<?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 Lucas Fernandez on Medium]]></title>
        <description><![CDATA[Stories by Lucas Fernandez on Medium]]></description>
        <link>https://medium.com/@lucas_fernandez?source=rss-e18ab2fabb8a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*-0AzbVd9cZTSikPB.</url>
            <title>Stories by Lucas Fernandez on Medium</title>
            <link>https://medium.com/@lucas_fernandez?source=rss-e18ab2fabb8a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 07:29:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lucas_fernandez/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[Will Quantum Computing follow Machine Learning Path]]></title>
            <link>https://medium.com/@lucas_fernandez/will-quantum-computing-follow-machine-learning-path-e556cb4f79ca?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/e556cb4f79ca</guid>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Fri, 04 Oct 2019 08:08:26 GMT</pubDate>
            <atom:updated>2019-10-04T08:08:26.103Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nv0-jcCx9DPP6QpEPIwtQA.jpeg" /><figcaption>IBM Quantum Computer, credit IBM</figcaption></figure><p>If you follow a bit of the specialized press, you have probably noticed that two weekends ago, on Sept, 24th 2019, <a href="https://www.ft.com/content/b9bb4e54-dbc1-11e9-8f9b-77216ebe1f17">The Financial Times stated that Google reached quantum supremacy</a>. Their claim is based on a now-removed scientific publication on the NASA website. Google and NASA teamed up on quantum computing research.</p><p>What is quantum supremacy? It is a crucial milestone in quantum computing, reached when a quantum computer achieves a task that could not have been performed by a classical supercomputer. In the case of this paper, the team claims they computed a validation proof on a random generation function that would have required 10,000 years to a standard super calculator. All of it was due to their 53-qubit specialized computer.</p><p>It’s not a general application. Nonetheless, it marks the passage of a major milestone; helping us to see more clearly the maturity point for reaching a potential roll-out of this technology.</p><p>I believe it highlights the importance of keeping an eye on this technology, not only from a product standpoint but also a security one. One of the first narrow applications would be to break asymmetric encryption, the mechanism over which most of IT security is built. We can easily guess that many agencies will put the financial funds in front of their interest in such a task.</p><p>Interestingly, since the release of this news, <a href="https://www.ft.com/content/cede11e0-dd51-11e9-9743-db5a370481bc">I have seen a bit of drama around Google’s allegation</a>. It would be untrue as they did not achieve to build a general-purpose quantum computer. I imagine that a few teams are frustrated by the fact that they could have claimed the milestone before Google.</p><p>However, in my eyes, its what makes the news so interesting. In machine learning, we did not reach the point of a General Artificial Intelligence able to understand, think by itself, and improve. Still, we are very good at applying the tools we have to a very narrow and specialized task to which Neural Networks definitively outperform human beings.</p><p>I guess we can expect the same kind of narrow applications for quantum. If you drill a bit into the technology, you will discover that there are already a few subsets of the technologies with technical limitations like quantum annealing. Nonetheless, they are near, if not mature, to be applied to solving business problems, thus unlocking new product possibilities that were impossible or too expensive to be offered so far, exactly like machine learning.</p><p>If you want to get more info about Quantum Computing I would recommend you:</p><ul><li><a href="https://www.youtube.com/watch?v=JhHMJCUmq28">https://www.youtube.com/watch?v=JhHMJCUmq28</a></li><li><a href="https://www.youtube.com/watch?v=zvfkXjzzYOo">https://www.youtube.com/watch?v=zvfkXjzzYOo</a></li><li>Chap 4 The Uncertainty Principle, A Brief History of Time, Stephen Hawking</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e556cb4f79ca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Arduino NeuralNetwork Car — Part 5]]></title>
            <link>https://medium.com/@lucas_fernandez/arduino-neuralnetwork-car-part-5-ad7ae3ea7588?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad7ae3ea7588</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[cars]]></category>
            <category><![CDATA[arduino]]></category>
            <category><![CDATA[sensors]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Mon, 13 May 2019 16:00:50 GMT</pubDate>
            <atom:updated>2019-05-13T16:00:50.366Z</atom:updated>
            <content:encoded><![CDATA[<h3>Arduino NeuralNetwork Car — Part 5</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k7FceI9AutsaJq-Yam-aRg.jpeg" /></figure><p>In the previous and last post, I described how to design and train properly the neural network that will drive the car. It is now time to write the code merging them all:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/e8662089726cc01f1f44a6e8a99c99f3/href">https://medium.com/media/e8662089726cc01f1f44a6e8a99c99f3/href</a></iframe><p>You should now be familiar with most of it; I let you upload it in the car’s Arduino and enjoy the experience of driving it and seeing it following your commands and breaking in front of obstacles.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*GkukRvrxT6Mt0IIPXmjJcA.gif" /></figure><p>I hope you enjoyed as much as I did this project and will continue to improve and tweak it. I have, for sure, a few things in mind that would require my 3D printer, a few extra parts, maybe even a webcam and an <a href="https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano/">NVIDIA Jetson Nano</a> that I will probably share with you in the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad7ae3ea7588" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Arduino NeuralNetwork Car — Part 4]]></title>
            <link>https://medium.com/@lucas_fernandez/arduino-neuralnetwork-car-part-4-30548d98d4bf?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/30548d98d4bf</guid>
            <category><![CDATA[cars]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[arduino]]></category>
