<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Marcelo De Oliveira Rosa Prates on Medium]]></title>
        <description><![CDATA[Stories by Marcelo De Oliveira Rosa Prates on Medium]]></description>
        <link>https://medium.com/@marcelodeoliveirarosaprates?source=rss-31b69a2db687------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*ClIDAU7HIt7hM7gp7FnrMg.jpeg</url>
            <title>Stories by Marcelo De Oliveira Rosa Prates on Medium</title>
            <link>https://medium.com/@marcelodeoliveirarosaprates?source=rss-31b69a2db687------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 03:44:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@marcelodeoliveirarosaprates/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Creating Mandalas with Floral Automata]]></title>
            <link>https://medium.com/@marcelodeoliveirarosaprates/creating-mandalas-with-floral-automata-f54219c9b837?source=rss-31b69a2db687------2</link>
            <guid isPermaLink="false">https://medium.com/p/f54219c9b837</guid>
            <category><![CDATA[design]]></category>
            <category><![CDATA[jupyter-notebook]]></category>
            <category><![CDATA[art]]></category>
            <category><![CDATA[cellular-automata]]></category>
            <category><![CDATA[generative-art]]></category>
            <dc:creator><![CDATA[Marcelo De Oliveira Rosa Prates]]></dc:creator>
            <pubDate>Wed, 25 Dec 2019 18:48:24 GMT</pubDate>
            <atom:updated>2019-12-28T14:10:50.142Z</atom:updated>
            <content:encoded><![CDATA[<p>The image below shows the process for constructing a <a href="https://en.wikipedia.org/wiki/Sierpi%C5%84ski_triangle">Sierpinski Triangle</a>. The Sierpinski Triangle is a <a href="https://en.wikipedia.org/wiki/Fractal">fractal,</a> meaning that its parts are self-similar to the whole. A Sierpinski Triangle is composed of three smaller Sierpinski Triangles, each one of which is composed of three even smaller Sierpinski Triangles and so on, to infinity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tnl9QkKSFkPxxycQcI70JA.png" /><figcaption>Sierpinski Triangle evolution. From <a href="https://en.wikipedia.org/wiki/Sierpi%C5%84ski_triangle#/media/File:Sierpinski_triangle_evolution.svg">Wikipedia</a>.</figcaption></figure><p>The Sierpinski Triangle is possibly one of the most well-known fractals, along with the <a href="https://en.wikipedia.org/wiki/Mandelbrot_set">Mandelbrot Set</a>. But what not everybody knows is that one can draw it using an astoundingly simple rule:</p><ol><li>Start with a row of black cells. Paint a single cell white:</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/683/1*0nC9D8IZDMHZx4TlPtIR3w.png" /></figure><p>3. Create a second row. In order to paint each cell of this row, look to the color of the cell at the same position (as well as its left and right neighbors) in the first row. Paint the cell white if the cell and its neighbors are in any of these configurations:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/907/1*TaCVlcxgBh34pSjrCOFyPw.png" /></figure><p>For any of the other possible configurations, paint the cell black:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/907/1*EvJuPjnNFwzLHQaPBNG3xQ.png" /></figure><p>A more compact way of visualizing this rule is the following:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x-BqxFJD5JKPCjPqv20wcA.png" /><figcaption>Rule 90</figcaption></figure><p>4. Repeat the process for the third row, fourth row, fifth row and so on. This is the result for the first 30 rows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bbe2Stv0VAxvytfv3JGVBQ.png" /></figure><p>For comparison, here are the first 200 rows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5eSjgkIdC3RipT4inqLWmg.png" /><figcaption>First 200 steps of CA rule 90 initialized with a single white cell</figcaption></figure><p>This example shows how simple rules can give rise to complex behavior. It also defines a <strong>cellular automaton</strong> without naming it. Each cellular automaton is defined by its <em>rule</em>, which corresponds to which color it assigns to each configuration of neighbors.</p><p>This rule, for example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NFmMF7WfY4H-OuI7CV6Rng.png" /><figcaption>Rule 30</figcaption></figure><p>Gives rise to a completely different pattern:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EdMvtJ2qhDRzNsT-uq1orQ.png" /><figcaption>First 200 steps of the execution of CA rule 30</figcaption></figure><p>This rule (called rule 30) is quite famous. It’s even argued that some natural patterns are derived from it, for example the pattern in the shell of the <a href="https://en.wikipedia.org/wiki/Conus_textile">Conus Textile</a>, a small venomous sea snail native to the tropical Indo-Pacific.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-RyCpNiotDkZkwwwhntugg.jpeg" /><figcaption>The pattern in the Conus Textile shell is reminiscent of CA rule 30. From <a href="https://en.wikipedia.org/wiki/Rule_30#/media/File:Textile_cone.JPG">Wikipedia</a></figcaption></figure><p>One can see similar sub-patterns such as all those pointing-down triangles of varied sizes, but the overall feel of the two patterns is quite different. That’s partially due to the fact that cellular automata are discrete and pixelated, while nature is continuous at the macroscopic level.</p><p>A way to make the CA execution look continuous is to blur it. If we blur the execution of rule 30 and draw the result in black-and-white, the output looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lOElwvodtnajuUKqqZNjgg.png" /></figure><p>Another thing you can do is to draw cellular automata on polar coordinates (I’ll explain later how this fits in my general concept for floral automata). Imagine taking the cellular automaton tape and bending it so it forms a circle. And then imagine stacking a series of smaller, concentric circles inside of it and having the cellular automaton’s evolution run in the inwards direction. Like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pndg87W7YLW2oymonurcUA.png" /><figcaption>Firts 10 steps of CA rule 90 drawn on polar coordinates. We start with a single white cell on the outermost layer and execute the CA in the inwards direction.</figcaption></figure><p>An advantage of drawing things in polar coordinates is that you can impose radial symmetry (which is pleasing to the eye) very easily. For example, instead of starting with a single white cell, we can paint a white cell every 8 cells. This is the result of running CA rule 90 on this initial tape:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qzl0gYCV_99948uXDQ3nfQ.png" /><figcaption>CA rule 90 run on an initial 128-sized tape where 1 in every 32 cells is white.</figcaption></figure><p>You may have understood by now why the title of this post is <strong><em>Floral </em></strong><em>Automata</em>. Just for comparison, let’s see how rule 30 looks like:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aZRkAJoBXZbkxZcLldYr4g.png" /><figcaption>CA rule 30 run on an initial 128-sized tape where 1 in every 32 cells is white.</figcaption></figure><p><em>Incredibly ugly</em>. That’s because, once again, CAs are discrete entities. Pixelated images already look bad on Cartesian coordinates, but they look even worse on polar as you approach the origin.</p><p>But never mind! Let’s merge the two insights (blurring and radial symmetry) to see what we can get.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HExwzEa8Z3R85fI5WxxNog.png" /><figcaption><strong>Blurred</strong> CA rule 30 run on an initial 128-sized tape where 1 in every 32 cells is white.</figcaption></figure><p>It’s almost unbelievable just how much we can improve the result just by blurring it. Here are some other examples:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6cfThaSmTEcO3W_-7t2IJw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5ct9FLrimgNuprT9bmRzfA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Oq5mVPDFLS857e1SlcXNQQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m4C0kSh5QLFywSGMAMjC0Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*du-N9CYtwidYcvfPVzoo5A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Akon1wC-zOALHpjNiK5_Bw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WyMzhhB9OxzolaxfVKc4JA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ff9S4mZKfCzuhajbWwkfwg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C9YWUBILW1y_Ejc2jcTK5Q.png" /><figcaption>Some other examples of floral automata. From the top right corner (left to right, top down): CA rules <strong>9</strong>, <strong>13</strong>, <strong>45</strong>, <strong>54</strong>, <strong>57</strong>, <strong>60</strong>, <strong>62</strong>, <strong>73</strong>, <strong>105</strong>, <strong>107</strong>.