<?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 Eder Santana on Medium]]></title>
        <description><![CDATA[Stories by Eder Santana on Medium]]></description>
        <link>https://medium.com/@edersantana?source=rss-f1bbe40e8ca9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*hmjDvVTn4L30IeubSEYuvA@2x.jpeg</url>
            <title>Stories by Eder Santana on Medium</title>
            <link>https://medium.com/@edersantana?source=rss-f1bbe40e8ca9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 21:26:57 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@edersantana/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[A week of Computer Vision at Seoul: My ICCV 2019 Highlights]]></title>
            <link>https://medium.com/@edersantana/a-week-of-computer-vision-at-seoul-my-iccv-2019-highlights-283b5bb8fe47?source=rss-f1bbe40e8ca9------2</link>
            <guid isPermaLink="false">https://medium.com/p/283b5bb8fe47</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Eder Santana]]></dc:creator>
            <pubDate>Thu, 07 Nov 2019 18:24:59 GMT</pubDate>
            <atom:updated>2019-11-08T01:41:01.170Z</atom:updated>
            <content:encoded><![CDATA[<h3>A week of Computer Vision in Seoul: My ICCV 2019 Highlights</h3><p><em>Note: Thanks to everyone that commented and suggested changes to the original </em><a href="https://docs.google.com/document/d/1X0g41PppRD2c5bkZAz61Pvl6DvkbP1icnAILKBzFlTU/edit?usp=sharing"><em>draft</em></a><em> of this post. Highlights are biased towards my current interests in information retrieval in large scale live videos.</em></p><p>I think <a href="http://iccv2019.thecvf.com/">ICCV2019</a> was one of the best conferences I’ve been so far. Some of my friends even said they liked it better than this year’s CVPR. It was pretty nice to see all the progress in computer vision. Also, Seoul is a wonderful city that I’d definitely visit again!</p><p>If I had to mention the top 3 most common topics I saw, I’d say compression/enhancement, neural architecture search and people tracking related work (person/face re-identification, crowd counting, etc). Honorable mention to GANs who are still pretty popular.</p><p>On the other hand, although I was pleased to see all the progress in video processing this year, I failed to learn about concerns of video processing at scale. Perhaps that’s left for future work. That said, here’s some of my highlights.</p><h4><strong>Bridging the Sim-to-Real gap in Computer Vision benchmarks</strong></h4><p>Karpathy’s take on the differences between computer vision research in academia vs industry and how to bridge the gap.</p><h3>Andrej Karpathy on Twitter</h3><p>We see more significant improvements from training data distribution search (data splits + oversampling factor ratios) than neural architecture search. The latter is so overrated :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7393gAlYvSWUoMT14Zf80Q.jpeg" /><figcaption>Karpathy presented Tesla’s Data Engine: how they organize their effort for progress in Machine Learning research for autonomous vehicles.</figcaption></figure><p>Coming from academia myself and now mostly focusing my research on real world problems, I found this one of my favorite talks. Also I like his suggested new kind of benchmark for industry:</p><ol><li>The val / test set is alive and carefully curated all the time based on QA.</li><li>It takes the form of “unit tests” instead of metrics</li><li>The org is split into two groups locked in a GAN-like battle: Group 1 tries to find images that break the network and drive down the % pass rate. Group 2 tries to improve the % pass rate.</li><li>Model uncertainty is a “first class citizen”</li><li>Instead of expected loss we care about maximum loss</li></ol><blockquote>Why are the COCO models allowed to improve alone? Introduce a second leaderboard for images that break SOTA models.</blockquote><p>Here’s another great quote on the topic of iterating in the training and test data instead of focusing solely in the models:</p><blockquote>It’s funny that people would consider this cheating in academia, because that’s what I do for my job everyday in industry.</blockquote><h4><strong>Re-localization part of Torsten Sattler talk at the Visual Localization Tutorial.</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v2cCn7aK6Mz36EWE8Ffx6Q.jpeg" /><figcaption>Image retrieval (the same tech we use for image search) works pretty well for localization in maps.</figcaption></figure><ul><li>(SIFT) feature-based methods still competitive</li><li>State-of-the-art on medium and large-scale datasets</li><li>Key idea of prioritization: Few unique / characteristic features are sufficient</li><li>Visibility filtering: Exploit co-visibility information from reconstruction stage</li></ul><h4>Action Recognition part of Jitendra Malik’s talk</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JT6lcf29Rb9zMjImSkq4Rg.jpeg" /></figure><p>Prof Malik’s talk made me curious to read about their <a href="https://arxiv.org/abs/1812.03982">SlowFast networks</a> in more detail. Here’s the part where he shows SlowFast nets can learn optical flow like features without expensive optical flow estimation</p><p>Our oral presentation</p><p><strong>Posters</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BhprpV02ODi5ydA0" /></figure><p><a href="https://arxiv.org/abs/1904.02750"><em>DecptionNet: Network-Driven Domain Randomization </em></a>was one of my overall favorites. Mostly for the concept of automatic data augmentation for training. I talked to Wadim and Sergey about possible ways to extend this approach to other domains and how one can automatically discover which invariances are relevant for the DeceptionNet to capture. Right now it relies on domain expertise.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SSi1LGsvmeqc6Yl8" /></figure><p>This is one of my favorite representatives of the video compression papers I saw. <a href="https://arxiv.org/abs/1908.05717"><em>Video Compression with Rate-Distortion Autoencoders</em></a><em>:</em> work on the entire video volume with 3D convolutions instead of the “find keyframes, encode diffs” approach. I think they still have the same issues with scalability of the other neural networks video compression approaches when we’re talking about YouTube/Twitch video scale…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xnRc7sn0mmHNU9J_" /></figure><p>I talked to my fellow Brazilian researcher Cézar de Souza, from Naver Labs Europe, about their paper on <a href="https://arxiv.org/abs/1906.07589"><em>Learning with Average Precision: Training Image Retrieval with a Listwise Loss</em></a>. And I think this is going to be one of the most useful papers from ICCV to me. In information retrieval, Average Precision is one of the main metrics for <a href="https://en.wikipedia.org/wiki/Learning_to_rank">Learning to Rank</a> performance. So, it’s obvious why I’d be interested in a paper that proposes a differentiable approximation of Average Precision as a Listwise loss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0wozVIax_JZg0yxI8bBtgA.png" /></figure><p>Also there was our oral presentation on<strong> </strong><a href="https://arxiv.org/abs/1904.08980">Exploring the Limitations of Behavior Cloning for Autonomous Driving</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=283b5bb8fe47" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ganhando experiência em uma nova área de trabalho sem pedir permissão pra ninguém]]></title>