            <category><![CDATA[tinn]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Mon, 06 May 2019 12:49:20 GMT</pubDate>
            <atom:updated>2019-05-14T15:30:36.287Z</atom:updated>
            <content:encoded><![CDATA[<h3>Arduino NeuralNetwork Car — Part 4</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jcFxv-CBdffEFwuWvMHeXg.png" /></figure><p><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-3-c3988913bb01">In the previous sections,</a> we worked on getting the command from the IR remote to offer the possibility to a user to drive the car, getting the data from our sensors and driving the motors. In this section, we will work on training the neural network in isolation to drive the car adequately. Then in the ultimate section, we will wrap all the codes together to be able to drive the car using a neural network.</p><p>The neural network will be composed of 6 inputs or neurons. The four first will be dedicated to the command sent by the user: forward, turn right, backward, turn left, and the two last for the obstacle sensor left and right. Following the best practice, the hidden layer will be composed of the same number of neurons the input layer plus one. Then the output layer will be formed of two neurons for the motor left and right where 0.0 mean backward, 0.5 stop and 1.0 forward. This will enable in the future a proportional driving of the engines.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*NlpVjn8LUUvploIP-FOFpg.gif" /></figure><p>To get the best potential accuracy from the neural network it is trained on the full list of situations that it will meet and their correct output:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/dda9db16bdbc9de2448f066ec8a39438/href">https://medium.com/media/dda9db16bdbc9de2448f066ec8a39438/href</a></iframe><p>But this is probably creating a strong overfit of the neural network and should be avoided.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/ab9a00200bb1c474e4e12f1830522abd/href">https://medium.com/media/ab9a00200bb1c474e4e12f1830522abd/href</a></iframe><p>After running 5000 iteration with a decaying learning rate here the weight that I got outputted from the training.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/710/1*vhagJh1-W-K-y3Y6pOSCLQ.png" /></figure><p>As you can see the generated output match the training dataset. So we can proceed to the Arduino to test more extensively the result with this code.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/0b8f29338791ff15825b3677144e522b/href">https://medium.com/media/0b8f29338791ff15825b3677144e522b/href</a></iframe><p>And as you can see the output is matching perfectly the the training dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/710/1*r6hkXBy_PQqc3BaejPxBPA.png" /></figure><p>It is quite simple to train the neural network using a powerful computer and then to transfer the weights inside the Arduino.</p><p>What fascinated me is that the behavior of the car can be changed by just changing the training dataset and updating the new weights. In an earlier version of the code, my training dataset was not accounting for the sensors, and the car was hurting the walls. I just had to extend the dataset, retrain and update the weights to fix the behavior without touching a line of code.</p><p>Potential improvement:</p><ul><li>Add a random noise on the input to reduce overfitting</li><li>Shuffle the training sample</li></ul><p>We can now move to the <a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-5-ad7ae3ea7588">final part</a> were we merge all the codes that you saw in a complete software that enables the car to run correctly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30548d98d4bf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Arduino NeuralNetwork Car — Part 3]]></title>
            <link>https://medium.com/@lucas_fernandez/arduino-neuralnetwork-car-part-3-c3988913bb01?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/c3988913bb01</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[arduino]]></category>
            <category><![CDATA[remote]]></category>
            <category><![CDATA[cars]]></category>
            <category><![CDATA[infrared]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Fri, 03 May 2019 07:22:09 GMT</pubDate>
            <atom:updated>2019-05-06T12:51:12.704Z</atom:updated>
            <content:encoded><![CDATA[<h3>Arduino NeuralNetwork Car — Part 3</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L6X08HWwZx4NrTjtP43SdA.jpeg" /></figure><p><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-2-b0e978ccc378">In the previous section</a>, I show how to interface and set up the sensors of our neural network controlled car project. In this section, I will cover the IR remote part of it that will be used to control the car. like usual the full code is on the proper <a href="https://github.com/kasey-/ArduinoNNCar/tree/master/3-remote">Github repo</a>.</p><p>The car kit that we use in this project offers two ways to control the car, by IR remote or by Bluetooth. The second is obviously very attractive as it would enable a full duplex (bi-directional) communication between the vehicle and the remote. But to keep things simples, we will start by using the remote. Later on, we will see if we want to extend the project by using a computer or a smartphone as a remote.</p><p>The IR remote is a simple device that can send commands to the remote car via an integrated IR LED.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4YbT7uCVBfIcuWmKn9tptw.jpeg" /></figure><p>I did not bother torturing myself with trying to interface natively with it and jumped directly on the most recent version of the lib that was used by the car manufacturer. Something like 10 years more recent… seriously guys …</p><pre>/*<br> * IRremote<br> * Version 0.11 August, 2009<br> * Copyright 2009 Ken Shirriff<br> * For details, see <a href="http://arcfn.com/2009/08/multi-protocol-infrared-remote-library.html">http://arcfn.com/2009/08/multi-protocol-infrared-remote-library.html</a></pre><p>The code is based on the <a href="https://github.com/z3t0/Arduino-IRremote">library IRRemote</a> example stripped down to its bare minimum to handle two tasks. First, get commands from the remote. Handle the key held by the user. As a user keep the button pressed on the remote we want to continue to execute the same command. But when he releases the button, we want the car to stop.