</figcaption></figure><p>There are some parameters with which we can play around here. For starters, the number of iterations of the CA (corresponding to the number of concentric circles in polar coordinates) defines the petal width:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YWxSCpj5jn7ymKP6Oabgbg.gif" /><figcaption>The animation shows the drawing of floral automaton rule 30 for different iteration depths (8–256).</figcaption></figure><p>We can also use more than 2 colors to draw the image after blurring. And we can use colors other than shades of gray:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/805/1*I9unaQudVjmEs1PG9zqtKQ.png" /><figcaption><strong>Blurred</strong> CA rule 30 run on an initial 128-sized tape where 1 in every 32 cells is white. This time the CA evolves outwards. We also use 3 colors to draw the figure.</figcaption></figure><p>Another parameter we can play with is the number of symbols in the CA rule. Up until now we have only painted a cell <strong>black</strong> or <strong>white</strong> depending on the configuration of neighbor cells—<strong> </strong>or <strong>0</strong> and <strong>1</strong>, as the colors themselves don’t matter. We can expand on that by allowing cells to be painted with more than one symbol. Below we can see some examples using three symbols: <strong>0</strong>, <strong>1</strong> and <strong>2</strong>. Have a look:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7HM6LXBfsb7RhfPnLPfYiw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rK5VYp2xTATtNeNyc9Y68g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RmjJKV9wlRO4va5voOulTA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CG69kervTMBn4QJF75RP1g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bU_kd0aM7a5ptg3ZQOAFMg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gk0onXND4NzSfcJuqKME1g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cpaRh1qpSd11TyCpQ7grjw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O8mPpa-AbBa2wcAXwUjyIg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*18wD4ebdhwEvc00nAJbMPg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EyK9BYXHBghXdS7IViJv7Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-17acVAKFpfqqLVjd-B-eA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SOw4kCu2wZZlO6l-lowaFA.png" /><figcaption>Examples of floral automata with three symbols. From the top right corner (left to right, top down): CA rules <strong>5089621433348</strong>, <strong>990299038253</strong>, <strong>6477086403745</strong>, <strong>598640925203</strong>, <strong>6017109809141</strong>, <strong>4252695808481</strong>, <strong>5732953757817</strong>, <strong>6036516668408</strong>, <strong>4486031731065</strong>, <strong>1473950395202</strong>, <strong>7000900136858</strong>, <strong>6106292208893</strong>.</figcaption></figure><h4>A small note on the rule numbers (or codes)</h4><p>At this point you might be asking what on earth are these numbers I’m using to identify each rule. What does <strong>5089621433348</strong> mean and why are the codes for 3-symbol CAs so large compared to the codes for 2-symbol CAs that we’ve seen before (such as rule <strong>30</strong>)?</p><p>Take for example rules <strong>0</strong>, <strong>1</strong> and<strong> 2 </strong>(note that <strong>white</strong> represents the symbol <strong>1</strong> and <strong>coral</strong> represents the symbol <strong>0</strong>):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SvXXGmyzkV7jXWSG9rGpXA.png" /><figcaption>CA rule 0 (2 symbols)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rw45SYsRMaA2iDvjozQdWw.png" /><figcaption>CA rule 1 (2 symbols)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Cpptmi8ijuXw_sIuU99gUQ.png" /><figcaption>CA rule 2 (2 symbols)</figcaption></figure><p>The sequence of the 3-neighbor configurations is always the same (top row), so we can identify a rule using only the sequence of produced symbols (bottom row). For example, rule <strong>0</strong> produces symbols <strong>00000000</strong> for the first 8 configurations (respectively). Rule <strong>1</strong> produces symbols <strong>00000001</strong> for the first 8 configurations, while rule <strong>2</strong> produces symbols <strong>00000010</strong>. Can you see the pattern? Rule 3 would produce <strong>00000011</strong>, rule <strong>4</strong> would produce <strong>00000100</strong> and so on. The code of each rule is simply the sequence of its symbol productions interpreted as a binary number (<strong>0</strong>, <strong>1</strong>, <strong>11</strong> and <strong>100</strong> are <strong>1</strong>, <strong>2</strong>, <strong>3</strong> and <strong>4</strong> in binary).</p><p>Ok, but what about when we have more than 2 symbols? For example rule <strong>4486031731065</strong>, which generates the 9th example in the last gallery (the red/cream/gray flower)?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uQ5WixfsVkpwLe7Edy8JaA.png" /><figcaption>CA rule <strong>4486031731065</strong> (3 symbols)</figcaption></figure><p>Well, the principle is the same. The only difference is that now we encode the sequence of produced symbols in <strong>base 3</strong> rather than <strong>base 2</strong>. The ternary number <strong>012201112010120020212212021</strong> corresponds to <strong>4486031731065</strong> in decimal. Rule numbers are (usually) larger with 3 symbols because there are a lot more of 3-neighbor configurations with 3 symbols compared with 2 (3³=27 vs 2³=8) and also because there is 1 more possibility for each symbol production. With 2 symbols, we can choose to fill each one of all 2³ positions with either a 0 or a 1. But with 3 symbols, we can choose to fill each one of all 3³ positions with either a 0, a 1 or a 2. So we have, in total, 2^(2³)=256 CAs with 2 symbols and 3^(3³) = 7,625597485×10¹² CAs with 3 symbols.</p><p>Ok, now back to the good part. Another parameter we can play with is the neighborhood size.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vW9fPGqWul6S8WgkldkcXw.png" /><figcaption>Example of a CA rule (2 symbols) with a neighborhood size of 5</figcaption></figure><p>Below are some examples using 2 symbols and neighborhood size 5:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gktr6Eu2Eie2RhPo67WIfg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FE2LF0Jvc5T5Cp-wvZK1eA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C5V_xrjsrexbvRHaQk27fw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*57h-OTwpBG9nZZ0MhVoDlQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*amJb4b5WJek0wEU-R0SxXA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7Q4GhqHPABxvOxS7LSLQyg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hnBwB4G3JfjWqNldbv0QHA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L3fmc5iQoRui2dLuodw--w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9D1xrDBYZSBEeSDp4ENwGg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d0TBNPCP3AmOOEb69bqT2g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6EoUd0PbC3-Z3mZwN1vbWw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fmh8w5pnWeCP6nox_n8blQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Jvw2QAQ2yrMwGzk0CxbuXA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EJyPOn9P8DWES9EgD6tIxQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Kbh_eQjWsxRQlnDA-3ebgQ.png" /><figcaption>Examples of floral automata with 2 symbols and neighborhood size 5.</figcaption></figure><p>Obviously we can increase the number of symbols and the neighborhood size at the same time, giving rise to even more complex patterns:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3DZBt6hLceorvZrvCqfcyw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JYIoyCdr5mZ29R6S94JxSw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S93q8vu1qXFZRnl8N8SycQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5ALsYsppR3jYe0rEk6DmQQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u3O32RG-R-DPHhEWe18D_w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ve95aLTn0PO7Jjv9VpxVMQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RdHieaDTRJzfwu7j9CXYow.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w50BkBu6WFa3N6p3TEjEXg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c9Y62yHDlgTeVzv3I2QziA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D1IFVK5ngyAndByo__ihIA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LzKt0KfbT691jvhaxNH8Ug.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FwH20An4ZCqmQVYswxLQLA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C1Am7spc4p0Coe5j-VnPvA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QMW0LnwBnla4uIKnoKbZ_g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JH9IvMx8AVQ6u26yLnlkZQ.png" /><figcaption>Examples of floral automata with 3 symbols and neighborhood size 5.