u00e1mos um gerador de dados que recebe uma imagem como entrada, aplica uma transforma\u00e7\u00e3o definida pelo utilizador e retorna as duas imagens transformadas juntamente com um alvo sint\u00e9tico. A fun\u00e7\u00e3o de transforma\u00e7\u00e3o \u00e9 definida na inicializa\u00e7\u00e3o do modelo, mas a magnitude da transforma\u00e7\u00e3o deve ser um valor aleat\u00f3rio que se encaixe no intervalo fornecido. Por exemplo, para ensinar um modelo a aprender a orienta\u00e7\u00e3o dos objetos, pode ser passada uma fun\u00e7\u00e3o de rota\u00e7\u00e3o juntamente com o intervalo de \u00e2ngulos poss\u00edveis. O gerador de dados selecionar\u00e1 dois \u00e2ngulos aleat\u00f3rios do intervalo fornecido e rodar\u00e1 a imagem de entrada considerando esses \u00e2ngulos, criando duas imagens transformadas diferentes. O gerador compara os dois \u00e2ngulos: se o \u00e2ngulo da primeira transforma\u00e7\u00e3o for superior ao \u00e2ngulo da segunda transforma\u00e7\u00e3o, o r\u00f3tulo sint\u00e9tico \u00e9 definido como 1, caso contr\u00e1rio, o r\u00f3tulo \u00e9 definido como 0. No final, o gerador produz as duas imagens transformadas e o r\u00f3tulo sint\u00e9tico correspondente.<\/p>\n<p>&nbsp;<\/p>\n<h3>Resultados de Base<\/h3>\n<p>Como um exerc\u00edcio explorat\u00f3rio, <a href=\"http:\/\/nilg.ai\/pt\/\" target=\"_blank\" rel=\"noopener\">NILG.AI<\/a> desenvolvida v\u00e1rias fun\u00e7\u00f5es de transforma\u00e7\u00e3o e treinado um modelo com um conjunto de dados muito pequeno (menos de 100 amostras) de imagens aleatoriamente recolhidas de <a href=\"https:\/\/unsplash.com\/\" target=\"_blank\" rel=\"noopener\">Unsplash<\/a>. As transforma\u00e7\u00f5es triviais inclu\u00edam:<\/p>\n<ul>\n<li aria-level=\"1\">Adi\u00e7\u00e3o de Desfoque<\/li>\n<li aria-level=\"1\">Rota\u00e7\u00e3o de imagem<\/li>\n<li aria-level=\"1\">Desvio de brilho<\/li>\n<\/ul>\n<p>A arquitetura do modelo siam\u00eas foi adaptada a cada fun\u00e7\u00e3o de transforma\u00e7\u00e3o, criando tr\u00eas modelos diferentes. Um \u00fanico ramo treinado do modelo siam\u00eas foi ent\u00e3o usado para calcular as previs\u00f5es em imagens \u00fanicas e estes s\u00e3o os resultados observados.<\/p>\n<p><b>Note<\/b>: The predictions correspond to the output of the last fully connected layer of the branch. These values should be interpreted as proxies of the magnitude of the transformations. To use these outputs as predictions of the real value of the transformation, e.g. rotation angle, the output should be calibrated.<\/p>\n<h5>Blur Addition<\/h5>\n<p>For the blur addition, we applied a gaussian blur filter with fixed window size and a variable sigma value. To test the model performance, we applied blur addition with different values of sigma to the same image and extracted the predictions from the model. In the image below, we can observe that the model output increases with the sigma, being able to distinguish the intensity of the transformation.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1452\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_sample.png\" alt=\"\" width=\"827\" height=\"519\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_sample.png 827w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_sample-300x188.png 300w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_sample-768x482.png 768w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_sample-600x377.png 600w\" sizes=\"(max-width: 827px) 100vw, 827px\" \/><\/p>\n<p>In the second test phase, we repeated the transformations in each image of the test set and extracted the correlation between the model output and the sigma. This model has a 95,5% correlation with the ground truth, having the potential to be used as a blur detector.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1453\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_correlation.png\" alt=\"\" width=\"440\" height=\"333\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_correlation.png 440w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_correlation-300x227.png 300w\" sizes=\"(max-width: 440px) 100vw, 440px\" \/><\/p>\n<p>Finally, we extracted the predictions for the original images of the test set, presenting some examples in the grid below. Since we were dealing with high-quality images, the model returned low values for blur detection (usually lower than -6.4), except for the image with the coffee filter which contains a significant background blur.