            <link>https://medium.com/@edersantana/ganhando-experi%C3%AAncia-em-uma-nova-%C3%A1rea-de-trabalho-sem-pedir-permiss%C3%A3o-pra-ningu%C3%A9m-a57fdfaae8ae?source=rss-f1bbe40e8ca9------2</link>
            <guid isPermaLink="false">https://medium.com/p/a57fdfaae8ae</guid>
            <category><![CDATA[motivação]]></category>
            <category><![CDATA[aprendizagem-autodirigida]]></category>
            <category><![CDATA[trabalho]]></category>
            <dc:creator><![CDATA[Eder Santana]]></dc:creator>
            <pubDate>Sun, 14 Oct 2018 23:07:06 GMT</pubDate>
            <atom:updated>2020-06-25T23:01:18.287Z</atom:updated>
            <content:encoded><![CDATA[<p>Quem nunca tentou começar uma nova carreira em um ramo da Ciência da Computação cujas posições de trabalho requerem experiência? Mesmo nas posições Junior, as expectativas podem ser altas. Eu observei isso nas áreas de pesquisa e desenvolvimento de machine learning e data science. Inclusive ouvi pessoas me perguntando como lidar com essa questão. Neste texto, eu organizo algumas das conversas que costumo ter sobre o assunto e mostro o exemplo de alguns amigos que não esperam a permissão de ninguém quando querem ganhar experiência.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*Vb4BnB_9RqLctOdur-8Mww.png" /><figcaption>source: <a href="https://unsplash.com/@pcalescu">pcalescu</a></figcaption></figure><p>David Ha uma vez escreveu online que se deparou com varias pessoas expressarem o desejo de trabalhar com pesquisa em machine learning (ou aprendizagem de máquina), mas rapidamente desistirem da ideia por não trabalharem no Google Brain ou outra empresa com foco em pesquisa. A sua resposta foi que as pessoas não precisam da permissão de ninguém para fazer pesquisa. David não fala tal coisa da boca pra fora, uma vez ele me confessou que começou a estudar machine learning enquanto tinha um trabalho muito exigente na Goldman Sachs no Japão. Ele tinha apenas 3 ou 4 horas pra aprender a programar, ler artigos, testar ideias e se aprimorar no que gosta. Tais limitações não davam muita margem a perda de tempo. Depois de um longo tempo fazendo pesquisa independente sob o pseudônimo de <a href="https://twitter.com/hardmaru">hardmaru</a> e publicando seus achados em seu blog, David de fato conseguiu uma oportunidade como residente no Google Brain.</p><p>O caso do David me soou muito interessante e eu observei em várias pessoas uma pergunta recorrente:</p><blockquote><em>Como fazer aquilo que queremos quando não temos a experiencia exigida nas ofertas de emprego?</em></blockquote><p>Eu acredito ter uma resposta para a pergunta:</p><blockquote>Enquanto ser bem recompensado exige um trabalho que valorize seus esforços, você não precisa de permissão para a ganhar a experiência necessária, fazer parte de uma nova área e conseguir a posição desejada.</blockquote><p>Minha resposta parte de um princípio simples, quando somos bons o suficientes naquilo que fazemos, não podemos ser ignorados! Este foi o conselho que que motivou <a href="https://twitter.com/FarzaTV">Farza</a> a parar de reclamar sobre a injustiça das entrevistas de emprego da Riot Games e outras empresas, e começar a investir tempo em projetos de código livre como seu famoso mood.gg e DeepLeague. Em certo momento, Farza se tornou reconhecido suficiente e os funcionários das empresas que ele queria ingressar ja usavam os aplicativos que ele desenvolveu.</p><p>Outros dois exemplos interessantes são os do <a href="https://twitter.com/fchollet">François Chollet</a> e <a href="https://en.wikipedia.org/wiki/George_Hotz">George Hotz</a>. François largou o doutorado no Japão e se mudou pros Estados Unidos para trabalhar em Startups. François começou a escrever uma biblioteca de código livre para Deep Learning chamada <a href="https://github.com/keras-team/keras">Keras</a> para ajudar em alguns projetos pessoais. Keras foi feita com tanto esmero que acabou se tornando a mais popular da área, ajudou François a conseguir um emprego na Google e hoje é parte fundamental do Tensorflow. Eu também acabei contribuindo para o código do Keras no início e isso me ajudou a ganhar uma visibilidade na industria.</p><p>O ultimo exemplo é o do George Hotz, que me mostrou a importância dos “projetos de final de semana”. George sempre que tem curiosidade sobre algum assunto dedica alguns finais de semanas para um projeto prático e permite que outras pessoas da sua startup façam o mesmo usando recursos da empresa. Por vezes George vai longe o suficiente a ponto de fazer transmissão ao vivo enquanto aprende e programa algo novo.</p><p>Concluindo os exemplos acima, podemos resumir nossa resposta para o ganho de experiência em 3 passos práticos:</p><ol><li><strong>Escolha uma área de interesse e colete material de estudo online.</strong> Como o David Ha na anedota acima arrume alguma forma de adquirir os conhecimentos básicos da área que quer ingressar. Sejam livros, artigos online, videos no YouTube, qualquer coisa. Separe algumas horinhas durante a semana e se dê algum tempo para aprender os princípios. A ideia não é ler tudo sobre o assunto, mas ganhar uma noção de onde está se metendo, o suficiente para agir no passo 2.</li><li><strong>Engage em projetos práticos da área de interesse.</strong> Com alguma noção das capacidades da nova área de interesse, fica mais fácil saber o que é possível no novo ramo. A história do Farza e George são bons exemplo aqui. Escolha projetos e problemas que tenham potencial educacional, mesmo que sejam problemas já resolvidos. O principal objetivo é se deixar levar pelo problema e acostumar o pensamento como alguém da área. Este projeto pode ser a implementação de um novo app para uma comunidade específica, uma competição no Kaggle, contribuição a uma biblioteca de código livre, etc. A meta neste ponto é construir um portfolio onde o contratante e contratado são você mesmo.</li><li><strong>Continue se aprimorando e ajude as pessoas a te acharem.</strong> No passo 3, o foco é usar a experiência do passo 2 e criar o tipo de material que você consumiu no passo 1. Assim se fecha o ciclo. Usando o exemplo dos colegas acima, neste passo você vai construir o tipo de blog, vídeos ou código livre que você poderia ser útil para as outras pessoas e você mesmo.