</figcaption></figure><p>It’s tempting never to end this post, as there are endless parameter combinations and as the number of CAs for each of them, while not infinite, is absolutely huge. But before I go, I wanna share with you the Jupyter Notebook I’ve been using to generate this images. It comes complete with a interactive tool for playing with the parameters yourself and creating your own floral automata. It’s available on <a href="https://github.com/marceloprates/Floral-Automata-Jupyter">Github</a> or directly at <a href="https://colab.research.google.com/github/marceloprates/Floral-Automata-Jupyter/blob/master/Floral%20Automata.ipynb">Google Colab</a>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/570/1*3_zDBImzW8cESGOKJwtrlw.png" /></figure><p>Also before I go, a final reminder that you can play with all of the parameters at the same time. I was able to obtain very nice results by randomizing the number of symbols, the neighborhood size, the number of colors and the direction (inwards / outwards) of the CA evolution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lldXhzgRKB8S1zcc-Vl3Vg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0RY0aVbTX8UhvO7A3kumDA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZbE72vd-qOqgqOx92GbgHg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Yhx0RB5U5E6T3sVqKckc1A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_evF1R3mAXXIu8OYlSUJqA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z3Uabpbg9ShBgaaL1QYjkA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4yaaGPzv4SrUBOpyryn1Dw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eyWerNBhC9mfel5tNJboYw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hESgQwcIbvIaR7PPW4fipw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H6mOFwN7Z4xie8KMptQiMg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8Ylj0rKsaAcAM-rW4kWOqQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YnNCzfvIt8CQIBcjxZpttw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l5TXpUyNJtlQVgJg6YKJFg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2haM8F5FEhp88epPuIFt5g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mWKEvhay3hCd5NDw6szsWQ.png" /></figure><p>Hope you’ve enjoyed!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f54219c9b837" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Khipu 2019 and the State of AI in Latin America]]></title>
            <link>https://medium.com/@marcelodeoliveirarosaprates/khipu-2019-and-the-state-of-ai-in-latin-america-ae9457b49405?source=rss-31b69a2db687------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae9457b49405</guid>
            <category><![CDATA[khipu]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[latin-america]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Marcelo De Oliveira Rosa Prates]]></dc:creator>
            <pubDate>Sat, 14 Dec 2019 19:34:21 GMT</pubDate>
            <atom:updated>2019-12-14T23:24:13.690Z</atom:updated>
            <content:encoded><![CDATA[<h3>Khipu 2019 and the State of AI Research in Latin America</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sa8Q0rSM_Xo-N7ju.jpeg" /><figcaption>Khipu 2019 Attendees, Organizers and Speakers pose for a group photo at the courtyard of the Facultad de Ingeniería (Universidad de la República).</figcaption></figure><p>Between Nov 11 and Nov 15 I had the opportunity of flying to the northeastern bank of the <a href="https://en.wikipedia.org/wiki/R%C3%ADo_de_la_Plata">Río de la Plata</a>, home to the southernmost capital in the Americas: <a href="https://en.wikipedia.org/wiki/Montevideo">Montevideo, Uruguay</a>. I was there to attend <strong><em>Khipu 2019</em></strong><em> </em>— the first-ever large scale Latin American Meeting in Artificial Intelligence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/311/0*o-N9d_j-FuvCOEGL.jpg" /><figcaption>Montevideo, Uruguay. <a href="https://en.wikipedia.org/wiki/Montevideo#/media/File:Collage_Landmarks_of_Montevideo.jpg">Source</a></figcaption></figure><p>The benefits of gathering the top Latin American AI researchers together for such an event are manifold, and <a href="http://khipu.ai/">khipu.ai</a> was able to summarize them in a powerful way. The two core ideas here are that:</p><ol><li>Home to many developing countries, Latin America can benefit hugely from AI in <strong>tackling social, environmental and economic problems more efficiently</strong>, and</li><li>As with any research field (and perhaps especially so), <strong>AI thrives on diversity of thought</strong>. Thus, inviting the many developing countries in Africa and Latin America to the discussion is advantageous not just for these continents but for science as a whole.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RhDwSWz5owASA8-OcHMHtw.png" /><figcaption>The principles behind Khipu 2019 (available at <a href="https://khipu.ai/">https://khipu.ai/</a>)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/367/1*qgHYvOAhqIEezSIXuU2vWg.png" /></figure><p>The event is not without its share of symbolism. Khipu AI is named after <strong>Khipus </strong>(or<strong> Quipus</strong>), which are ancient recording devices, made from llama or alpaca wool and used for communication and counting purposes by the <a href="https://en.wikipedia.org/wiki/Inca_Empire">indigenous peoples of the Andes</a>. “Khipu” means “knot” in the <a href="https://en.wikipedia.org/wiki/Quechua_language">Quechua Language,</a> and that’s because numbers were represented in it by a series of knots along a string. The Incas used a decimal position-based number system just as ours: <strong>1</strong> knot in the 100s place plus <strong>3</strong> knots in the 10s place plus <strong>2</strong> knots in the 1s place makes for <strong>100+30+2=132</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/625/0*OAgL5oIfEcf49_DQ.jpg" /><figcaption>An example of a <em>quipu</em> from the Inca Empire, currently in the <a href="https://en.wikipedia.org/wiki/Larco_Museum">Larco Museum</a> Collection. <a href="https://en.wikipedia.org/wiki/Quipu#/media/File:Inca_Quipu.jpg">Source</a></figcaption></figure><p>As recording and communication devices, Khipus played an important part in the administration of the <a href="https://en.wikipedia.org/wiki/Inca_Empire">Inca Empire</a> and the <a href="https://en.wikipedia.org/wiki/Kingdom_of_Cusco">Kingdom of Cusco</a>. This is due to the fact, as Nando de Freitas beautifully put it during the event, that <em>“The original Khipu was a tool which extended the cognitive abilities of the Andean peoples which developed it”</em>. Five hundred years later, artificial intelligence strives to do the same. For this reason, I think Khipu is a very fitting and powerful symbol for AI in Latin America.</p><p>In the days following the arrival of hundreds of Latin American researchers in Montevideo, turmoil was gathering across the Andes Mountains in the Bolivian capital La Paz. Under suspicious conditions, the electoral polls declared current President Evo Morales the winner with <strong>47.08%</strong> of the votes, slightly above the required ten-point margin over its nearest competitor, Ex-President Carlos Mesa. An <a href="https://www.washingtonpost.com/world/bolivia-to-hold-new-elections-after-protests-and-international-criticism/2019/11/10/4778e842-03b2-11ea-ac12-3325d49eacaa_story.html">audit</a> by the <a href="https://en.wikipedia.org/wiki/Organization_of_American_States">Organization of the American States</a> suggested “clear manipulation” in the election as widespread protests claiming electoral fraud raged (Morales was running for a fourth term, backed by the Supreme Tribunal of Justice but in clear violation to the Bolivian Constitution). In the following days, Morales <a href="https://www.bbc.com/news/world-latin-america-50365340">announced new elections</a> but <a href="https://www.nytimes.com/2019/11/10/world/americas/evo-morales-bolivia.html">was forced to step down by the military</a>. After a series of resignations, the fifth in the line of succession, <a href="https://elpais.com/internacional/2019/11/12/america/1573566340_453048.html">Senator Jeanine Áñez became interim President</a>. <a href="https://apnews.com/d88e67e214b04dcd92dc58caa884cdaf">Morales was forced to seek political asylum in Mexico</a> (and later Argentina).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/307/1*AS2OodXF_AnSZg2tKVNFqg.png" /><figcaption>Anti-Morales protesters burning down a Wiphala flag, associated with the native peoples of the Andes.</figcaption></figure><p><a href="https://apnews.com/530280a8d9674f58ad19af8d3f00edee">Anti-Áñez protesters were met with police brutality</a> while her supporters targeted indigenous peoples and burned their Wiphala flag. The shift of power elevated a politician <a href="https://elcomercio.pe/mundo/latinoamerica/jeanine-anez-estos-son-los-agresivos-tuits-contra-originarios-e-indigenas-que-borro-la-presidenta-interina-de-bolivia-noticia/">with a history of racist remarks</a> to the Presidency of Bolivia, a country where 41% of the population identifies as indigenous.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/711/1*jA7XzfPulRotlhx21Mvttg.png" /><figcaption>Bolivian attendees with Nando de Freitas at Khipu 2019.