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1454\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_raw_images.png\" alt=\"\" width=\"821\" height=\"519\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_raw_images.png 821w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_raw_images-300x190.png 300w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_raw_images-768x485.png 768w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/blur_raw_images-600x379.png 600w\" sizes=\"(max-width: 821px) 100vw, 821px\" \/><\/p>\n<h5>Image Rotation<\/h5>\n<p>We repeated the same test for the rotation model, this time changing the angle value, getting the results below.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1455\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_sample.png\" alt=\"\" width=\"827\" height=\"519\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_sample.png 827w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_sample-300x188.png 300w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_sample-768x482.png 768w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_sample-600x377.png 600w\" sizes=\"(max-width: 827px) 100vw, 827px\" \/><\/p>\n<p>The correlation value for this model was 0.37, which is understandable considering the difficulty of the task. Recognizing if an object is tilted and identifying the correspondent angle implies prior knowledge of the object itself, which is hard to teach to a model with a set of only 100 images representing different objects. Therefore, to use this model as a rotation corrector, we might need to use a lot more data or put more constraints on the image selection, e.g. select images of interior decoration, only.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1456\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_correlation.png\" alt=\"\" width=\"440\" height=\"333\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_correlation.png 440w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/rot_correlation-300x227.png 300w\" sizes=\"(max-width: 440px) 100vw, 440px\" \/><\/p>\n<p>Since this task requires a higher knowledge of the objects, it can also be used as a secondary task of a multitask learning model. This strategy can help the model to better learn the features of the objects of interest avoiding extra labeling costs.<\/p>\n<h5>Brightness Deviation<\/h5>\n<p>The same analysis was made for the brightness model. For this example, the transformation relies on manipulating the value (V) on the HSV color representation by adding a random number. When testing the model for the same transformed image, we can observe that the predicted target increases with the added value.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1457\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_sample.png\" alt=\"\" width=\"827\" height=\"519\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_sample.png 827w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_sample-300x188.png 300w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_sample-768x482.png 768w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_sample-600x377.png 600w\" sizes=\"(max-width: 827px) 100vw, 827px\" \/><\/p>\n<p>Analyzing the performance of the model for the overall test set, we can confirm the correlation between the model outputs and the ground truth since the correlation rate is around 77,5%. This model can be used for brightness correction or image quality assessment. In the next section, you will see some practical examples of where to use these models.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1458\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_correlation.png\" alt=\"\" width=\"440\" height=\"333\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_correlation.png 440w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_correlation-300x227.png 300w\" sizes=\"(max-width: 440px) 100vw, 440px\" \/><\/p>\n<p>Once again, the model predictions were extracted for the original images (shown below). These predictions are also eloquent since the model returned positive values for the brightest images and negative values for the darkest ones, in particular, the coffee filter and the bridge pictures.