</li></ol><p>No texto acima nós observamos o exemplo de algumas pessoas que não esperaram a permissão ou a oportunidade de trabalho perfeita para se treinar em uma nova área de interesse. Depois nos organizamos os exemplos em 3 passos para que todos possam repetir os exemplos. O interessante desta abordagem é que ela pode ter um efeito colateral muito positivo e interessante. Enquanto que a falta de experiência formal pode parecer uma desvantagem, quando desenvolvemos experiência de maneira independente ganhamos o mérito de auto-didata e provamos uma capacidade de iniciativa e independência, talvez a experiencia mais importante de todas. Por fim, vale notar que nenhum dos meus amigos que citei acima fizeram um doutorado (George nem a faculdade terminou), mas todos eles demonstraram trabalhos de mais impacto que vários dos que vi em programas de ensino formais ou em empresas de prestígio.</p><p>Eu adoraria ouvir sobre a jornada de crescimento de outros Brasileiros, principalmente nas áreas de machine learning, data science e inteligência artificial. Contem aqui nos comentários ou no Twitter sobre suas jornadas de aprendizado auto-ditada. Eu mesmo estou numa jornada parecida e escrevo aqui pra ganhar mais experiencia e contato com outros profissionais do meu país. Até a próxima!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a57fdfaae8ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MVE Series: Playing Catch with Keras and an Evolution Strategy]]></title>
            <link>https://medium.com/@edersantana/mve-series-playing-catch-with-keras-and-an-evolution-strategy-a005b75d0505?source=rss-f1bbe40e8ca9------2</link>
            <guid isPermaLink="false">https://medium.com/p/a005b75d0505</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[keras]]></category>
            <category><![CDATA[evolutionary-algorithms]]></category>
            <dc:creator><![CDATA[Eder Santana]]></dc:creator>
            <pubDate>Sun, 28 Jan 2018 04:04:12 GMT</pubDate>
            <atom:updated>2018-01-28T04:18:40.704Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*NjcSrsT-dSyEjweCVwwXAg.gif" /><figcaption>Keras model evolved to play catch.</figcaption></figure><p>Whenever I want to learn a new machine learning method, I write a single file example with a Minimum Viable Experiment (MVE). Last time I did an <a href="https://gist.github.com/EderSantana/c7222daa328f0e885093">MVE to learn RL and Q-Learning</a>. This time I want to learn <a href="https://arxiv.org/abs/1703.03864">Evolution Strategies</a>.</p><p>Before I show you the code I came up with, let us talk a bit about a fun recurring equation. Say you want to create a model of a variable <strong>y</strong> (ex. price of a stock, say if an image shows a cat or a dog, the steering of a self driving car, etc). You go ahead and use a model <strong>f(x)</strong> to approximate <strong>y</strong>. Where <strong>x</strong> are inputs (ex. previous stock values, a jpeg image, a video of the road, etc). That model <strong>f</strong> also have parameters <strong>w</strong> to be tuned. Just like in a neural network that at this point everybody heard about.</p><p>A simple way to measure how well your model performs is to calculate the average value of <strong>L = (y-f(x))^2</strong>. Then, you can update the parameters <strong>w</strong> using information about the gradient of L: <strong>grad(L) = -2(y-f(x))*grad(f(x))</strong>.</p><p>Note that there are two main terms multiplied in the last equation. A quality measure <strong>-2(y-f(x))</strong> and a model sensitivity <strong>grad(f(x))</strong>. If we forget precise formulations for a bit and just think about matching possible quality measures and model sensitivity, we can see that this same equation is used elsewhere, especially in RL. For example, the REINFORCE algorithm also updates model parameters according to <strong>R*grad(log(p(x))</strong> where <strong>R</strong> is a reward measure and <strong>log(p(x))</strong> is a log-probability of certain actions of a policy function.</p><p>Taking this to an extreme, where we assume that a model sensitivity is solely defined by small random jitters <strong>n</strong> we add to our model parameters <strong>w</strong>, we get <strong>grad(L) = R * n.</strong> Where again <strong>R</strong> are rewards defined by completing a given task. Obviously that approximation is not really precise and should have a lot of errors when estimating good values to update <strong>w</strong>. Good thing about it is we can usually compensate for lack of precision with exhaustive trial, error and lots of data.</p><p>And that is it. We can now try to solve the same catch game from the previous post, but this time with a much simpler formulation. The trade off is more compute. Now you can <a href="https://github.com/EderSantana/EvolutionMVE">check out the code</a>. It is pretty short and should be simple to understand, but let me know if something isn’t clear.</p><p><em>PS: After you are done understanding the basics of evolution strategies, I’d recommend using more serious code bases like </em><a href="https://github.com/openai/evolution-strategies-starter"><em>OpenAI ES</em></a><em>, </em><a href="https://github.com/hardmaru/estool"><em>estools</em></a><em> or </em><a href="https://github.com/alirezamika/evostra"><em>evostra</em></a><em> for further experiments.</em></p><p>PS1: Check out<a href="https://twitter.com/hardmaru"> hardmaru</a>’s <a href="http://blog.otoro.net/2017/10/29/visual-evolution-strategies/">A Visual Guide to Evolution Strategies</a> for more evolution strategies and nice visualizations.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a005b75d0505" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How do self driving cars drive? Part 1: Lane keeping assist]]></title>
            <link>https://medium.com/@edersantana/how-do-self-driving-cars-drive-part-1-lane-keeping-assist-581f6ff50349?source=rss-f1bbe40e8ca9------2</link>
            <guid isPermaLink="false">https://medium.com/p/581f6ff50349</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[control]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Eder Santana]]></dc:creator>
            <pubDate>Sun, 04 Jun 2017 17:30:37 GMT</pubDate>
            <atom:updated>2017-06-06T06:10:44.059Z</atom:updated>