</figcaption></figure><p>Unfortunately, it’s not only Bolivia who’s experiencing a political crisis. Virtually every South American country is either experiencing civil unrest (<strong>Chile</strong>, <strong>Peru</strong>, <strong>Colombia</strong>, <strong>Ecuador</strong>), under the threat (<strong>Brazil</strong>, <strong>Bolivia</strong>) or reality (<strong>Venezuela</strong>) of a dictatorship or at the very least under severe economic problems (<strong>Argentina</strong>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j7KMkou1nZw4ftZiUrwZ_A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/304/1*WSkMm_wuEossM7guDpPZrA.png" /></figure><p>When we paint on a map of South America all countries experiencing political crises (red) and all countries experiencing severe economic problems (yellow), the figure shows a continent plunged into its most chaotic moment in decades. It is not without irony that Khipu’s host, Uruguay, is an exception to the rule: an example of democracy for Latin America and for the world.</p><p>It’s obviously not easy to explain why 2019 was such a difficult year for Latin America, but some authors attribute it to the volatility of the prices of commodities. The economies of many Latin American countries prospered under oil, for example. But the demand for oil won’t keep increasing forever as it did in the last decades.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/689/0*nvUQqjboG_g1Kk0x.png" /><figcaption>Historical trend of the prices of different commodities. <a href="https://www.weforum.org/agenda/2015/04/what-100-years-of-commodity-prices-tells-us-about-the-future/">Source</a></figcaption></figure><p>What does seem to generate increasing revenue year after year is <strong>automation</strong> and <strong>artificial intelligence</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/733/0*Fp_Jmkzo4xSOoKI7.jpg" /><figcaption>Projected AI revenue by region (2015–2024). <a href="https://emerj.com/ai-sector-overviews/valuing-the-artificial-intelligence-market-graphs-and-predictions/">Source</a></figcaption></figure><p>AI is a clear way out of the crises for Latin America’s developing countries. But unfortunately, Latin American representation in AI is quite limited, as the graph below exemplifies. In the last decade Latin America was able to publish <strong>only 11 papers out of &gt;3000 (~0.37%) at NeurIPS</strong>, the most prestigious deep learning conference in the world. In my opinion, this is a testament to the monumental importance of events such as Khipu. We need to increase the representation of Latin America in AI, and we need to do it fast.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/648/1*2YNrqmdV8UlGFo7njm65mg.png" /><figcaption>World map distorted according to the representation of each country in terms of NeurIPS publications. <a href="http://www.deeplearningindaba.com/blog/missing-continents-a-study-using-accepted-nips-papers">Source</a></figcaption></figure><h3>Khipu Program</h3><p>This below is my Google Calendar activities during the Khipu week. It was a very intense event, with a large number of talks you could simply not miss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iTzIXwz_oi46dtG3rsKUiQ.png" /></figure><h4>Monday (Nov. 11):</h4><p>From the first day we had some amazing talks, starting with Guillermo Sapiro (Uruguayan <strong>🇺🇾</strong>) <a href="http://tv.vera.com.uy/video/54702">talking</a> about his and his colleagues work on using machine learning to speed up the diagnostic of autism in infants. René Vidal (Chilean 🇨🇱) presented a very interesting <a href="http://tv.vera.com.uy/video/55357">talk</a> about the mathematics of deep learning optimization.</p><p>We also had parallel practical sessions on <strong>Convolutional Neural Networks</strong> and <strong>Optimisation for Deep Learning</strong>. The practicals were ministered using Colab Notebooks, all of which can be acessed <a href="https://github.com/khipu-ai/practicals-2019">here</a>. I enjoyed the optimisation session very much.</p><p>Monday ended on a very satisfying note with a show featuring Uruguayan tango!</p><h3>Mauricio Mazuecos on Twitter</h3><p>Finishing day 1 of @Khipu_AI #khipu2019 with some tango! 🖤 https://t.co/6KDHel5dCv</p><h4>Tuesday (Nov. 12):</h4><p>The second day included a very anticipated talk by Ian Goodfellow. Previously a researcher at Google Brain, now working at Apple and considered by many as the father of Generative Adversarial Networks (GANs), Ian spoke about some very useful applications of GANs in many areas, including the effort by Apple on using GANs to train an eye-gaze predictor in the absence of a large enough amount of labelled examples. The idea is to train a GAN to map from a synthetic, CG simulation of an eye (gazing in a particular direction) to a photo-realistic image. This can be done by training a GAN to generate samples from a distribution of real eye pictures and coupling it with a self-regularization loss whose purpose is to minimize the difference between the synthetic and generated images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/495/1*5lghW79yrdCGxfH_8aukIw.png" /></figure><p>Tuesday ended with an amazing panel on <a href="https://drive.google.com/file/d/1Qc4NXeXJj-whjqmZIMjLr6RPbQOFgfnt/view">“How to Write a Great Research Paper”</a> featuring Nando de Freitas, Claire Monteleoni, David Lopez-Paz and Martin Arjovsky.</p><h3>lesly🤖🍫👩🏾‍💻@ #DevFest19 🇳🇮🇵🇭🇧🇴 on Twitter</h3><p>Start writing! Identify your key idea! Tell one story! #Khipu2019 #howtowriteagreatpaper @Khipu_AI</p><h4>Wednesday (Nov. 13):</h4><p>Luciana Benotti (Argentinian <strong>🇦🇷</strong>) spoke about challenges faced by Latin American researchers, which include expensive conferences, decreasing funds for research and low access to cutting-edge equipment. She chose to devote her talk to tell the success stories of her students, summarizing the coping mechanisms they’ve successfully implemented to produce impactful research — such as collaborating with researchers from rich countries to obtain bilateral funding or addressing hard, but low-resource tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r_cbnxW4HqGj2GoFvdNgSg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/152/1*LQO2WhMKhCPAsLp-9Zy8wA.png" /></figure><p>Wednesday also featured one of the highlights of Khipu 2019: a talk by Yoshua Bengio, one of the 3 recipients of the 2019 Turing Award along with Geoffrey Hinton and Yann LeCun and considered by many one of the fathers of modern deep learning. Bengio spoke about his perspectives on the future of AI research. According to Bengio, we have made a lot of progress in the recent years with insights such as <strong>attention </strong>and<strong> GANs</strong>, but are still far behind with respect to human level intelligence. Among other things, that’s because modern DL models are trained to solve very specific tasks, while human intelligence is general. Bengio suggested that a way forward is to focus on learning agents capable of interacting with the world as humans do. He pinpointed “meta-learning” (learning to learn) as a key direction to improve the state-of-the-art in machine learning, and devoted a lot of attention to the subject of causality (humans learn causal relationships about the world, but current ML focuses on correlations). On a final note, Bengio spoke about the importance of discussing the ethics of AI and making sure that AI is used for social good.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/258/1*9XpD9HjrFvTPat1wsjSzBQ.png" /></figure><p>Following a brief discussion about the purpose, history and cultural heritage of the original Khipus, Nando de Freitas presented a stellar lecture on Reinforcement Learning, building up from the basics. Nando briefly commented on the gaps in Latin American mathematical education while presenting the lecture, and stopped it more than once to cover concepts lacking in our formal education.</p><h4>Thursday (Nov. 14):</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/236/1*Yb7A-nnuqQag9RNX3z7DRQ.png" /></figure><p>The fourth day included my favorite talk of Khipu 2019, presented by David Lopez-Paz. It was named “Causality and Generalization” and spoke a lot about what the field of ML can gain by learning to train models which capture causal relationships instead of (possibly spurious) correlations. David spoke about how modern image classification tools fail at identifying causal relationships even in very simple tasks. For example, the “cattle” class is so heavily correlated with grasslands that modern tools will fail to recognize a cow in a beach scenario.