</p><p>Thanks to the lib, recognizing the command the code is simple</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/6b46b7e14555774d952c3d7c66f259a4/href">https://medium.com/media/6b46b7e14555774d952c3d7c66f259a4/href</a></iframe><p>And we use the execution time trick with the function millis to create a virtual watchdog. If we receive the code which indicates the user is keeping is finger pressed, we continue to execute the last command. If we do not receive a command since one cycle (the remote sends a command every 100ms), we stop the car.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/81c026518a1e4e272ed7c5665bb8e01a/href">https://medium.com/media/81c026518a1e4e272ed7c5665bb8e01a/href</a></iframe><p>Or just turn a LED on and off for a more showoff effect :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*FRTd3AB5iLmfMiFFHyCPvg.gif" /></figure><p>Could be improved:</p><ul><li>the IRRemote lib offers the possibility to emulate a remote. If funny, this feature is totally useless for our project but still compiled and packaged with the rest of the lib to the remote. Stripping this part from the code or finding an alternative lib that would just receive and decode commands would be more efficient. Nonetheless, so far this is not an issue, 30% of the memory of the microcontroller are used by the current implementation.</li><li>the remote allows only to send one command a time, we could use the numeric pad to merge commands like “up” and “right” at the same time. This would enable richer input for the neural network, allowing to drive more behaviors</li><li>the usage of a <a href="https://github.com/contrem/arduino-timer">timer library</a> or a more <a href="https://github.com/tinkerspy/Automaton/wiki">advanced state library</a> could clean up the code of the call to millis</li></ul><p>We are at the end of the hardware interfacing of the vehicle. <a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-4-30548d98d4bf">The next post</a> is dedicated to creating and training the neural network that will be used to drive the car, based on the user and sensors input.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c3988913bb01" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Arduino NeuralNetwork Car — Part 2]]></title>
            <link>https://medium.com/@lucas_fernandez/arduino-neuralnetwork-car-part-2-b0e978ccc378?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0e978ccc378</guid>
            <category><![CDATA[sensors]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[arduino]]></category>
            <category><![CDATA[obstacles]]></category>
            <category><![CDATA[cars]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Mon, 29 Apr 2019 12:54:03 GMT</pubDate>
            <atom:updated>2019-05-03T07:28:06.134Z</atom:updated>
            <content:encoded><![CDATA[<h3>Arduino NeuralNetwork Car — Part 2</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A4AOquXMM0VKqJEbZlDfRQ.jpeg" /></figure><p><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-1-45debeed6eb">In the previous section</a>, I show how to control the motors of our neural network controlled car project. In this section, I will cover all the sensors available in the kit, how to put them into practice and we will select the ones making the most sense to feed the neural network. All the source code for this section is available on the <a href="https://github.com/kasey-/ArduinoNNCar/tree/master/2-sensor">ArduinoNNCar</a> Github repo.</p><p>The car kit available here is provided with three types of sensors:<br>* Ultrasonic distance<br>* IR obstacle detectors<br>* IR line detector</p><p>We can already sense that we could use them two create two types of robots, one a 100% autonomous following lines or another one piloted avoiding collisions. In this series of posts, we will work on the 2nd one. But I will review the basis of the IR line detector anyway.</p><p><strong>Ultrasonic Distance Sensor</strong></p><p>The ultrasonic distance sensor is a very classical sensor type HC-SR04.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A4AOquXMM0VKqJEbZlDfRQ.jpeg" /></figure><p>Essentially, the sensor is composed of two elements an ultrasound emitter and a receptor. Activate the output trigger for ten microseconds which create a pulse sent by the ultrasonic HP, then measure the time it takes to get the echo on the microphone output. Divide it by two and the speed of sound, and you get your distance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/456/1*ZEhEcmKhkGVgxhLSEEkeDA.png" /></figure><p>Bellow the pseudo code with the pin used on the car kit</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/cbfe67c4d0af43b74151df0696154d22/href">https://medium.com/media/cbfe67c4d0af43b74151df0696154d22/href</a></iframe><p>Note: the timeout is set up to trigger at more than 50 cm as we do not care of such usage in the context of driving our car.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*pwcqCfV09OQdow1bOMg5qQ.gif" /></figure><p>As you can see on the signal acquisition below the sensor is quite unstable, and from time to time the sensor does not capture the distance. To clean it up I found the easiest solution is to acquire the measure from the sensor twice and to take the bigger distance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*uqV-dV9ny67nR2-_GTPkEQ.gif" /></figure><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/3ff048f80a3e6882913e33c0a7bb76bb/href">https://medium.com/media/3ff048f80a3e6882913e33c0a7bb76bb/href</a></iframe><p>Sources:<br>* <a href="https://dronebotworkshop.com/hc-sr04-ultrasonic-distance-sensor-arduino/">https://dronebotworkshop.com/hc-sr04-ultrasonic-distance-sensor-arduino/</a></p><p><strong>IR Obstacle Detectors</strong></p><p>The IR obstacle detectors are composed of two elements. An IT emitter and a receptor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NViI_00XvGgPvaBtL-3NPg.jpeg" /></figure><p>The output of the sensor is activated by the quantity of light received by the receptor. Two potentiometers adjust the sensor: the first one define the frequency of oscillation of the emitter and the second one the power of emission.