            <content:encoded><![CDATA[<p>You’re sold on self driving cars (SDC), right? You’ve read so much about the future of mobility and how disruptive self driving cars can be. You even keep coming back to this blog… Now, let’s try to get some first principles about how they work and hopefully see what we can do to contribute to self driving progress.</p><p>Putting it simply, self driving car is about getting a 4k pounds, 4 wheeled robot to follow a path. That path can be calculated globally, say <a href="https://voyage.auto/">go from A to B like a taxi</a> and/or locally, <a href="http://comma.ai/">going forward in a highway without departing the lane</a>. In the process, the car should also adapt its speed to not hit objects and plan its path to avoid obstacles.</p><p>Now, let us assume the simplest case possible. We have a car in a highway without any other vehicles around and we want it to go forward keeping itself in the center of the lane. We will discuss more general cases in other posts or something… For now, see Figure 1.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*hkG4vVmAQ8ItmAQuJIPhwQ.png" /><figcaption>Figure 1. Highway lane keeping assist. We want the car to drive itself in the center of the lane.</figcaption></figure><p>If you ask me, a human, how I’d solve the problem of keeping a car in the center of a lane. I’d say I would verify how far I’m from left and right lanes. If I’m too close to the left lane, I’d steer the vehicle to the right and vise-versa. I’d use my eyes and spatial intelligence to estimate how far each lane is. Same thing for speed. If the car is too slow I’d press the gas, if it too fast I’d press the brakes. How much force or torque I’d place on the pedals and the steering wheel was heuristically learned since I was 18, back in Brazil. Let us forget about manual gears for now, since millennials don’t know about it anyways…</p><p>The intuition above involved actuators (steering wheel and pedals), state measurement (how far I am from each lane and my current speed) and the knowledge of a desired state (keep myself in the middle of the lane as best as I can while following the speed limits). All that is represented in the block diagram of Figure 2.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/723/1*L2NjVcHmoFWu_VueQpF-kA.png" /><figcaption>Figure 2. Block diagram of the closed loop controls for self driving cars.</figcaption></figure><p>Assume next that actuation is granted because we hacked our cars to steer it programmatically. The problem reduces to measure where the car is in the lane and how to control it to the desired state.</p><h3>Sensors: getting the car to find where it is</h3><p>There are several ways to calculate how far the car is from each lane. If we want to use computer vision, we can place cameras pointing forward or sideways to see the road. Classic computer vision uses <a href="https://medium.com/udacity/udacity-self-driving-car-nanodegree-project-1-finding-lane-lines-719ac1adbed9">edge detectors</a> to estimate the lanes. We can also use deep learning to find the lanes — to train that we will need ground truth positions that comes either from hand labelling (not scalable) or from other sensors like LiDARs (better). LiDARs themselves are pretty good for finding lane lines because they reflect light differently from asphalt (surprise!).</p><p>Once we segment out the lanes in the image, we can use geometry to calculate their distance to the car. Assuming the cameras are stable in the car, we can figure a pixel to inches transformation. I may write later about computer vision and geometry for SDC, but if you are really curious already, check out <a href="https://www.amazon.com/Multiple-View-Geometry-Computer-Vision-ebook/dp/B00AKE1QK4/ref=sr_1_1?ie=UTF8&amp;qid=1496594497&amp;sr=8-1&amp;keywords=multiple+view+geometry+for+computer+vision">this book</a>. In a <a href="http://www.cv-foundation.org/openaccess/content_cvpr_2016_workshops/w3/papers/Gurghian_DeepLanes_End-To-End_Lane_CVPR_2016_paper.pdf">paper by Ford</a>, they used sideway cameras and calculated the position of the lanes using deep learning and hand labeled images.</p><p>Speed measurement is a bit easier to estimate and already provided by the car internals.</p><h3>Controls: actually self driving</h3><p>Once we know where car is in the road and how fast it is going, we have to control it to the desired position and speed. Another simplification is to assume we can divide the problem in two: Longitudinal (how much throttle and brake) and Lateral (how much torque on the steering wheels) controls. That break down will not always work. But we may be able to do that in a highway, where curves are gentle.</p><p>The simplest controls method is the Proportional Integral Derivative (PID) controller. Note that there are several other ways to do controls, but we will stick to this one for simplicity. To understand how PIDs work save me some words and watch the (very good) video below:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F4Y7zG48uHRo%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D4Y7zG48uHRo&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F4Y7zG48uHRo%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/65fa368d76624b9312a0547f7f5fac7d/href">https://medium.com/media/65fa368d76624b9312a0547f7f5fac7d/href</a></iframe><p>If you watched the video you know that there are 3 parameters to be defined for each controller. In a perfect world, we would know exactly how each actuation command moves the car. In other words, we would know the function that gets steering wheel angles and speed as inputs and outputs the car movement. In real life, we don’t know that function and modelling car dynamics is a kinda hard. See <a href="https://www.amazon.com/Science-Vehicle-Dynamics-Handling-Braking/dp/9401785325/ref=sr_1_1?ie=UTF8&amp;qid=1496594695&amp;sr=8-1&amp;keywords=the+science+of+vehicle+dynamics">this book</a> if you believe that <em>Physics or it didn’t happen</em>.</p><p>We need that vehicle dynamics function to define the best PID. Since that model of the car might not be perfect, in practice engineers usually have to do hand tuning of the PID until the car drives well enough. If the PID or whatever controls strategy was used to steer the car was not well tuned, we may get a system that wobbles the car between lanes.</p><p>Machine learning can be used for estimating a good vehicle model. When we have a good vehicle model, there exist other controls methods that may work better, like <a href="https://en.wikipedia.org/wiki/Linear%E2%80%93quadratic_regulator">LQR</a>, <a href="https://en.wikipedia.org/wiki/Model_predictive_control">model-predictive control</a>, etc. Check out what the <a href="http://heli.stanford.edu/">Stanford helicopter project</a> did for instance.</p><h4>A note about end-to-end self driving cars</h4><p>In the context of SDCs end-to-end driving refers to using deep learning to get a camera feed as input and output the driving decisions a human would take to drive the car. That doesn’t work well yet. Yes, we can learn to regress images to steering wheel angles using deep learning. That doesn’t mean, for instance, that the deep neural network learned the physics of the problem. This is an innovative strategy and kinda works for defining steering wheel commands that goes to the PID loops as the desired path instead of <em>center of the lane</em>. This is particularly useful when there are no lanes to be detected. But neural nets generalize only so well as the test data follows a similar distribution to the training data. For example, neural nets have a hard time doing position recovery once the car start departing the lane. Think about it, we don’t usually collect self driving car data falling asleep and waking up to recover the right position, neither do we drive recklessly to get the car in a weird positions in public roads for training purposes.