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9n2kxWsaGNc0x3JAwKPKRg.png" /></figure><p>David spoke at length about how the biggest lie in ML is assuming that the test distribution will be similar to the train distribution. To circumvent this issue, causal models are desperately needed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/770/1*Y0IG0lCbcBJMxGO6QK6RQQ.png" /><figcaption>David’s highlights on causality in ML</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*at6nh1rkfXa474bCBc_WwQ.png" /></figure><p>One of the highlights of the event was the Women in AI panel, hosted by Google. It featured Brazilian 🇧🇷 researcher Sandra Avila, Google Scientist Chelsea Finn, Uruguay’s 🇺🇾 Ex-Education minister and current dean of Facultad de Ingeniaria Maria Simon, Apple’s Giuilia Pagallo (Venezuelan 🇻🇪), Uruguay’s 🇺🇾 current Minister of Industry, Energy and Mining Guillermo Moncecchi and DeepMind’s Nando de Freitas. Maria Simon spoke at length about the need for integrating non-minority groups when discussing about diversity in research, arguing that it’s no use talking only to women about the importance of gender diversity: <em>they’re already convinced</em>. When I registered for this event, I have to tell I was afraid I’d be the only man in the room, but I wasn’t. It was very nice to see so much men interested in hearing more about gender diversity in AI.</p><h4>Friday (Nov. 15):</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TgD7t3d1Ve1IyLZG.jpg" /><figcaption>Google AI lead Jeff Dean</figcaption></figure><p>Friday was the last day of the event, but it included some very interesting talks. The current lead of Google AI, Jeff Dean, which is responsible for many of Google’s pivotal projects such as MapReduce and Tensorflow, presented a talk about the company’s efforts to use AI for social good in healthcare and for environmental or education purposes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*BYGJC6UxM6Ku88PV.jpg" /></figure><p>Oriol Vinyals spoke about DeepMind’s efforts to achieve superhuman performance on the game of <strong>Starcraft</strong>, a far more difficult benchmark than the Chinese boardgame of Go (DeepMind achieved superhuman status after defeating the long-reigning champion, Lee Sedol), detailing the model’s architecture (which includes a ResNet, a Transformer, Pointer networks, among other architectures) and the company’s strategy for training and evaluating it against human players.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*h0S1stRgDB11lM6lRz3RWg.png" /></figure><p>Finally, Jeff Dean returned to the stage to give a historical perspective of the current decade (2010s) with respect to machine learning. It was a beautiful thing to see just how much the ML community has achieved in so little time. Jeff brought graphs showing the exponential rate of the # of ML papers produced each year. Moore’s law may have ended, but it is being replaced by something else.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/497/1*tKDapRijSUKSmlpFoxMRoA.png" /></figure><p>From 2011 to 2016, the state-of-the-art in image classification rocketed from a <strong>26%</strong> error rate to <strong>3%</strong> (a 4x drop in a 5 year period). In the meantime it achieved superhuman performance, as the human error rate is typically <strong>5%</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/594/1*xg-BowNFGdkARgZ4oa2Q1Q.png" /></figure><p>Dean talked about the Grand Engineering Challenges for the 21st Century, and how Google is working to solve a large number of them using AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/580/1*8wKW3KuF4sBw18In0xO_Ag.png" /></figure><p>Google has collected a series of achievements in the healthcare field, such as managing to diagnose diabetes from retinal images better than trained ophthalmologists. The researchers also realized that the same technique could be used to predict a set of health indicators, such as cardiovascular risk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/844/1*9I_jucVQHLvoU4kYUPSbbA.png" /></figure><p>Jeff also spoke about something which I’m particularly excited about (as my research focuses on Graph Neural Networks for symbolic problems). Traditionally, we use machine learning to tackle problems for which a solution is hard or virtually impossible to articulate in a procedural manner. We cannot describe how to classify an animal as a cat or a dog given a matrix of pixels, so we use convolutional neural networks to do that. A very interesting research direction is training models to solve problems for which we <strong>already have</strong> an explicit solution, but one which consumes <strong>too much resources</strong>. Jeff exemplified this with a work which has managed to predict the chemical properties of molecules with similar accuracy and 10 thousand times faster than traditional techniques.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r6IvQ30-pXXqS86ymeNg9Q.png" /></figure><p>He also talked about Google’s efforts on advancing the state-of-the-art in connectomics (that is, reconstructing the network structure of an animal’s brain given brain imaging slices).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/396/1*V_OhXBNun2knfVA0Jbn-bQ.png" /><figcaption><a href="https://www.smithsonianmag.com/smart-news/weve-put-worms-mind-lego-robot-body-180953399/#B5dToGgrBUHxMkr2.99">We’ve Put a Worm’s Mind in a Lego Robot’s Body</a></figcaption></figure><p>Today we can already port a C. Elegans brain into a Lego Mindstorms robot, but using state-of-the-art DL techniques we can expect to be able to do the same for animals with much larger brains in the near future.</p><p>Jeff also talked a little bit about the <em>attention revolution</em> in sequence-to-sequence models and natural language processing, with a focus on the transformer architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/631/1*b3YDb5xe_6F6rj0xj-BPqw.png" /></figure><p>Another topic in the talk was <strong>AutoML</strong>, and I have to admit that at this point I feel like the audience of ML engineers felt like we’re living in a science fiction scenario. AutoML consists in automating the part of ML research devoted to finding efficient DL architectures. This is done by training a reinforcement learning agent to navigate the space of model architectures to find the ones best suited for training.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/422/1*SRYqlE1Pk4Dz8m2oGmsk7A.png" /></figure><p>To me, the surprising thing is how AutoML manages to find architectures yielding models which are not only <strong>more accurate</strong>, but also <strong>faster</strong> than those produced by ML experts. AutoML has produced a totally different Pareto curve than those we get from human experimentation. The trade-off between speed and accuracy is much better for AutoML.</p><p>Jeff spoke about how ML is reinventing how we build computers, from TPUs to small chips used for efficient inference (e.g. Google’s Coral).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/633/1*K4WEFsLr8ascXF7eiug9Cg.png" /></figure><p>Finally, Jeff talked about Google’s vision for the future of ML: not necessarily bigger and more accurate models requiring a lot of memory and computation nor smaller and less accurate models easily deployable in a small device, but an hybrid solution. Gigantic models, but very sparsely activated and relying heavily on intelligent routing (also learned). These models would be composed of many sub-modules, each specialized for a given task, and one would be able to activate them on-demand depending on the input.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/598/1*pGZeQBl_LpPIMHBOqdSVvw.png" /><figcaption>Interior of Club Uruguay</figcaption></figure><p>The event ended with a magnificent dinner at Club Uruguay. Almost all attendees, organizers and speakers were there having fun and dancing to a variety of Latin American genres spanning from Tango through Reggeaton to Brazilian Funk. I don’t think I’ve ever experienced a similar vibe on a research conference — but once again, I’ve never attended a meeting with so much Latin Americans. We know how to party. And I’m hopeful that we can do it again at Khipu 2020 (maybe in Brazil this time?)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uJU-Cpv2UYVRyd3kErDC3Q.png" /><figcaption>Brazilian attendees at Khipu 2019.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ae9457b49405" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Google Translate e o Viés de Máquina]]></title>
            <link>https://medium.com/@marcelodeoliveirarosaprates/google-translate-e-o-vi%C3%A9s-de-m%C3%A1quina-f1e294d7f4f1?source=rss-31b69a2db687------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1e294d7f4f1</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-bias]]></category>
            <category><![CDATA[aprendizado-de-maquina]]></category>
            <category><![CDATA[viés-de-máquina]]></category>
            <dc:creator><![CDATA[Marcelo De Oliveira Rosa Prates]]></dc:creator>
            <pubDate>Tue, 22 Oct 2019 22:04:11 GMT</pubDate>
            <atom:updated>2019-10-22T22:08:28.193Z</atom:updated>