<\/p>\n<h3><img decoding=\"async\" class=\"aligncenter size-full wp-image-1459\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_raw_images.png\" alt=\"\" width=\"821\" height=\"519\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_raw_images.png 821w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_raw_images-300x190.png 300w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_raw_images-768x485.png 768w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/06\/bright_raw_images-600x379.png 600w\" sizes=\"(max-width: 821px) 100vw, 821px\" \/><\/h3>\n<h3>Use Cases<\/h3>\n<p>Self Supervised learning can be very useful to pre-train encoders to be used by other models or to be used as an extra task, making the final model more robust. However, there\u2019s also a lot of potential for these models to be used directly as predictors, and here are a few examples of where to use them.<\/p>\n<h5>Cuidados de sa\u00fade<\/h5>\n<p>Image quality assessment for medical imaging &#8211; trivial models like a blur, brightness, and crop detectors help can assess image quality in real-time, being useful to select the sample to be analyzed (by another model or by an expert).<\/p>\n<h5>Real State<\/h5>\n<p>Image standardization in the website &#8211; brighter and more colorful images are more attractive to prospects and, taking pictures in poor lighting conditions can influence the propensity of engagement with the image. With a brightness model as the one proposed above, it is possible to assess this feature and correct it automatically.<\/p>\n<h5>Car Dealership<\/h5>\n<p>Similar to the previous use case, the pictures that show the product (the car, in this case) may influence the propensity of a user to become a buyer. Trivial computer vision models like brightness quantifiers and blur detectors can be used to filter which images have enough quality to be published on the website.<\/p>\n<h3>Conclus\u00e3o<\/h3>\n<p>As we saw in this post, you don&#8217;t always need a large dataset to build a model that meets the needs of your business.<\/p>\n<p>If you\u2019re interested in making your company&#8217;s decisions data-driven but you\u2019re not sure you have a data structure prepared for that, contact us at info@nilg.ai, and let&#8217;s discuss some ideas!<\/p>\n<p>If you want to learn more about it, enroll to our course:<\/p>\n  \n\n <div class=\"author-cta\">\n\t\t<div class=\"author-cta-img\">\n\t\t    \n\t\t    <img decoding=\"async\" width=\"1024\" height=\"906\" src=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/07\/Web-Rafael.png\" class=\"attachment-full size-full\" alt=\"Rafael Cavalheiro NILG.AI\" srcset=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/07\/Web-Rafael.png 1024w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/07\/Web-Rafael-300x265.png 300w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/07\/Web-Rafael-768x680.png 768w, https:\/\/nilg.ai\/wp-content\/uploads\/2022\/07\/Web-Rafael-600x531.png 600w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/>\t\t    <\/div>\n\n<div class=\"author-cta-content\">\n\t<h3>Quer discutir esta ideia mais a fundo?<\/h3><p>Marcar uma reuni\u00e3o com <strong>Rafael Cavalheiro<\/strong><\/p>\t<a class=\"cta_btn\" onclick=\"Calendly.showPopupWidget('');return false;\"  \n\">Conhece o Rafael<\/a>\n\t\t\t\n\t<a href=\"https:\/\/nilg.ai\/pt\/?post_type=team&p=1650\" class=\"author-cta-link\">Saber mais<\/a>\n\t\t\t<\/div>\n\t<\/div>\n\n<div class=\"thinkific-product-card\" data-btn-txt=\"Learn more!\" data-btn-txt-color=\"#ffffff\" data-btn-bg-color=\"#1b9eea\" data-card-type=\"card\" data-link-type=\"landing_page\" data-product=\"1787642\" data-embed-version=\"0.0.2\" data-card-txt-color=\"#7d7d7d\" data-card-bg-color=\"#ffffff\" data-store-url=\"https:\/\/learn.nilg.ai\/embeds\/products\/show\">\n<p><noscript><a href=\"https:\/\/learn.nilg.ai\/courses\/the-machine-learning-spectrum\" target=\"_blank\" rel=\"noopener\">Inscreva-se j\u00e1!<\/a><span id=\"mce_marker\" data-mce-type=\"bookmark\" data-mce-fragment=\"1\">\u200b<\/span><\/noscript><\/p>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u201cThe more I see, the less I know\u201d might be a saying, but it does not apply to AI models. It\u2019s well known that the performance of an artificial neural network is highly dependent on the volume and on the diversity of the data that was shown to the model. This happens because exposing the [&hellip;]<\/p>\n","protected":false},"author":5,"featured_media":1468,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[53],"tags":[81,45,90],"class_list":["post-1451","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technical","tag-deep-learning","tag-machine-learning","tag-self-supervised"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.8 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Teaching Models With Free Data - NILG.AI<\/title>\n<meta name=\"description\" content=\"An overview of self-supervised learning methods using siamese neural networks. Training models without labels for real-life applications.