</p><p>I think that learning rules of physics from video is one of the most important lines of Machine Learning research right now. But there is only so much we can learn from a static dataset as used to train end-to-end self driving cars. <a href="https://medium.com/@edersantana/towards-a-scientific-method-for-self-driving-cars-research-4db960833677">I argued before</a> that something like <a href="https://blog.openai.com/roboschool/">OpenAI Roboschool</a> would be a much smarter and principled approach.</p><h3>Open problems</h3><p>If you were paying attention until now, you saw how many simplifications we had to make: highway only, separate Longitudinal and Lateral controls, imperfect vehicle dynamics model, simple PID controls, ground truths to find lanes in the road, etc.</p><p>Out of those problems, the ones regarded as the hardest are the perception problems, i.e. finding the lanes and drivable paths. That is fair to say since engineers know controls well enough for quite sometime now. If machine learning/computer vision people can come up with good vehicle state and surroundings estimation, controls people are competent enough to know exactly what to do next. This helps to understand why most of the startups in SDC pitch themselves as localization or perception companies.</p><p>That said, the way we discussed SDC in this post will only help solve the simplified highway problem. The full story will require dynamic path planning, handling other vehicles and people around in the streets. But let’s leave further discussions to future posts.</p><h4>Acknowledgments</h4><p>Thanks to <a href="http://sam.khalandovsky.com/">Sam Khalandovsky</a> for suggestions to the draft.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=581f6ff50349" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Towards a Scientific Method for Self-Driving Cars Research]]></title>
            <link>https://medium.com/@edersantana/towards-a-scientific-method-for-self-driving-cars-research-4db960833677?source=rss-f1bbe40e8ca9------2</link>
            <guid isPermaLink="false">https://medium.com/p/4db960833677</guid>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[scientific-method]]></category>
            <dc:creator><![CDATA[Eder Santana]]></dc:creator>
            <pubDate>Sat, 20 May 2017 18:04:28 GMT</pubDate>
            <atom:updated>2017-05-22T20:50:44.915Z</atom:updated>
            <content:encoded><![CDATA[<p>We need well designed methods for progress in both Academia and industry. For example, the recent Deep Learning trend started in Academia as an approach to improve Machine Learning performance in well known tasks. After years of Academic validation, Deep Learning graduated to a systems problem and is widely adopted and developed in industry. Even after that stage, there is still huge efforts to make Deep Learning research and development even more <a href="http://ai-on.org/">democratic</a> and <a href="https://keras.io/">accessible</a>.</p><p>On the other hand, Machine Learning <a href="https://arxiv.org/abs/1604.07316">as</a> <a href="https://arxiv.org/abs/1608.01230">used</a> <a href="https://arxiv.org/abs/1704.07911">for</a> self-driving cars is a piece in an incredibly complex system with many moving parts. We propose breaking that problem into pieces to make it easier to measure progress.</p><h3>Scientific Method</h3><p>A simplified version of the Scientific Method can be presented in 6 steps:</p><ol><li>Define a question</li><li>Gather information</li><li>Form hypothesis</li><li>Test hypothesis with <strong>reproducible</strong> experiments.</li><li>Analyze results</li><li>Repeat!</li></ol><p>Current self-driving car research, when compared to the general scientific method, requires further clarity. First, the question made in Step 1 “how to autonomously drive a car from A to B” is too broad and Academic unfriendly. Equipping a vehicle with the necessary hardware components, enabling algorithmic sensing and control, is difficult and cost prohibitive. Cost is also a major factor in information gathering (Step 2).</p><p>Most critical is the lack of well defined standards for Step 4. Without such standards progress is difficult to measure, generates confusion, and causes over defensive laws.</p><h3>Proposal</h3><p>The contribution we make here is a set of improved driving tasks. The tasks do not focus explicitly on image segmentation, odometry estimation, road sign classification, etc, as they are already well maintained by the Robotics, Machine Learning, and Computer Vision communities. See for example, the <a href="http://www.cvlibs.net/datasets/kitti/">Kitti</a> and <a href="https://www.cityscapes-dataset.com/">Cityscapes</a> datasets.</p><p>Towards this goal we discuss a standardized <em>driver’s school for self-driving cars</em>. Both in real and simulated roads, and breaking down the problem of self-driving in several smaller tasks. The specific implementations of such driver’s school can follow the approach developed by OpenAI in their <a href="https://gym.openai.com/">Gym</a> and <a href="https://blog.openai.com/roboschool/">Roboschool</a> environments. In OpenAI’s approach simulated environments and public performance leaderboards are used to benchmark the performance of controls algorithms, mostly for Reinforcement Learning research.</p><p>Real world environments can be track circuits, such as the annual <a href="http://selfracingcars.com/">selfracingcars.com</a>, and public streets or road segments where the same number of miles are driven in different days and hours of the day. Standard software should be required to measure performance and ensure public credibility.</p><p>Below, we list smaller driving tasks that can be used to benchmark self-driving car improvement. We believe those tasks are straightforward enough to be tackled in both Academic and business settings.</p><h4>Lane maneuvers</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tLr40EqRVTlukXbrFVm57Q.png" /><figcaption>comma one lane keeping assistant (source: electrek.co)</figcaption></figure><p>Lane keeping involves getting the vehicle centered within a selected lane. The problem can be broken down into lane perception and path prediction. Which can be easily tackled with physics models of cars and large scale lane annotated datasets.