            <content:encoded><![CDATA[<p><em>OBS.: Esse post é um resumo. O artigo científico completo, co-autorado por Pedro Avelar e Luis Lamb, está disponível (em inglês) no arXiv</em><a href="https://arxiv.org/abs/1809.02208"><em>: https://arxiv.org/abs/1809.02208</em></a><em>.</em></p><p>Na última década, as tecnologias digitais têm sido revolucionadas por uma área de estudo da ciência da computação conhecida como “aprendizado de máquina”. Em termos simples, o objeto de estudo da área de aprendizado de máquina consiste em sistemas capazes de aprender a resolver problemas por meio de exemplos. Essa perspectiva vai na contramão da tradição clássica da inteligência artificial, que consiste na descrição explícita (passo-a-passo) da resolução de um problema pelo programador. Obviamente, aprendizado de máquina (ou <em>machine learning</em>, em inglês), tem muitas vantagens sobre as técnicas tradicionais, pois nos permite produzir um sistema capaz de resolver problemas <em>cujo passo-a-passo da resolução eventualmente nós mesmos não entendemos</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/594/0*yr4YyXHx6YPDW3vN" /><figcaption>Alguns exemplos de imagens dos dígitos 0–9 desenhados à mão, amostradas do dataset <a href="http://yann.lecun.com/exdb/mnist/">MNIST</a>.</figcaption></figure><p>Considere por um instante o problema de classificar um número desenhado à mão. É bastante difícil descrever exaustivamente, passo-a-passo, de que maneira analisar os pixels da imagem de modo a obter uma classificação. Mesmo no caso do dígito zero (um dos formatos mais simples), é preciso considerar todas as possíveis irregularidades e variações na tentativa de se desenhar um círculo. No entanto, hoje em dia esse problema pode ser resolvido de maneira muito rápida e simples usando técnicas de aprendizado de máquina. Basta que tenhamos acesso a um conjunto de dados contendo múltiplos exemplos de imagens anotadas com os dígitos correspondentes e é possível treinar um sistema para mapear as imagens para as suas respectivas classes, sem que seja necessário programar a regra de maneira explícita.</p><p>Vários problemas práticos de relevância social e/ou comercial têm sido atacados usando técnicas de aprendizado de máquina nos últimos anos. Um exemplo notório é reconhecimento facial (reconhecer uma pessoa através de uma foto do seu rosto), uma aplicação usada por sites de redes sociais como o Facebook. Possivelmente o exemplo mais notório de aprendizado de máquina hoje em dia são os próprios mecanismos de busca (ex. Pesquisa Google), que são treinados com dados de usuários para montar um “perfil de busca” para cada usuário e aprender a condicionar os resultados das pesquisas de acordo com o que o usuário já pesquisou no passado. Um treinamento semelhante é adotado por sites de redes sociais (ex. Facebook e Instagram), que são treinados com dados de bilhões de usuários para montar um <em>feed</em>s com o objetivo de maximizar o tempo que o usuário passa conectado na plataforma.</p><p>No entanto, recentemente estudiosos não somente da inteligência artificial como também de vários outros campos — incluindo as ciências humanas, têm alertado para o fato de que o aprendizado de máquina é uma faca de dois gumes. Ainda que hoje em dia as técnicas de aprendizado de máquina estejam por trás de tecnologias e indústrias que movimentam bilhões de dólares todo ano, a dependência progressiva da nossa sociedade em sistemas treinados por meio de exemplos traz consigo diversos problemas. Um dos problemas mais notórios, de evidente impacto social, é o que chamamos <em>viés de máquina</em>. Ele se refere ao fenômeno pelo qual sistemas treinados via exemplo acabam herdando vieses presentes nos dados de treinamento. Em 2017, o sistema <a href="https://www.google.com/photos/about/">Google Photos</a>, que usa reconhecimento facial para organizar de maneira automática a biblioteca pessoal de fotos de cada usuário, foi acusado de <a href="https://www.theverge.com/2015/7/1/8880363/google-apologizes-photos-app-tags-two-black-people-gorillas">classificar pessoas de tom de pele escuro como gorilas</a>. A acusação provocou reações muito fortes, pois cabe a interpretação de que o sistema foi <em>explicitamente programado</em> para classificar pretos e pardos dessa maneira. Não é assim: um entendimento básico de aprendizado de máquina nos permite entender que esse comportamento na verdade tem relação com o fato de que os engenheiros responsáveis pelo sistema o alimentaram com um conjunto de treinamento enviesado, que provavelmente tinha baixa representação de pessoas pretas ou pardas em proporção a caucasianos. A partir do momento que o sistema não é alimentado com esses exemplos, a sua capacidade de identificar pretos ou pardos como seres humanos se torna limitada, e algum nível de erro é esperado. Deste modo, a questão se torna um pouco mais sutil: em se tratando de aprendizado de máquina, os possíveis preconceitos dos engenheiros por trás do sistema não se manifestam ativa, mas passivamente. A indiferença com relação à representatividade de todas as etnias no conjunto de treinamento se manifesta de maneira não-antecipada no produto final, com implicações possivelmente traumáticas para os seus usuários e para a sociedade como um todo. Esse exemplo se soma a um número de outras acusações recentes, como a de um <a href="https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing">preditor de reincidência criminal que, se sugere, acusa pretos e pardos como tendo maior probabilidade de reincidir num crime, se comparados com caucasianos com as mesmas características</a>; e o <a href="http://www.huffpostbrasil.com/entry/iphone-face-recognition-double_us_5a332cbce4b0ff955ad17d50">sistema de reconhecimento do Iphone X, acusado de não ser capaz de distinguir duas pessoas asiáticas diferentes uma da outra</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/690/1*Fdc2j4tSnm8cpIEmg2XwqA.png" /><figcaption>Variações desta captura de tela viralizaram nas redes sociais em 2017. Nele, podemos observar uma sequência de frases escritas no idioma húngaro, que admite pronomes de gênero neutro (no caso, o pronome “ő” pode ser lido tanto como “ele” quanto como “ela”) sendo traduzidas para o português, onde os pronomes têm gênero. O fato curioso é que o sistema traduz algumas profissões com gênero feminino e outras com gênero masculino. O fato de “enfermeira”, “professora”, “faxineira” serem traduzidos com pronome feminino enquanto “médico”, “cientista”, “engenheiro” serem traduzidos com pronome masculino sugere a existência de algum tipo de viés de gênero. Essa captura de tela foi feita em 22/10/2019.</figcaption></figure><p>Em 2017, algumas capturas de tela viralizaram nas redes sociais sugerindo que o sistema de tradução automática <a href="https://translate.google.com/">Google Translate</a> também exibia viés de máquina, específicamente viés de <em>gênero</em>. Na prática, o que se observava era que, quando traduzíamos sentenças de um idioma com suporte para pronomes de gênero neutro (por exemplo, no idioma Húngaro, o pronome “ő” pode ser lido tanto como “ele” quanto como “ela”) para um idioma como o português ou o inglês, em que os pronomes têm gênero, o sistema traduzia cada sentença com pronome masculino ou feminino dependendo do contexto. Por exemplo, a sentença em Húngaro “ő nővér” (“ele/ela é enfermeiro(a).”) era traduzida como “ela é enfermeira.”, enquanto a sentença “ő orvos.” (“ele/ela é médico(a)”) era traduzida como “ele é médico.”. Esse fenômeno curioso levantou acusações de que o sistema exibe viés de gênero, mas é impossível condená-lo ou inocentá-lo a partir de apenas algumas capturas de tela. Então nos propomos a selecionar um conjunto de idiomas que admitem pronomes de gênero neutro, coletar uma lista volumosa de 1020 profissões e pedir que o Google Translate traduzisse sentenças no formato “ele/ela é um(a) &lt;profissão&gt;” de cada um desses idiomas para o inglês. Isso nos permitiu coletar estatísticas sobre a proporção de vezes que uma sentença era traduzida com pronome feminino, masculino ou neutro (o português não tem pronome neutro, mas o inglês sim: “it”) para cada profissão. Além disso, como as profissões podem ser agrupadas em categorias (por exemplo, profissões da área médica, profissões de ciência e tecnologia, etc.), também pudemos calcular estatísticas dentro de cada categoria.</p><p>A tabela abaixo detalha quais foram os idiomas escolhidos (coluna “Testada”) e porquê. De maneira geral, escolhemos idiomas que suportam pronomes de gênero neutro, com algumas exceções nos idiomas persa, nepalês e coreano, que produziram dificuldades técnicas.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/634/1*SWhGoVbnuJJ1i2g7Qzsw8w.png" /></figure><p>A tabela abaixo detalha quais foram as profissões selecionadas, obtidas do U.S. Bureau of Labor Statistics (Departamento Americano de Estatísticas Laborais) em <a href="https://www.bls.gov/cps/cpsaat11.htm.">https://www.bls.gov/cps/cpsaat11.htm.</a> Esses dados dizem respeito ao mercado de trabalho americano. A última coluna (“Participação Feminina”) mostra a proporção de trabalhadoras do gênero feminino na respectiva categoria.