</p><p>I share here the process I followed to calibrate the sensor as it was a pain to do:</p><ol><li>Completely cover the sensor you are not calibrating to avoid any disturbance from one sensor to the other</li><li>Put something at 4–5 cm from the sensor</li><li>Put the led power at his maximum. You can use your smartphone camera to see when the LED start to turn on as smartphone usually camera do not filter IR</li><li>Start to adjust the frequency potentiometer until you know the trigger activating, continue to turn the potentiometer until the trigger is not activated anymore and come back to a midrange position</li><li>position your obstacle at 20 cm and adjust the power potentiometer until you detect it and the led fade off when you remove the obstacle</li><li>Then hide the sensor and calibrate the second one.</li></ol><p>Note: the sensors are pretty unstable and calibrating them is a frustrating process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*siEfh-V4vxL29pCtvfy-Pw.gif" /></figure><p>As the output of the sensor is digital (0/5v) the code to read if the sensor is triggered or not is straightforward:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/954a2f060f5a0e7eb7bd4d05fef10b8a/href">https://medium.com/media/954a2f060f5a0e7eb7bd4d05fef10b8a/href</a></iframe><p>Sources:<br>* <a href="http://irsensor.wizecode.com/">http://irsensor.wizecode.com/</a></p><p><strong>IR Line Detector</strong></p><p>The IR line detector is equally simple than the previous sensor with zero setups.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3pgNjWt5XOdcSkjJKCN_8w.jpeg" /></figure><p>It uses an integrated sensor which outputs 5v if a white line is detected otherwise zero. The setup of the sensor is done mechanically: the physical distance between the sensor and the line is adjusted.</p><p>Bellow the code to interface with it</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/3a94a626a90c5b6c257dd372210bf49b/href">https://medium.com/media/3a94a626a90c5b6c257dd372210bf49b/href</a></iframe><p>Source:<br>* <a href="https://showmeyourcode.org/how-to-use-tcrt5000-with-arduino/">https://showmeyourcode.org/how-to-use-tcrt5000-with-arduino/</a></p><p>After a bit of reflection, using the ultrasonic distance sensor is tempting. But as we want to avoid obstacles in the first iteration of the code, we will restrict ourselves in the usage if the obstacle detector. Simpler is better. In a later iteration, we will use the ultrasonic sensor.</p><p>Could be improved:<br>* a bit of filtering on all the types of sensor would create a more resilient signal. For example, the <a href="https://github.com/thomasfredericks/Bounce2">Bounce2 library</a> could ensure we have clean signals fed into the neural network in a later stage<br>* a 3D printed cone around the LED would improve the precision by protecting the sensor from external disturbance and keep the light focalized on the frontal area</p><p><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-3-c3988913bb01">The next section</a> will be dedicated to the understanding of the IR remote available in the car kit to send the commands to the robot and drive it through the neural network.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0e978ccc378" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Arduino NeuralNetwork Car — Part 1]]></title>
            <link>https://medium.com/@lucas_fernandez/arduino-neuralnetwork-car-part-1-45debeed6eb?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/45debeed6eb</guid>
            <category><![CDATA[l298n]]></category>
            <category><![CDATA[cars]]></category>
            <category><![CDATA[motor]]></category>
            <category><![CDATA[arduino]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Sat, 27 Apr 2019 12:58:43 GMT</pubDate>
            <atom:updated>2019-04-29T12:55:54.208Z</atom:updated>
            <content:encoded><![CDATA[<h3>Arduino NeuralNetwork Car — Part 1</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pp70NDbmUvqHJ3rgjNmkOQ.jpeg" /></figure><p><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-0-43c30e4d4886">In the previous section</a>, I outlined our neural network controlled car project. In this section, I will quickly cover the motor part of the car as we will need to drive it at the end of the project.</p><p>The booklet and schematics of the car are quite limited in information about the motor drivers. But a quick look to the schema and on the driver chip informs us that both of the sides of the cars’ motors are connected to a Dual H-bridge L298N chip.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uwVGtH4Wi-RadrvnQjJmSA.jpeg" /></figure><p>To speed up the implementation in the Arduino I checked the libraries available and settled for the <a href="https://github.com/AndreaLombardo/L298N">L298N by AndreaLombardo</a> which is established and super lightweight. After downloading it from the Arduino IDE, a quick review of the schema of the car to know which of the Arduino’s pin drive the driver and we are ready to go. The following code illustrates how I tested the motors driving.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/31098975bcc5bb89f747810d8fe075e7/href">https://medium.com/media/31098975bcc5bb89f747810d8fe075e7/href</a></iframe><p>Good to know:<br>* The L298N is for one H bridge, to drive the full driver we need to create two instances. That’s handy if you want to drive four or six motors with two or three L298N.<br>* The L298N motor speed can be driven by a PWM, so the max speed is limited to the max PWM which is 255</p><p>The code:<br>1. Create the two instance of L298N to drive the two sides of the car<br>2. Setup: stop the motors and define the speed at maximum<br>3. Run a test sequence: M1 + M2 Stop, then forward, backward for each side then simultaneously.</p><p>You can now drive pretty easily the motors and so the car.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*jHR_d7NaNdfJmHkBoaK41w.gif" /></figure><p>Could be improved:<br>* alternate the direction of rotation of the motors to turn the car like a tank<br>* implement an advance steering of the car by reducing the speed of the motors inside the curve<br>* modify the car by adding another L298N or a driver able to handle four motors, to drive each motor independently would enable a smoother control in the curves and would make more interesting the usage of the neural network as it would do it.</p><p><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-2-b0e978ccc378">The next section</a> will be dedicated to the understanding of the sensors available in the car kit.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45debeed6eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Arduino NeuralNetwork Car — Part 0]]></title>