</p><p>The problem of lane keeping should also involve comfort measures to avoid wobbling between lanes and safety to perform curves at proper speed.</p><p>Other lane maneuvers are lane changing, overtaking, exit taking, etc.</p><h4>Road event detection and prediction</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2BePFVi876pOyPR6Amamrg.jpeg" /><figcaption>Tesla Autopilot Demonstration (source: <a href="https://www.tesla.com/videos/autopilot-self-driving-hardware-neighborhood-long">https://www.tesla.com</a>)</figcaption></figure><p>Detecting and predicting events on the road can be treated as classification problems. Here we need a large scale and public dataset of driving footage where the relevant events are labeled. Many relevant events exist but for the sake of brevity we list the most interesting:</p><ol><li>Lane departure</li><li>Collision prediction</li><li>Cut-in and cut-out prediction</li><li>Pedestrian action recognition</li></ol><p>Essentially, a driving version of the ImageNet dataset and the competition around it are needed.</p><h4>Massively Multiplayer and Multimanufacturer Online Simulator (MMMOS)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*b3Z2_HE80-pfhDznJcXm4Q.jpeg" /><figcaption>GTA V taxi mission (source: guides.gamepressure.com)</figcaption></figure><p>We shouldn’t expect that all self-driving cars on the roads run the exactly same software. But we should require that they can operate well when driving with humans and different self-driving algorithms.</p><p>A desired solution for this problem should be a separate party that maintains an online environment where both open ended worlds (such as World of Warcraft and GTA V) and task specific worlds (ex: GTA V taxi missions or League of Legends) are inhabited by human players and AI from different manufacturers.</p><h3>Conclusions</h3><p>Measuring progress within any field is diffcult but can be made easier through discussion of well defined tasks. We hope that our proposal, which we make no claims at being extensive, provides a good starting point towards measurable self-driving research progress. Furthermore, we believe the discussed tasks can bring value to both Academic and Industrial settings. The value provided here is two fold: progress between industry and acaedmia is shared, measured, and open for discussion with the public.</p><h4>Acknowledgement</h4><p>Thanks to <a href="https://medium.com/u/d4975234c148">Norm</a> for carefully reviewing and improving the draft.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4db960833677" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[So your company wants to do AI?]]></title>
            <link>https://medium.com/@edersantana/so-your-company-wants-to-do-ai-824b0fbf8b1a?source=rss-f1bbe40e8ca9------2</link>
            <guid isPermaLink="false">https://medium.com/p/824b0fbf8b1a</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Eder Santana]]></dc:creator>
            <pubDate>Sun, 19 Mar 2017 16:50:18 GMT</pubDate>
            <atom:updated>2017-04-25T02:30:43.599Z</atom:updated>
            <content:encoded><![CDATA[<p>Machine Learning, and more so Deep Learning, is so popular now that it is being referred as AI itself. Gladly, your startup just got funded or your new team budget was approved! Now you will be doing Deep Learning as well. You already had fun with Keras, Imagenet, etc. That is truly exciting! Now here are a few things to consider when getting started for real in your business. I’ll illustrate my suggestions with some anecdotes from my time working on <a href="https://electrek.co/2016/08/03/comma-ai-self-driving-car-data-tesla-google/">self-driving cars</a> at<a href="http://comma.ai/"> comma.ai</a> with <a href="https://en.wikipedia.org/wiki/George_Hotz">George Hotz</a> early last year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/638/1*rNBmxYyMRihMgGMerc0v-A.gif" /><figcaption>Help, AI! Where should I go?</figcaption></figure><h4>1) Don’t let data leave your engineers hanging</h4><p>Deep Learning is a data first science. The whole reason for your team or startup to exist is to make sense of a dataset. Think about it, you can only ship that sweet AI Bitcoin Chatbot if you first make sense of text! You can only automate multimedia collage for your next Snapchat Stories clone if you understand images, video, and so on.</p><p>You should not consider data handling just a side aspect of your work. Don’t do a poor job at it. For example, if it takes you “just 15 minutes” to prepare and load a dataset, those are 15 minutes you have wait every single time you find a better model decision or spot a bug in you Tensorflow code.</p><p>The rules here are simple. Version your dataset and preprocess everything once and use it several times. Tools like <a href="http://www.celeryproject.org/">Celery</a>, <a href="https://github.com/spotify/luigi">Luigi</a> and others are your friends. If you are in a large team where tasks need to be submitted to a cluster, consider a data server solution that feeds batches to the model training workers. Please, please! Do not let your team members have to wait for the whole dataset to be downloaded to a worker before they even know there is a bug in their model.</p><p><strong>Fun Story</strong>. comma.ai probably has the second or third largest driving dataset out there. In the early days of comma.ai, to train the model that drives the car, several hours of video were loaded to a big beefy machine with &gt;700 Gb of memory. Everytime he needed to train on more data, George would Prime Now another 100gb of RAM. I was hired to work on better versions of that model, but I didn’t want to have to wait 15 minutes to load up the data. Instead, I got pieces from open source projects for a<a href="https://github.com/commaai/research/blob/master/server.py"> simple ZMQ server</a>. There was no data letting us hanging anymore, we could scale to larger training sets and cheaper machines. The model training was bound only by the GPU and its researcher.</p><h4>2) Start with something you can visualize</h4><p><strong>Fun Story. </strong>In my exit interview I asked George for advice on how to be more productive as an engineer (trust me, he’s the most productive I’ve ever met, I took every single opportunity to learn the heck out of him). He suggested to always start by building things that would help me visualize what I’m doing. <a href="https://www.youtube.com/watch?v=eGl6kpSajag">George had followed this advice himself previously</a>. Also, all of George’s IPython Notebooks had sliding widgets to quickly visualize how parameters affect results while prototyping.