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/692/1*JHkisvwxbpkAUlLLcfpGdA.png" /></figure><p>Na tabela abaixo, um resumo dos resultados obtidos. Para cada categoria de trabalho, podemos comparar a proporção de sentenças traduzidas com pronome feminino, masculino ou neutro. De maneira geral, o sistema traduz usando pronomes masculinos com muito maior frequência. Também Podemos observar que em categorias estereotípicas masculinas, como Ciência, Tecnologia, Engenharia e Matemática, a proporção de traduções femininas diminui enquanto a de traduções masculinas cresce.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/568/1*904Kkwx_5zPHHdnUEjBy4w.png" /></figure><p>Os resultados nos levam a crer que o Google Translate (que é treinado também a partir do feedback dos usuários) incorporou viés de gênero dos dados de treinamento. Existe a possibilidade, ainda, que o Google Translate esteja apenas refletindo a proporção de participação feminina em cada categoria. Por exemplo, talvez o sistema traduza com pronome feminino sentenças sobre vagas na saúde com 23% porque a categoria de saúde é composta (aproximadamente) de 23% de trabalhadoras mulheres. Conduzimos testes estatísticos para verificar essa possibilidade, no entanto, e ela não se sustenta. Em quase todos os casos, a participação feminina numa vaga de trabalho é maior do que a probabilidade do Google Tradutor traduzir uma sentença sobre aquela profissão usando o pronome feminino. Ou seja: aparentemente, o viés de gênero do sistema vai além da assimetria no mercado de trabalho. Possivelmente ele reflete os estereótipos associados às profissões. Por exemplo, ainda que vagas em Ciência, Tecnologia, Engenharia e Matemática sejam estereotipicamente consideradas masculinas (e traduzidas como feminino somente 4% das vezes), a participação feminina é de 31%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/693/0*KxBF4xug5QfyYt1U" /><figcaption>Comparação entre uma captura de tela do Google Tradutor antes (à esquerda) e depois (à direita) da atualização que oferece traduções em ambos os gêneros, para algumas linguagens.</figcaption></figure><p>Nosso estudo foi publicado inicialmente em Setembro de 2018. Desde então, a Google tem tomado providências para mitigar o efeito do viés no sistema, por exemplo oferecendo traduções em ambos os gêneros masculino e feminino. Como pesquisadores, reconhecemos que o viés de gênero presente no Google Tradutor é acidental e não proposital. Além disso, reconhecemos que não é prático selecionar conjuntos de treinamento sem viés para treinar esse tipo de sistema, pois eles provavelmente não existem. No entanto, existem técnicas simples e eficientes que são capazes de remover viés de um sistema treinado. Nosso objetivo é o de que o nosso trabalho possa servir de alerta sobre os perigos que o viés de máquina representa, e de convite para que as empresas que fazem uso aprendizado de máquina se esforcem para remover ativamente o viés dos seus sistemas treinados.</p><p>Da minha perspectiva pessoal, os resultados desse estudo reforçam a necessidade de promovermos a diversidade de gênero (mas não só ela) na computação. Com a escalada da automação, o impacto do aprendizado de máquina na sociedade se torna cada vez maior, mas a identificação dos seus vieses permanece sendo uma atividade essencialmente humana e nossa empatia infelizmente é limitada. Não podemos esperar que uma equipe homogênea seja capaz de se colocar na pele de todas as minorias que compõem a nossa sociedade. Por isso é imperativo que as integremos no desenvolvimento dos softwares que pautam as suas vidas.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1e294d7f4f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Colorindo a Antiga Porto Alegre com #DeOldify]]></title>
            <link>https://medium.com/@marcelodeoliveirarosaprates/colorindo-a-antiga-porto-alegre-com-deoldify-bc8ad095b563?source=rss-31b69a2db687------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc8ad095b563</guid>
            <category><![CDATA[deoldify]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[porto-alegre]]></category>
            <dc:creator><![CDATA[Marcelo De Oliveira Rosa Prates]]></dc:creator>
            <pubDate>Mon, 07 Oct 2019 22:39:10 GMT</pubDate>
            <atom:updated>2019-10-07T22:52:14.712Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bPmIOSqnr6jsoxFJD3_aKQ.png" /><figcaption>Gasômetro em pleno funcionamento, visto do Guaíba em 1969. <a href="http://antigaportoalegre.no.comunidades.net/fotos-1961-1980">Fonte</a></figcaption></figure><p>Colorização de fotografias antigas é um problema importante, inclusive com relevância comercial (vide seriado “Segunda Guerra em Cores”, disponível na Netflix). Entretanto, se efetuada à mão e dependendo do nível de detalhe da fotografia, ela pode acarretar num processo longo e complicado. Esse trabalho pode ser simplificado por algoritmos de coloração que completam o serviço a partir de pequenas “dicas” do usuário, que se responsabiliza por indicar a cor de algumas regiões manualmente. Mesmo que técnicas tradicionais de coloração digital sejam capazes de delimitar automaticamente as regiões que serão coloridas, não é possível fugir da necessidade de recorrer a input humano para identificar qual cor será pintada em cada região. Isso porque esse conhecimento, além de empírico, é muito vasto e portanto não está ao alcance dos algoritmos modernos. Certo?</p><p>Há 5 anos, responderíamos essa pergunta com um estrondoso e definitivo “sim”. Entretanto, o panorama da tecnologia foi transformado durante a última década e permanece sendo continuamente reinventado devido ao advento do <em>Deep Learning</em>. Situado dentro da área de estudo de <em>Machine Learning</em> (aprendizado de máquina), que se encarrega do estudo de algoritmos que aprendem a resolver problemas a partir de exemplos, deep learning se refere ao desenvolvimento destes algoritmos por meio de redes neurais artificiais — modelos matemáticos vagamente inspirados no funcionamento de redes neurais biológicas.</p><p>Essas redes neurais artificiais são “treinadas” a partir de uma vasta quantidade de exemplos, e são capazes de reter o conhecimento empírico que nós, humanos, capturamos tão bem. Elas são definidas por um número gigantesco de parâmetros. Cada combinação de parâmetros produz um resultado potencialmente diferente, e resultados teóricos garantem que, respeitadas algumas condições, existe pelo menos uma combinação de parâmetros capaz de produzir uma rede que soluciona o problema. O treinamento consiste em definir uma função de erro, ou <em>loss</em>, que penaliza os erros do modelo, e calibrar os parâmetros iteradamente até que a loss se torne suficentemente pequena.</p><p>No nosso caso específico, os exemplos correspondem a centenas de milhares de pares de imagens. Cada par é produzido a partir da decomposição de uma imagem colorida qualquer na sua luminância (isto é, a sua versão preto-e-branco) e apenas as suas cores.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dETEPdj-eb9GFXE3vA0dzg.png" /><figcaption>À esquerda, a luminância da imagem. À direita, o mapa de cores (matiz e saturação) produzido pelo algoritmo.</figcaption></figure><p>O objetivo da rede é aprender um mapeamento entre a primeira e a segunda imagem, ou seja: ser capaz de produzir, a partir de uma imagem em preto-e-branco, uma camada adicional de cores a ser sobreposta a ela de modo a produzir uma imagem colorida. Guardadas as devidas proporções, isso corresponde a aprender as pinceladas de tinta que um artista sobreporia à fotografia original num processo artesanal de coloração. A maneira mais direta e intuitiva de treinar uma rede com esse objetivo é usando uma “loss perceptual”, ou seja, uma função de erro que penaliza a diferença entre o valor real e o valor produzido pela rede na cor de cada pixel da imagem. Infelizmente, essa penalização não é suficiente para resolver o problema, e é fácil de entender o porquê. Algumas imagens podem admitir múltiplas respostas (por exemplo, um fusca pode ser pintado tanto de azul quanto de vermelho, desde que seja bem colorizado), mas a loss perceptual enviesa o modelo para uma única. Um modelo treinado com X fotos de fuscas azuis e X fotos de fuscas vermelhos será penalizado ao desviar tanto da cor azul quanto da vermelha, e portanto optará por chutar uma cor no meio do caminho, com tom roxo. Quando usamos loss perceptual, a rede acaba aprendendo a usar esse tipo de estratégia num nível muito maior, abrangendo todas as classes de objetos, e acaba chutando de maneira predominante a cor verde, que por estar no ponto médio do espectro de cores, estatisticamente mantém o erro numa faixa moderada — porém ainda insuficiente para resolver o problema.