            <link>https://medium.com/@lucas_fernandez/arduino-neuralnetwork-car-part-0-43c30e4d4886?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/43c30e4d4886</guid>
            <category><![CDATA[arduino]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[cars]]></category>
            <category><![CDATA[tinn]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Thu, 25 Apr 2019 19:46:36 GMT</pubDate>
            <atom:updated>2020-04-01T04:36:55.785Z</atom:updated>
            <content:encoded><![CDATA[<h3>Arduino NeuralNetwork Car — Part 0</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DW7GUugJQdll0ViPyrNbEg.jpeg" /></figure><p>A few weeks ago I saw two videos on youtube. The first one was google presenting their adaptation of <a href="https://www.youtube.com/watch?v=xFbLHf0lIcw">Tensorflow for Microcontroller</a> and the second was presenting the work of Lantern embedding <a href="https://www.lantern.systems/projects">neural network in a hardware project to detect kickback on circular saws</a>. Suddenly I realized that I never checked how to mix two of my favorites topics: hardware and machine learning.</p><p>The common idea about neural networks is that they consume a ton of computing power. For deep neural networks I totally agree, you can forget the idea of performing a convolutional neural network on an Arduino. But simple things should be achievable. So I started to dig into the net … and long story short, it works super well !!!!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*V02cFI6NkqX14CwOI5_hCQ.gif" /></figure><p>In the present case, I am using an IR remote to send instructions to the car. A neural network analyses the user command plus the car’s sensors and drives the motors accordingly to the training dataset. For proof, I made a few improvements over the behavior of the car, and I did so just by extending the training dataset, retraining, then updating the weights in the car’s neural network.</p><p>In this series of posts, I will cover every aspect of the project in the order in which I explored them:</p><p><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-0-43c30e4d4886">0. run a minimal neural network on an Arduino</a><br><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-1-45debeed6eb">1. control the engine on the car</a><br><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-2-b0e978ccc378">2. read all the sensors</a><br><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-3-c3988913bb01">3. understand the commands from the remote</a><br><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-4-30548d98d4bf">4. car neural network</a><br><a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-5-ad7ae3ea7588">5. tie everything together</a></p><p>If you want to follow along, you can buy the same $70 <a href="https://www.amazon.com/UNIROI-Ultrasonic-Infrared-Tracking-Required/dp/B07CVS1LBT/">kit than me</a>. It contains everything that you need to run the same experiment. All the code is available on a dedicated public <a href="http://github.com/kasey-/ArduinoNNCar">Github repo</a>. I organized the folders following the same order that the part of this series.</p><p>Finding a library able to run on an Arduino was easy. An Arduino can run code in C++ or C and after a few minutes of GitHub research, I found the perfect candidate: <a href="https://github.com/glouw/tinn">Tinn</a></p><blockquote>Tinn (Tiny Neural Network) is a 200 line dependency free neural network library written in C99.</blockquote><blockquote>Tinn is great for embedded systems. Train a model on your powerful desktop and load it onto a microcontroller and use the analog to digital converter to predict real-time events.</blockquote><p>So let’s start hand-on by testing the library on a computer. <a href="https://hackaday.com/2018/04/08/tiny-neural-network-library-in-200-lines-of-code/">Hackaday</a> did a quick example of an implementation that I adapted:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/6939fddd24679710c29548f6ccee8dba/href">https://medium.com/media/6939fddd24679710c29548f6ccee8dba/href</a></iframe><p>To run it you need to compile using make and execute the produced binary</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*9Hj_BR5ouisq8XqUsvV_XA.png" /></figure><p>As you can see the output of the neural network match the training dataset.</p><p>Let’s now run it into an Arduino. To do so, you need to import the Tinn.c and Tinn.h files in the Arduino IDE by including the library as an extern C. Then you can use it without any modification. Hereafter the code to display on the serial terminal the prediction of the neural network given the training input.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/c637da88d233cb7092276dfa800bddd7/href">https://medium.com/media/c637da88d233cb7092276dfa800bddd7/href</a></iframe><p>The result is displayed on the Arduino serial output. As you can see we get the same values than the ones outputted by the “train” binary in our computer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*75i1PjX7pF7D_of3s2SKHw.png" /></figure><p>We have now bootstrapped our project. We can train the neural network on a computer then run it on an Arduino. I hope to see you in the <a href="https://medium.com/@lucasfernandez_74743/arduino-neuralnetwork-car-part-1-45debeed6eb">next post</a> where we will explore how to drive the motors of the car.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=43c30e4d4886" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Two blueprints for agile specifications]]></title>
            <link>https://medium.com/@lucas_fernandez/functional-specifications-for-agile-teams-ecb16895d4b3?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/ecb16895d4b3</guid>
            <category><![CDATA[functional-specifications]]></category>
            <category><![CDATA[blueprint]]></category>
            <category><![CDATA[user-stories]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[epic]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Wed, 30 May 2018 16:05:50 GMT</pubDate>