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/nilg.ai\/pt\/202202\/modelos-de-ensino-com-dados-gratuitos\/\" \/>\n<meta property=\"og:locale\" content=\"pt_PT\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Teaching Models With Free Data - NILG.AI\" \/>\n<meta property=\"og:description\" content=\"An overview of self-supervised learning methods using siamese neural networks. Training models without labels for real-life applications.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/nilg.ai\/pt\/202202\/modelos-de-ensino-com-dados-gratuitos\/\" \/>\n<meta property=\"og:site_name\" content=\"NILG.AI\" \/>\n<meta property=\"article:published_time\" content=\"2022-02-23T20:34:21+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2025-03-17T12:39:39+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/02\/pexels-erik-mclean-4582541-e1656153098676.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"2048\" \/>\n\t<meta property=\"og:image:height\" content=\"1365\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"Paulo Maia\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@nilg_ai\" \/>\n<meta name=\"twitter:site\" content=\"@nilg_ai\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Paulo Maia\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"10 minutos\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/\"},\"author\":{\"name\":\"Paulo Maia\",\"@id\":\"https:\/\/nilg.ai\/#\/schema\/person\/a860b9395bd25c5090160a3ac30e980d\"},\"headline\":\"Teaching Models With Free Data\",\"datePublished\":\"2022-02-23T20:34:21+00:00\",\"dateModified\":\"2025-03-17T12:39:39+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/\"},\"wordCount\":1640,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/nilg.ai\/#organization\"},\"keywords\":[\"Deep Learning\",\"Machine Learning\",\"Self-Supervised\"],\"articleSection\":[\"Technical\"],\"inLanguage\":\"pt-PT\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/\",\"url\":\"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/\",\"name\":\"Teaching Models With Free Data - NILG.AI\",\"isPartOf\":{\"@id\":\"https:\/\/nilg.ai\/#website\"},\"datePublished\":\"2022-02-23T20:34:21+00:00\",\"dateModified\":\"2025-03-17T12:39:39+00:00\",\"description\":\"An overview of self-supervised learning methods using siamese neural networks. Training models without labels for real-life applications.\",\"inLanguage\":\"pt-PT\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/\"]}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/nilg.ai\/#website\",\"url\":\"https:\/\/nilg.ai\/\",\"name\":\"NILG.AI\",\"description\":\"Create ever-improving businesses with AI\",\"publisher\":{\"@id\":\"https:\/\/nilg.ai\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/nilg.ai\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"pt-PT\"},{\"@type\":\"Organization\",\"@id\":\"https:\/\/nilg.ai\/#organization\",\"name\":\"NILG.AI\",\"url\":\"https:\/\/nilg.ai\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"pt-PT\",\"@id\":\"https:\/\/nilg.ai\/#\/schema\/logo\/image\/\",\"url\":\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/03\/logo.svg\",\"contentUrl\":\"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/03\/logo.svg\",\"caption\":\"NILG.AI\"},\"image\":{\"@id\":\"https:\/\/nilg.ai\/#\/schema\/logo\/image\/\"},\"sameAs\":[\"https:\/\/twitter.com\/nilg_ai\",\"https:\/\/youtube.com\/@nilg_ai\",\"https:\/\/www.linkedin.com\/company\/nilg-ai\/\"]},{\"@type\":\"Person\",\"@id\":\"https:\/\/nilg.ai\/#\/schema\/person\/a860b9395bd25c5090160a3ac30e980d\",\"name\":\"Paulo Maia\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"pt-PT\",\"@id\":\"https:\/\/nilg.ai\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/9e3e7cb701fd6dcee37f7a4e9f73ef5a27f5741062c4cb179724a1052ea23086?s=96&d=mm&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/9e3e7cb701fd6dcee37f7a4e9f73ef5a27f5741062c4cb179724a1052ea23086?s=96&d=mm&r=g\",\"caption\":\"Paulo Maia\"},\"url\":\"https:\/\/nilg.ai\/pt\/author\/paulo\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Modelos de Ensino com Dados Gratuitos - NILG.AI","description":"Uma vis\u00e3o geral dos m\u00e9todos de aprendizagem auto-supervisionada utilizando redes neuronais siamesas. Treinar modelos sem r\u00f3tulos para aplica\u00e7\u00f5es da vida real.