</p><p>Luckily for Deep Learning we have<a href="https://www.tensorflow.org/get_started/sum%E2%80%A6"> Tensorboard</a>, the recently introduced <a href="https://github.com/facebookresearch/visdom">Visdom</a>, and some other tools to help us visualize results. I believe Data Science in general is best served with visualization driven development. Consider developing a good rapport with your problem at every step of the way. You don’t necessarily have to learn d3.js to get useful visualizations. Unless you’re one of those with javascript fetish.</p><h4>3) Define your validation/hard cases dataset early on</h4><p>I put the fun visualization stuff as number 2 to give you a break after freaking you out about preparing data. But if you want to avoid being just a monkey on the typewriter, randomly adding more layers to your neural networks, you have to define how to measure progress. Ask what metrics correlate well with better deliverables and which data you should be tracking. This might go beyond the simple “random 10% of the data left out for validation”. A validation dataset optimally has the same statistical properties as production. That same production can be tracked for hard, edge and failure cases that will make future validation sets. Therefore, your validation set may be evolving and should version it just like your training set.</p><p><strong>Fun Story</strong>. I learned that for the hard cases and validation set of a self-driving car you can consider those moments you had to take control back while driving. But there is no better validation test than sending an experienced controls engineer to the road to minutely judge the quality of your self-driving system. If you’re in that business,<a href="http://www.recode.net/2017/1/26/14401452/tesla-lawsuit-chris-urmson-sterling-anderson-autopilot"> try</a> <a href="http://www.theverge.com/2016/4/18/11454858/george-hotz-self-driving-car-tesla-employee-poach">poaching</a> <a href="https://electrek.co/2016/04/19/exclusive-apple-tesla-vp-chris-porritt-car-project/">one</a> <a href="https://electrek.co/2016/04/22/baidu-self-driving-car-rd-center-silicon-valley-tesla-autopilot/">from</a> Tesla (I love you Elon, I’m just kidding! And props to Tesla).</p><h4>4) Premature scaling is the main reason of death of early stage startups</h4><p>“Don’t try to teach me that, I’m pretty sure I listen to more startup podcasts than you!” says you. True, but the new thing here is you should consider GPUs and training hardware as employees in that equation. Once you hire/buy more than you need you will spend a lot of energy finding use for your surplus. Managing clusters can be hard and large scale HPC for deep learning is a topic of research itself. My suggestion here is to always make sure all your GPUs are constantly in use before you buy another one. You can spend as big as Google once you are as productive and profitable as Google.</p><p>If your team and company are already big enough, be serious about who you hire to work on the infrastructure. If you hire 10x researchers and leave them hanging, in the best case they will build the infrastructure that is good enough for them, in the worst case they will simply quit! That is not what you want.</p><p><strong>Fun Story. </strong>Once I left the office without getting all my GPUs busy, Niel (comma’s VP of Phone App) gave me such a disappointed look that I developed OFG (Overnight Free GPU) phobia. It is a common issue nowadays.</p><p>Yeah! Working on AI can be both challenging and really fun. Make sure you are a bit thoughtful about handling resources and visualizations and you should be fine. But let me know in the comments if I missed any of your main concerns. As a sane community, let’s make sure together that AI back it up all its hype!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=824b0fbf8b1a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AlphaGo, Tesla, Oculus and the next step of AI in everyday life]]></title>
            <link>https://medium.com/@edersantana/alphago-tesla-oculus-and-the-next-step-of-ai-in-everyday-life-15dd6a0062f4?source=rss-f1bbe40e8ca9------2</link>
            <guid isPermaLink="false">https://medium.com/p/15dd6a0062f4</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Eder Santana]]></dc:creator>
            <pubDate>Wed, 23 Mar 2016 02:41:48 GMT</pubDate>
            <atom:updated>2016-03-23T02:41:48.739Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PUrH1Q8oUBVIVU8wrP4cxA.jpeg" /><figcaption>image credit: DeepMind, Gazette Review, Techradar</figcaption></figure><p>Humans just got beaten by computers again. This time in the board game Go, playing against a program called AlphaGo that was developed by Google DeepMind. Here we ponder possible further applications of the technology behind AlphaGo in our lives. We will do this thinking talking about products such as those developed by Tesla, Oculus, and other modern tech companies.</p><p>To understand how AlphaGo works, first let us talk about Machine Learning.</p><blockquote>In Machine Learning we write program architectures and let the specific architectural parameters be adapted from data to better solve problems.</blockquote><p>For example, neural networks are now the most popular of such architectures, the one behind AlphaGo. Training neural networks requires an architecture (number of layers, number of weights per layer, types of connections, etc) and an algorithm to adapt the network weights using example pairs of inputs and desired outputs. For example, if we want to differentiate dogs from cats in an image, we could write a neural network and show it several examples of cat and dog pictures (which would not be hard to get online) and ask the network to classify them.</p><p>Afterwards, the learning algorithm corrects the weights every time the network makes a classification mistake. The main difference between classifying cats and dogs and winning a game of Go is the number of actions required to solve the problem. Given a picture as input our cat/dog network’s classification is either right or wrong and we can tell it how to fix its mistake. On the other hand, given a picture of a board in a Go match, there are many possible actions a player can take and although an experienced player may know what to do by reasoning and intuition, we may never know the optimal action. In such a complicated scenario, training a network to play Go starting from random weights, making random actions, and using only the information about who wins the game in the end is just too hard of a problem, taking infeasibly long to successfully train a network (assuming that is possible at all).</p><p>To make the problem tractable the AlphaGo team used a good old idea:</p><blockquote>To bootstrap its bot, the AlphaGo team first used human movements as targets instead of the end result of the game.</blockquote><p>In other words, instead of playing to win the match, the early versions of AlphaGo were trained to simply replicate a good player’s move in a given position. This reduces the complexity of learning Go moves to something as simple as telling cats from dogs. Looking at the game ata move by move scale, DeepMind scientists could tell AlphaGo’s neural network if it did reproduce the human move and, if it didn’t, they could tell how wrong it was.