            <atom:updated>2018-05-31T14:03:34.322Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HYabs9UmGucX5joGvgB6mA.jpeg" /><figcaption>Joy Oil Co Ltd — City of Toronto Archives on Wikipedia</figcaption></figure><p>It will soon be ten years since I started doing project management in a form or another. One of the significant activities that I perform is translating consumers’ needs in requirements that can be efficiently converted into technical implementations by engineers. Over the years, I documented myself about most of the conventional methodologies of specifications from UML to User Stories in Waterfall and Agile context. The question of how I specify software often come in discussions, so I thought it would be interesting to share my current process here and to exchange about what you use on your side.</p><h3>Core principles</h3><ol><li>The specifications are here as a reference for the developer but must also act as support to exchange around the feature.</li><li>The specification must evolve to follow the new version of the product.</li><li>They are not a static artifact consumed and forgotten.</li><li>The specification is here to detail the context in which the feature is implemented. Knowing the context (personas, system, goals, …) enables us to do all the micro-adjustment during the implementation in an autonomous manner.</li></ol><h3>Organisation</h3><p>In an agile fashion, I group my features following this taxonomy:</p><ol><li>An <strong>initiative</strong> is a group of related epics that, when delivered, will provide a new set of customers features. In my previous company, it was taking us between three and six months to deliver a full initiative on a product. Initiatives are a bit unusual and are not described more in detail in this article.</li><li>An <strong>epic</strong> is a group of user stories with a common goal. Usually, by convenience, I group all the user stories around the management of one resource like users or baskets in the same epic.</li><li>A <strong>user story</strong> is the single element of specification below which you cannot go. It is a usable feature or a subset of a usable feature that can be delivered to a customer. I try to cut my user stories as small as possible to be sure that if an unimportant element cannot be delivered in one sprint, its essence will be available in any case to a customer. For example, in resources listing, I always split pure listing from the sugar-like research and filtering that can wait for the next sprint.</li></ol><h3>Epic</h3><p>I use epic specs to define as best as possible the context of the user stories an epic contains. To achieve that, I share the following.</p><ul><li><strong>Purpose</strong> of the epic: Why is it there, and what do users want to achieve through its usage?</li><li><strong>Personas</strong>: This is the list of personas that will use the product. Code school in the first free chapter of <a href="https://www.codeschool.com/courses/the-elements-of-web-design">The Elements of Web Design</a> provides an excellent basic persona template that I advise you to watch.</li><li><strong>Characteristic</strong> of the epic: If the epic is related to the management of a resource, this is a table that summarises all the attributes of the resource. This is a direct legacy of the <a href="https://www.uml-diagrams.org/property.html">UML property</a> system.</li><li><strong>Illustration</strong>: I illustrate the resource or business process modeled in the epic so the developer can see the business logic inside the software. Pictures, state diagrams, YouTube videos, prototypes, and competitive products are excellent sources of illustrations.</li></ul><p><a href="https://docs.google.com/document/d/1XkJHr6KPqjo_JrVjfyLuhdxnrPgqm-nWGN7yQG9PZK4/edit?usp=sharing">Google doc Epic Blueprint</a></p><h3>User Stories</h3><p>I use user stories to define as best as possible the way the system should behave without influencing the designer or the developer regarding the design or technical solutions. To achieve that I share the following things:</p><ul><li><strong>User stories,</strong> while writing, I strive to define the received beneficial parts of the user stories as they are explicit in the intent of the features implemented.</li><li><strong>Source of the request</strong>, usually a customer ticket. This can give context about the feature.</li><li><strong>Rights</strong> will be involved to be able to perform the user story<br>The initial situation is how the system is before the use case starts, for example, resources registered or not, etc.</li><li><strong>Use case step by step,</strong> I write one line per action and try to define the nominal navigation path through the app without detailing it to let the UX part do its job.</li><li><strong>Exceptions,</strong> I list all the exceptions that I think could occur and how to manage them. I do not make explicit the trivial ones but focus on exotic cases like resources deletion conflicts, etc — all the things that the developer could forget to handle.</li><li><strong>Final situation,</strong> all resources created, resources updated, and events sent are defined here.</li><li><strong>Final pixels,</strong> if there are views involved in the user story, they must be ready when the implementation is started and linked to the user story as they give a massive context to the software that must be produced.</li></ul><p><a href="https://docs.google.com/document/d/130fajXr2yxS91-S2wos8JLmga2_jLEiDvpKVd4Ctglo/edit?usp=sharing">Google doc User Story Blueprint</a></p><p>Notes:</p><ul><li>When I model a process interaction instead of user interaction, I switch on a <a href="https://jtbd.info/replacing-the-user-story-with-the-job-story-af7cdee10c27">job story</a>. The template is practically 100% identical.</li><li>The estimate of the user story is not defined here as it is usually held by the ticket who enable the management of the user story.</li></ul><h3>Lesson Learned</h3><p>Writing specifications is an investment that, as a PM, you must make to let your team members focus as autonomously as possible on doing their jobs efficiently while you work on other aspects like sales and marketing. Specifications will never remove all need for clarification and exchange around the product, but they will save you from a lot of back and forth with your team to ensure the product is right. Plus, they will hold the memory of arbitrations done in the past, which will help you avoid rechallenging choices constantly. The more time you invest in writing specifications, the better they will be during implementation. However, keep in mind that perfection is achieved only through practice. Personally, it takes me around two hours to specify a straightforward epic. This time does not include the exploration previously done.</p><p>I am very curious to know how you handle specifications, so if you share your methods in the comments section, I will be more than happy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ecb16895d4b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vision, Google’s first step for innovation]]></title>