</p><p>O modelo <a href="https://github.com/jantic/DeOldify">#DeOldify</a>, proposto por Jason Antic, ataca o problema usando treinamento adversário, uma técnica poderosa usada em muitos modelos generativos. Já que a loss perceptual não funciona tão bem, a ideia é propor uma loss de mais alto nível, algo que poderia ser chamado de loss “conceitual”. Ao invés de penalizarmos as diferenças no nível da <em>percepção visual</em> (pixels), vamos penalizar as diferenças no nível da <em>conceitualização</em> da imagem (carros, edifícios, cachorros, etc). Infelizmente não existe nenhuma operação simples capaz de conceitualizar imagens nesse nível abstrato, mas é aí que as redes neurais oferecem ajuda mais uma vez: podemos <em>treinar</em> um segundo modelo para atuar como função de loss. O propósito desse modelo, chamado <em>discriminante</em>, seria separar imagens coloridas autênticas de imagens coloridas por um algoritmo. Se tivermos esse modelo à disposição, podemos usá-lo para treinar o modelo generativo, cujo novo objetivo seria “enganar” o discriminante, ou seja, produzir imagens tão indistinguíveis quanto possível de imagens autênticas. Na prática, ambos os modelos são treinados juntos, com propósitos opostos: o discriminante é treinado para “denunciar” imagens produzidas pelo gerador, enquanto o gerador é treinado para enganar o discriminante. Ao longo do tempo, a competição rende bons resultados, e o gerador aprende a produzir colorações verossímeis. O #DeOldify implementa alguns truques adicionais, treinando ambas as redes de maneira separada antes de dar início ao treinamento adversário, que é bastante custoso computacionalmente pois precisa ser rodado por um número suficientemente alto de iterações para que ambos os modelos atinjam um ponto de equilíbrio.</p><p>A título de experimento, produzi colorações de dezenas de imagens em preto-e-branco da antiga Porto Alegre (décadas 1920–1970) usando o #DeOldify e produzindo resultados muito satisfatórios. Por exemplo, essa fotografia do Arroio Dilúvio ainda limpo e banhável:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/550/1*hge7n9s7BhQFlQ448lKzuQ.png" /><figcaption>Arroio Dilúvio limpo. Fonte: <a href="http://g1.globo.com/rs/rio-grande-do-sul/noticia/2016/03/g1-compara-mudanca-ao-longo-dos-anos-de-cinco-locais-de-porto-alegre.html">http://g1.globo.com/rs/rio-grande-do-sul/noticia/2016/03/g1-compara-mudanca-ao-longo-dos-anos-de-cinco-locais-de-porto-alegre.html</a></figcaption></figure><p>Ou ainda, a Praça da Matriz nas décadas de 1910 e 1920:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NBb3Og-CkzaiJ8Ajew4eLA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*2IbLWVtIJi8brtOUoB8PUg.png" /><figcaption>Praça da Matriz nas décadas de 1910 (esq) e 1920 (dir). Fonte: <a href="http://lealevalerosa.blogspot .com/2010/03/pracas-e-parques-de-porto-alegre.html">http://lealevalerosa.blogspot .com/2010/03/pracas-e-parques-de-porto-alegre.html</a></figcaption></figure><p>Aterro que daria origem aos atuais Bairro Praia de Belas e Parque Marinha do Brasil:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*-Mpg4AslRcfbBf5kLpOOZQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HMZh_kUgNCuBjDeQpvOT_w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*3CYg4Ir_HSHgdbtI1MsEgA.png" /><figcaption>Aterro que daria origem ao bairro Praia de Belas nos anos de 1976 (esq), 1959 (cen) e 1950 (dir). Fontes: <a href="http://lealevalerosa.blogspot.com/2015/11/aterros-sobre-o-guaiba-porto-alegre.html">http://lealevalerosa.blogspot.com/2015/11/aterros-sobre-o-guaiba-porto-alegre.html,</a> <a href="https://prati.com.br/fotosantigas/fotos-antigas-porto-alegre">https://prati.com.br/fotosantigas/fotos-antigas-porto-alegre,</a> <a href="http://www.libretos.com.br/aguas-do-guaiba/index.php/aterro-sul">http://www.libretos.com.br/aguas-do-guaiba/index.php/aterro-sul</a></figcaption></figure><p>Atual Redenção à época das comemorações do Centenário Farroupilha, em 1935:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ag103L6F6l7QcIRHBXi7Zg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*3md3pHt5DQfsIbuGcLkMcQ.png" /><figcaption>Pórtico de entrada da Exposição do Centenário Farroupilha, no então Parque da Redenção (esq) e Lago da Redenção durante a Exposição do Centenário Farroupilha, 1935. Pavilhão do Estado de São Paulo pode ser visto no canto direito (dir). Na época, diversos pavilhões temporários foram erguidos no Parque. Fonte: <a href="http://www.jornalja.com.br/do-campo-da-varzea-ao-parque-farroupilha-redencao-completa-210-anos/">http://www.jornalja.com.br/do-campo-da-varzea-ao-parque-farroupilha-redencao-completa-210-anos/</a></figcaption></figure><p>Ou mesmo a própria Redenção nas décadas de 1950–1960:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*N5XTPQfAwvzI5FSUQa_sVw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8WXe0zXhfDpCMEJmSugkTA.png" /><figcaption>Redenção nas década de 1950 (esq) e 1960 (dir). Fonte: <a href="http://www.jornalja.com.br/do-campo-da-varzea-ao-parque-farroupilha-redencao-completa-210-anos/">http://www.jornalja.com.br/do-campo-da-varzea-ao-parque-farroupilha-redencao-completa-210-anos/</a></figcaption></figure><p>Maria Fumaça e bondinhos no Centro de Porto Alegre:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/833/1*lsM4GDf3LGKW4cvcLoku4g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/837/1*2A82GsGE5eXxKWOCRu9AOw.png" /><figcaption>Bondinho na Praça XV da década de 1950 e Maria Fumaça no Centro de Porto Alegre na década de 1970. Fontes: <a href="http://antigaportoalegre.no.comunidades.net/fotos-1941-1960">http://antigaportoalegre.no.comunidades.net/fotos-1941-1960,</a> <a href="http://antigaportoalegre.no.comunidades.net/fotos-1961-1980">http://antigaportoalegre.no.comunidades.net/fotos-1961-1980</a></figcaption></figure><p>Além de muitas e muitas fotos das enchentes de 1941 e 1967:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*riRFvnpumV0WcLnP2dgmJA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*IC7JwZiYphE1TmdEL7_ALw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/638/1*NmjwmUagaaB102JYZtqtkg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*g1ee1-DhM1bv94jz7UMbog.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/839/1*gRvC3y3SI1TcVb98LJIgJA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/843/1*z649SAcDDeT8C28dx0D8oA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*GK3s_2FNOVgy838TJyvGOw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*wSbBjRFk91FQV-PMxLSnaw.png" /><figcaption>Fotos das enchentes de 1941 e 1967.</figcaption></figure><p>Avenida Farrapos entre as décadas de 1940 e 1950:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/842/1*OWMwO7Tdu7BANahWVJBv1w.png" /><figcaption>Avenida Farrapos entre as décadas de 1940 e 1950 (provavelmente). Fonte: <a href="http://antigaportoalegre.no.comunidades.net/fotos-1941-1960">http://antigaportoalegre.no.comunidades.net/fotos-1941-1960</a></figcaption></figure><p>Ponte de Pedra antes da recanalização do Arroio Dilúvio, que passava por ali:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0Y7xX4cN_F_7jTEt6HWOwA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/620/1*qzuKFHInlF0bhAD4BXVEcA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/670/1*V8YeSLYU9z5DL4yqgxo8kA.png" /><figcaption>Ponte de Pedra em 1955 (esq), em data desconhecida (cen) e 1950 (dir). Fontes: <a href="https://www.skyscrapercity.com/showthread.php?p=161762400">https://www.skyscrapercity.com/showthread.php?p=161762400,</a> <a href="http://g1.globo.com/rs/rio-grande-do-sul/noticia/2016/03/g1-compara-mudanca-ao-longo-dos-anos-de-cinco-locais-de-porto-alegre.html">http://g1.globo.com/rs/rio-grande-do-sul/noticia/2016/03/g1-compara-mudanca-ao-longo-dos-anos-de-cinco-locais-de-porto-alegre.html,</a> <a href="https://prati.com.br/fotosantigas/fotos-antigas-porto-alegre/146?page_number_0=43">https://prati.com.br/fotosantigas/fotos-antigas-porto-alegre/146?page_number_0=43</a></figcaption></figure><p>O arquivo conta também com fotos bastante antigas que mostram uma Porto Alegre quase irreconhecível, como nessa foto da Praça Otávio Rocha num remoto ano de 1930 no qual Zepellins ainda pairavam sobre Porto Alegre:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/517/1*tnDT1f7wyKMa5tGaV08PRA.png" /><figcaption>Praça Otávio Rocha, década de 1930.</figcaption></figure><p>Ou nessas mostrando os Bairros Teresópolis na década de 1920 e o então Bairro Colônia Africana (atual Rio Branco) em data desconhecida perto do início do Século XX:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qU4kn9E5ckur9RmJODZKqw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/799/1*UUjUA7tBMqnt-bFeIAF2aw.png" /><figcaption>Antigo Bairro Colônia Africana (atual Rio Branco) em data desconhecida perto do início do Século XX e Bairro Teresópolis na década de 1920. Fontes: <a href="https://www.sul21.com.br/cidades/2017/04/colonia-africana-e-o-constante-afastamento-da-cultura-negra-em-porto-alegre/">https://www.sul21.com.br/cidades/2017/04/colonia-africana-e-o-constante-afastamento-da-cultura-negra-em-porto-alegre/,</a> <a href="https://www.flickr.com/photos/fotosantigasrs/11013207255">https://www.flickr.com/photos/fotosantigasrs/11013207255</a></figcaption></figure><p>O álbum completo, disponibilizado no Imgur, pode ser acessado em <a href="https://imgur.com/gallery/vHuKjy9">https://imgur.com/gallery/vHuKjy9</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc8ad095b563" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>