</p><p>Obviously, acting like a human is not always the best choice for winning the game, but once AlphaGo could reproduce human moves it had a better start than having not clue about the game at all through random initializations. From that better starting point, AphaGo entered its next learning stage:</p><blockquote>AlphaGo went from playing to match humans to playing to beat them.</blockquote><p>The more data the better is common knowledge. The more human experience AlphaGo could soak up the better its start. But to go to the next level, AlphaGo needed more data than humans could provide. Luckily AlphaGo could also play against itself for thousands of games at a rate much faster than the Go community could play. With such large amounts of data, one could use a simple statistical regression to calculate what are the moves most likely to win the game.</p><p>And that was another proof of the importance of data collection. What AlphaGo did differently was showing that data can be reliably generated by AI once it catches up with human behavior, even if in a limited aspect. Human copying AI thus has the potential to introduce an exponential inflection point in the data collection process.</p><p>Go rules are fairly simple, deterministic, and discrete, exactly the opposite of the real world where behaving humans live. So let us do a thought experiment and try to extrapolate that idea: since humans are generating data all the time about how to do things, let us think what if we could dump all that information into learning agents.</p><blockquote>Every moment in our lives we are labelling data. When we drive to work, when we clean the house, or stack toolboxes in the garage. Every human behavior is a data point about how to do something and here we propose that all that could be used to bootstrap neural networks.</blockquote><p>That idea takes us to the next hero of our story, Tesla, that by now we all know is developing self driving cars. Although Google, Baidu, and others are also working in self-driving cars, Tesla has a massive advantage:</p><blockquote>Tesla has human driven cars in the market (packed with sensors).</blockquote><p>All it has to do now is to use all its sensors to record videos or LIDAR scans of surroundings, as well as destination information picked up on the on-board computer and possibly some overall traffic analytics. Using all that information as an input picture of the “board”, Tesla’s self-driving cars neural networks could be initialized by learning to reproduce human drivers in a milliseconds base.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F1zpjcsGiZJY%3Fstart%3D600%26feature%3Doembed%26start%3D600&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D1zpjcsGiZJY%26feature%3Dyoutu.be%26t%3D10m&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F1zpjcsGiZJY%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/5cef1025b6ae6d699c3f73bdafbfcef5/href">https://medium.com/media/5cef1025b6ae6d699c3f73bdafbfcef5/href</a></iframe><p>Nvidia showed at GTC 2015 a video of a remote controlled car trained to simply move like a human would given its front camera stream.</p><p>Last December, Bloomberg also featured George Hotz that applied similar ideas in his own self-driving car.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FKTrgRYa2wbI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DKTrgRYa2wbI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FKTrgRYa2wbI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d7aa851e147bd7373220a2dbbdf4c63b/href">https://medium.com/media/d7aa851e147bd7373220a2dbbdf4c63b/href</a></iframe><p>One can argue that driving is not about taking humans actions each second, but about safely taking people from an origin point to a destination. On the other hand, just like AlphaGo did, Tesla cars could drive themselves for thousands or millions of trips on simulated roads. Simulated and real roads don’t quite offer the same risks and such, but that is a possibility we shouldn’t discard.</p><p>On the following article Jack Clark reports that Google indeed uses both real and simulated roads information for training its self-driving cars.</p><p><a href="http://www.bloomberg.com/news/articles/2016-03-18/why-google-wants-to-sell-its-robots-reality-is-hard">Why Google Wants to Sell Its Robots: Reality Is Hard</a></p><p>But as far as I know we still have to see something about how Tesla uses human drivers data to enhance its self-driving system. I would also not be surprised if every modern car company start shipping their cars with powerful sensors for “parking assistance” ;-).</p><p>When discussing simulated environments, virtual and augmented reality (VR and AR) must be brought into the conversation. Companies like Oculus have can impact everyday life AI by collection ubiquitous human behavior data. That data can come from immersive technologies such as Oculus Rift, MS HoloLens, Meta, and soon to be everywhere Internet of Things sensors. Want to train a robotic arm to reach and grasp objects? Just spin it off from Oculus or <a href="https://www.leapmotion.com/">LeapMotion</a> data:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FdbYP4bhKr2M%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DdbYP4bhKr2M&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdbYP4bhKr2M%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/b9db574730e1942ed0c556ae354b20af/href">https://medium.com/media/b9db574730e1942ed0c556ae354b20af/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FGOTRXM4O9L0%3Ffeature%3Doembed&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DGOTRXM4O9L0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FGOTRXM4O9L0%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/4516dc18d7fd6f1eef178478ebd2c528/href">https://medium.com/media/4516dc18d7fd6f1eef178478ebd2c528/href</a></iframe><p>There are plenty of opportunities. Products that stimulate humans to generate data abound. The difference now is that such slow processes should be accelerated after a certain point, where AI can kick in and start generating simulated and real world data. I wouldn’t be surprised if all product centered companies start to develop or partner with automation companies.</p><p>Since the potential is there, here is a question for the reader:</p><blockquote>Which kind of human activity would you like to automate? How would you engage people to start producing labeled data? Finally, how would you scale up the data generation process using AI?</blockquote><p>Comment your start-up ideas below.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=15dd6a0062f4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>