            <link>https://medium.com/@lucas_fernandez/vision-googles-first-step-for-innovation-5f9842e30430?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f9842e30430</guid>
            <category><![CDATA[moonshots]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[vision]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Tue, 22 May 2018 17:22:56 GMT</pubDate>
            <atom:updated>2018-05-23T07:36:37.576Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KiBL_6Pq9Qu9T2wmCETCsg.jpeg" /><figcaption><a href="https://unsplash.com/photos/U2uKrI4lci8?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">NASA</a> on <a href="https://unsplash.com/search/photos/nasa?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Since a year ago when I read this article in the <a href="https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-the-perfect-team.html">New York Times</a> publishing the finding of Google about how to compose the perfects teams, I am following with careful attention the new addition made on the <a href="https://rework.withgoogle.com/">rework platform </a>by Google. I love their publication about management and organization shaping. So, I thought you could be interested in what Google published about the perfect environment to foster innovation.</p><p><a href="https://rework.withgoogle.com/print/guides/5748556159778816/">Google insists on developing the five following points</a> to help <a href="https://www.linkedin.com/pulse/make-innovation-flow-through-companies-lucas-fernandez-mba/">innovation flow</a>:</p><ol><li><strong>Vision</strong>, make sure everyone knows where the organization is heading</li><li><strong>Autonomy</strong>, allow employees to define their own work as much as possible</li><li><strong>Intrinsic motivation</strong>, hire naturally curious people who like to learn</li><li><strong>Risk-taking</strong>, enable employees to feel psychologically safe to take risks and try new ideas</li><li><strong>Connection</strong>, make it easy for employees to find partners and work together</li></ol><p>The first point about having a vision is, from my point of view, a must have and one of the first questions I ask entrepreneurs. Carefully building one for which the organization can stand is vital for a company. Sadly, sharing it rightly is a too-often forgotten duty of management despite critical statistics:</p><blockquote>“According to a 2016 Society for Human Resource Management report, 76 percent of working adults need to feel their job is meaningful to engage with their company. If the meaning isn’t clear, employees can fall victim to the gravity of purposeless employee experience.” — <a href="https://www.inc.com/elizabeth-dukes/how-to-get-employees-invested-in-your-mission-why-it-matters.html">Inc</a></blockquote><p>Vision <a href="https://youtu.be/u4ZoJKF_VuA?t=7m32s">brings alignment</a> through the whole company on common goals and priorities.</p><p>The best illustration for a vision being able to align an entire country on a precise target was made by John F. Kennedy on May 25, 1961, when he gave the Moon Shoot Speech to Congress and said the following:</p><blockquote>“I believe that this nation should commit itself to achieving the goal, before this decade is out, of landing a man on the moon and returning him safely to the Earth.” — John F. Kennedy</blockquote><p>Two years later, JFK was assassinated in Dallas, but eight years later, on July 20, 1969, NASA’s Apollo 11 mission landed the first humans on the moon. The goal defined by JFK was clear and left no ambiguity about what should be done to achieve the desired output. In the meantime, it was creating the right stretch to push every organization engaged in the space run to innovate to achieve the target. I let you picture the number of questions this statement answered about the level of safety, technology to invent to allow a safe return, etc.</p><p>To help you define, stretch, and follow your mission, Google offers two tools:</p><ol><li><a href="https://rework.withgoogle.com/guides/managers-set-and-communicate-a-team-vision/steps/create-a-vision-with-the-team/">a two days workshop</a> about how to build a mission, vision, and values which offers an excellent structure to this highly intangible output.</li><li><a href="https://rework.withgoogle.com/guides/set-goals-with-okrs/steps/develop-team-OKRs/">the OKR system</a> by which each employee define his objectives and key-result to sustain the company goals. Of course, to work, your OKRs will need to rely on a strong vision and mission.</li></ol><p>I advise you to do a mission and vision exercise for yourself if the purpose of your organization is not explicitly stated. When done, share it with the rest of your teams, display it and repeat it until they all know it by heart. If innovation is one of your core values, it will kick off a shift at the heart of your organization. By creating an environment favoring innovation as I describe it in <a href="https://medium.com/@lucasfernandez_74743/make-innovation-flow-through-companies-f9761a2d9fd5">my previous article</a> and by refining the management of your like Google suggested it, it should help you to get an edge against the competition.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f9842e30430" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My Five Inspirational Sources of Innovation]]></title>
            <link>https://medium.com/@lucas_fernandez/my-five-inspirational-sources-of-innovation-c093d950b6c8?source=rss-e18ab2fabb8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/c093d950b6c8</guid>
            <category><![CDATA[tools]]></category>
            <category><![CDATA[academic]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[research-and-development]]></category>
            <category><![CDATA[trends]]></category>
            <dc:creator><![CDATA[Lucas Fernandez]]></dc:creator>
            <pubDate>Mon, 14 May 2018 13:58:41 GMT</pubDate>
