<?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 Denys Kovalenko on Medium]]></title>
        <description><![CDATA[Stories by Denys Kovalenko on Medium]]></description>
        <link>https://medium.com/@deniskovalenko?source=rss-fe46ce9ae72e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*GtSY5l5PmnfCbgUeRihqqg.jpeg</url>
            <title>Stories by Denys Kovalenko on Medium</title>
            <link>https://medium.com/@deniskovalenko?source=rss-fe46ce9ae72e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 03:23:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@deniskovalenko/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[Data Science seminar on Autonomous Vehicles @ University of Tartu]]></title>
            <link>https://medium.com/@deniskovalenko/data-science-seminar-on-autonomous-vehicles-46f02b0828e9?source=rss-fe46ce9ae72e------2</link>
            <guid isPermaLink="false">https://medium.com/p/46f02b0828e9</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[tesla]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Denys Kovalenko]]></dc:creator>
            <pubDate>Sun, 18 Mar 2018 20:06:45 GMT</pubDate>
            <atom:updated>2018-03-18T20:39:46.370Z</atom:updated>
            <content:encoded><![CDATA[<p>On March 14 Data Science Seminar was held in Tartu, with companies sharing their knowledge in the field of self-driving vehicles. The highlight of the event was video-conference with Andrej Karpathy — head of AI at Tesla and well-known member of deep learning community.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9QapIOWS_cUQ2ahSXIWCzw.jpeg" /></figure><ul><li>First talk was by Tallinn University of Technology — <strong>TTU&amp;Silberauto</strong> partnered to develops own self-driving car project (they have shown software at NorthStart AI conference), as a platform for students and a testing concept of smart infrastructure on TTU Campus.</li><li><strong>Starship (delivery robot startup, with robots operating daily in 10 cities as of March 2018) —</strong> has demonstrated how software used to localize robot with respect to pavement has evolved.</li><li><strong>Milrem Robotics (Lauri Tammervesk)</strong> — a company from Tartu is developing autonomous vehicle for off-road driving (mainly in forest) with military use cases in mind. They are able to recognize individual trees in forest from far away and to identify whether it’s possible to drive through it or not. They use Lidar, radar and stereo cameras. Challenges — trees change their appearance during the year (but don’t change location). They are already able to segment and classify “world-view”.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UvDOBTSGWPOVlGKe." /></figure><ul><li><strong>Allan Aksiim — rules &amp; laws for tech no-one understands</strong>. The claim is that in Estonia by 2030 90% of kilometers driven would be done by self driving cars (though it mainly means taxis/trucks). There are several problems in legislation, especially in EU: can software be considered a driver? EU laws are slow to change. Currently, the hack that allows testing of autonomous vehicles is that the only requirement for test permission is that vehicle is able to brake in case of an obstacle, and to pass control to human driver if something weird happens.</li><li><strong>Andrej Karpathy.</strong></li></ul><p>This was a fun talk, full of jokes and references about bright future under the umbrella of Musk’s universe. Mostly goes over general trends in switching to deep learning in more use cases, and gives overview of challenges for Tesla autopilot system.</p><ol><li>Autopilot is basically regression problem, that has to predict best steering angle and acceleration for given picture of the world, gathered through 8 RGB cameras, GPS, radars and ultrasonic sensors, as well as sound — think of honk or sound of car rapidly braking nearby.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*szKdFJk_sBk9_rVd." /></figure><p>In second part of talk Andrej gave historic perspective on image recognition, from manual feature extraction and SVMs to CNNs, and a transition from Software 1.0 to Software 2.0, covering his blog post — <a href="https://medium.com/@karpathy/software-2-0-a64152b37c35">https://medium.com/@karpathy/software-2-0-a64152b37c35</a>. Instead of C++ codebase written by humans, we will have (more) weights of neural networks, figured out by optimization, searching through program space. We’ve seen such examples with AI becoming world’s best player at Chess/Go , speech recognition/synthesis switching to neural networks and same with robotics.</p><p>Then he goes over pros and cons of such replacement:</p><p>Pros:</p><ul><li>Homogeneous computational base (basically matrix multiplication + ReLU) that allows to have constant runtime and memory usage</li><li>Specialized hardware allows to have ASICs for Deep Learning and chips with weights of NN wired in, enabling cheap and low-power speech synthesis devices, image classifiers etc.</li><li>It’s possible to adjust performance easily by allowing a bit smaller accuracy.</li></ul><p>Cons:</p><ul><li>Data labeling is hard</li><li>Classes are imbalanced</li></ul><p>He goes through challenges of Tesla and other autonomous vehicle projects, although without explaining how they solve it.</p><p>Problem number one is imbalanced classes — 90% of collected data comes from highways, 90% of recognized vehicles are personal cars and just 0.001% — trams, same with orange color of traffic light — there are much less labeled training samples.</p><p>Also really big issue is learning how to deal with previously unseen objects and situations — if something drops on the road from car driving in front, is it safe to drive? Think of plastic bag vs ladder falling on the road in front of the car.</p><p>Another problem is that speed limit signs, lane marking and others might have additional information (like hours when it applies) and varies a lot in different countries.</p><p>Going away from difficult input data — general issue is that imitation learning has compound error:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*EFIYhDKO0ogD93oZ." /></figure><p>And, of course, humans are imperfect drivers. If a person sees car in front braking fast, because of reaction time she will start braking with ~250ms lag, and car will also by default train to wait for 250ms after seeing a danger and brake only then. This isn’t desirable behavior at least.</p><p>So, how do we train cars to be better than imperfect human labelers?</p><p>Since joining Tesla Andrej’s focus was to replace as much Software 1.0 with Software 2.0.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7dI55a7zgXWXN8uc." /><figcaption>Current state</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gXawncnjTA_w9Kti." /><figcaption>Goal — to have much more use cases covered by Software 2.0 — machine learning algorithms</figcaption></figure><p>Currently, Tesla car has around 500 Mb binary blob with regular software, and trained ML models exported as weights into C++ code.</p><p>Here is an example of such shift on use case of classifying whether car is parked.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r6fGHggcNzuUilhi." /></figure><p>Also, Tesla has challenges training “social agreements” — in the merger of 2 roads people follow zipper pattern — car from the left lane, car from the right lane and so on. Tesla just jump in if it’s possible.</p><p>Despite all of that, Tesla cars driving around is world’s biggest deployment of robots so far — 250k, and their drivers are serving as free data labelers for Tesla autopilot, so they have a big advantage compared to competitors, as in Deep Learning — the more data, the better.</p><p>Thanks to Institute of Computer Science of UT for organizing awesome event and thanks to all speakers!</p><p>Looking forward to see self-driving future!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=46f02b0828e9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>