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/nilg.ai\/pt\/202202\/modelos-de-ensino-com-dados-gratuitos\/","og_locale":"pt_PT","og_type":"article","og_title":"Teaching Models With Free Data - NILG.AI","og_description":"An overview of self-supervised learning methods using siamese neural networks. Training models without labels for real-life applications.","og_url":"https:\/\/nilg.ai\/pt\/202202\/modelos-de-ensino-com-dados-gratuitos\/","og_site_name":"NILG.AI","article_published_time":"2022-02-23T20:34:21+00:00","article_modified_time":"2025-03-17T12:39:39+00:00","og_image":[{"width":2048,"height":1365,"url":"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/02\/pexels-erik-mclean-4582541-e1656153098676.jpg","type":"image\/jpeg"}],"author":"Paulo Maia","twitter_card":"summary_large_image","twitter_creator":"@nilg_ai","twitter_site":"@nilg_ai","twitter_misc":{"Written by":"Paulo Maia","Est. reading time":"10 minutos"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/#article","isPartOf":{"@id":"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/"},"author":{"name":"Paulo Maia","@id":"https:\/\/nilg.ai\/#\/schema\/person\/a860b9395bd25c5090160a3ac30e980d"},"headline":"Teaching Models With Free Data","datePublished":"2022-02-23T20:34:21+00:00","dateModified":"2025-03-17T12:39:39+00:00","mainEntityOfPage":{"@id":"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/"},"wordCount":1640,"commentCount":0,"publisher":{"@id":"https:\/\/nilg.ai\/#organization"},"keywords":["Deep Learning","Machine Learning","Self-Supervised"],"articleSection":["Technical"],"inLanguage":"pt-PT","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/","url":"https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/","name":"Modelos de Ensino com Dados Gratuitos - NILG.AI","isPartOf":{"@id":"https:\/\/nilg.ai\/#website"},"datePublished":"2022-02-23T20:34:21+00:00","dateModified":"2025-03-17T12:39:39+00:00","description":"Uma vis\u00e3o geral dos m\u00e9todos de aprendizagem auto-supervisionada utilizando redes neuronais siamesas. Treinar modelos sem r\u00f3tulos para aplica\u00e7\u00f5es da vida real.","inLanguage":"pt-PT","potentialAction":[{"@type":"ReadAction","target":["https:\/\/nilg.ai\/202202\/teaching-models-with-free-data\/"]}]},{"@type":"WebSite","@id":"https:\/\/nilg.ai\/#website","url":"https:\/\/nilg.ai\/","name":"NILG.AI","description":"Crie neg\u00f3cios em constante melhoria com IA","publisher":{"@id":"https:\/\/nilg.ai\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/nilg.ai\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"pt-PT"},{"@type":"Organization","@id":"https:\/\/nilg.ai\/#organization","name":"NILG.AI","url":"https:\/\/nilg.ai\/","logo":{"@type":"ImageObject","inLanguage":"pt-PT","@id":"https:\/\/nilg.ai\/#\/schema\/logo\/image\/","url":"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/03\/logo.svg","contentUrl":"https:\/\/nilg.ai\/wp-content\/uploads\/2022\/03\/logo.svg","caption":"NILG.AI"},"image":{"@id":"https:\/\/nilg.ai\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/twitter.com\/nilg_ai","https:\/\/youtube.com\/@nilg_ai","https:\/\/www.linkedin.com\/company\/nilg-ai\/"]},{"@type":"Person","@id":"https:\/\/nilg.ai\/#\/schema\/person\/a860b9395bd25c5090160a3ac30e980d","name":"Paulo Maia","image":{"@type":"ImageObject","inLanguage":"pt-PT","@id":"https:\/\/nilg.ai\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/9e3e7cb701fd6dcee37f7a4e9f73ef5a27f5741062c4cb179724a1052ea23086?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/9e3e7cb701fd6dcee37f7a4e9f73ef5a27f5741062c4cb179724a1052ea23086?s=96&d=mm&r=g","caption":"Paulo Maia"},"url":"https:\/\/nilg.ai\/pt\/author\/paulo\/"}]}},"_links":{"self":[{"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/posts\/1451","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/comments?post=1451"}],"version-history":[{"count":10,"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/posts\/1451\/revisions"}],"predecessor-version":[{"id":4283,"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/posts\/1451\/revisions\/4283"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/media\/1468"}],"wp:attachment":[{"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/media?parent=1451"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/categories?post=1451"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nilg.ai\/pt\/wp-json\/wp\/v2\/tags?post=1451"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}