            <atom:updated>2018-05-14T13:58:53.600Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xnDjT_FF-5OBCrTgHBF6aA.jpeg" /><figcaption>Mikael Kristenson on Unsplash</figcaption></figure><p>In the many years that I have been thinking about and conceiving new products, I have matured a process to keep myself connected with innovation in our fast-paced society. I share here my five major sources of insight so you too can stick closer to innovation.</p><p>From the broadest to the most focused, I consider five sources of insight:</p><ol><li>Megatrends</li><li>Industry trends</li><li>Academic</li><li>Consumers</li><li>Internal</li></ol><h3>Megatrend, trends at the societal scale</h3><p>Our society is shaped by major trends. It is essential to keep yourself updated to understand in which very big picture your company, customers, and products could live in the future. Megatrends give you an idea of where the society could evolve in the medium and long-term. Most of the big organizations publish their analyses of what society could become at many levels: growth, society, technology, etc. (<a href="https://blog.euromonitor.com/2017/09/introduction-top-8-global-megatrends.html">Example</a>)</p><p>I advise you to check them on a yearly basis and to keep an idea of the future environment of your products in the corner of your head.</p><h3>Industrial trends, at the scale of your industry sector</h3><p>To stay connected with your industry, I advise you to register yourself on the industry news websites and newsletter, so that the information comes to you. These websites will keep you informed about the latest trends in your specific industry, along with major evolutions of the competitions. A weekly review of them works fine for me.</p><p>Most of the websites you can register with free information, sell yearly deep insight reports about the industry, and draw current pictures and prognostics. From my experience, such reports can be of great value, as they back up your business model hypothesis with numbers. Furthermore, considering the time it takes to gather such figures, the investment is definitively worth it, if you bargain the report :)</p><p>Finally, fairs are also an excellent source of insight at your industry scale to get fresh insight and exchange information directly with competition about their roadmap, trying their product, etc. Besides, many of these fairs organize roundtables and presentations, giving you the opportunity to complete your big picture, and also, exchanging with the speakers. I advise you to follow the major ones in your industry and clear your agenda ahead of time to be able to do at least one or two per year.</p><h3>Academic Publications</h3><p>Too often, the industrial sector underlooks the academic world. But academic journals regularly publish papers related to your industry that can reveal themselves to be inspirational when they are not directly plug-and-play. For example, at EtixLabs we directly adapted a publication about downsampling to produce a state of the art export function that, as far as I know, none of our competitors offers.</p><p>I advise you to take a subscription to a few academic journals related to your industry (I own a subscription to the <a href="https://onlinelibrary.wiley.com/journal/15405885">JPIM</a>) and to follow a few curators and select the most exciting paper for yourselves. In addition, twice a month, following the section related to your industry on <a href="https://arxiv.org/">arxiv.org</a> can be a valuable source of publications.</p><h3>Consumer insight &amp; competition analysis</h3><p>The fourth source of information: consumers and competitors. Consumers will always tell you what they need. The art of product management is to be able to extract the meaning and read between the lines to find the right features to offer them. This simple point deserves a long article :)</p><p>Competition is also a nice source of innovation. There are three ways to exploit it. The first one: you can follow your competitors’ website and check every time they publish new features and see how the customers are reacting to the features. To do so, you can register yourself for their newsletter. But, marketing analytics tools like <a href="https://www.talkwalker.com/">Talkwalker </a>are excellent to follow the trends and the reactions of customers. Ultimately, you can also follow the patents they publish. Used on a quarterly basis, <a href="https://patents.google.com/">Google Patent</a> is a fantastic tool for that. Even if you cannot copy patents, disclosing the very details of the protected invention is a great source of inspiration.</p><h3>Internal &amp; personal researches</h3><p>Internal and personal researchers are, of course, the fascinating ones for me. From my experience, internal innovation can come from two things:</p><p>First, they are the result of brainstorm workshops like during a Design Sprint or way after when recrawling the museum of ideas that you previously archived. I advise you to save all the ideas produced during design thinking activities. As <a href="https://en.wikipedia.org/wiki/Jeanne_Liedtka">Jeanne M. Liedtka </a>says:</p><blockquote><em>“[…] I said table, not kill, because the idea that doesn’t make it today may be one of tomorrow’s biggest winners. And, we really want to keep an inventory of ideas ready to implement as situations change.” — Jeanne M. Liedtka</em></blockquote><p>Second, it also happens that sometimes, ideas emerge from existing unexpected product capabilities. The best illustration that I know of is my speed sensor in our Tecbak’s eFoosball tables. While I wanted to detect the goal during the game, by playing a bit with my goal sensor, I discovered that I could derivate from the intensity of the signal by the speed of the goalball. This discovery made me able to add a new feature to my product without doing any hardware modification.</p><h3>Conclusion</h3><p>To conclude, all those sources of innovation are excellent but require time and discipline to be followed efficiently. I advise you to strive to automate the tracking as much as possible to receive a regular report by email to save time to deepen the few topics that matter. Keep in mind that the goal of this game is to keep yourself ahead of your competitors by giving you time to experiment before them, so you should, of course, adapt the frequency of pulling pieces of information at the pace of your industry.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c093d950b6c8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>