<?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 Ankit Nayan on Medium]]></title>
        <description><![CDATA[Stories by Ankit Nayan on Medium]]></description>
        <link>https://medium.com/@ankitnayan?source=rss-99989d9afc6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*O7mC5wiExx53OIJsRQTXZg.jpeg</url>
            <title>Stories by Ankit Nayan on Medium</title>
            <link>https://medium.com/@ankitnayan?source=rss-99989d9afc6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 17:40:18 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ankitnayan/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[How Deep Learning can help you post right content on Social Media?]]></title>
            <link>https://medium.com/@ankitnayan/deep-learning-can-help-you-post-right-content-on-social-media-53d63fd210e?source=rss-99989d9afc6------2</link>
            <guid isPermaLink="false">https://medium.com/p/53d63fd210e</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Ankit Nayan]]></dc:creator>
            <pubDate>Wed, 20 Sep 2017 14:47:46 GMT</pubDate>
            <atom:updated>2017-09-20T17:59:48.621Z</atom:updated>
            <content:encoded><![CDATA[<p>In a pilot with one of the US based companies, who were managing the social media profile of companies in automotive industry, we laid out a goal to identify which types of images shared in social media gave more engagement. The different image variations to be identified were:</p><ul><li>Types of vehicle — Eg. concept car, vintage car, racing car, bike, truck, etc</li><li>Different viewpoints of the vehicle — front view, side view &amp; back view</li><li>Different parts of the vehicle — Eg. Focus on wheels, engine, interiors, etc</li><li>Background scene of the vehicle, like whether the vehicle was in street, racing track, in mountain, in forest, with sea in background, etc</li></ul><p>Here are a few sample images:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*GaXmu8vhXj3cKHfNMDXc4w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*wkeepk-QgvJ2b6c_6P4KhA.jpeg" /></figure><p><strong>What can Deep Learning do here?</strong></p><p>Deep Learning, in the field of computer vision, has been extensively used for object recognition. The ImageNet project is a large visual database designed for use in visual object recognition software research (<a href="http://www.image-net.org/challenges/LSVRC/">link</a>) where deep learning and its variants are used extensively to achieve state-of-art results.</p><p>We used deep learning models to identify types of vehicles with their viewpoints and also to identify background scenes. Now we have attributes assigned to each image which can be used to predict which types of images were liked more by users and produced better engagement.</p><p><strong>Building the model</strong></p><p>Basic understanding to build model for object classification can be found at paper<em> “ImageNet Classification with Deep Convolutional Neural Networks” by Krizhevsky &amp; Hinton</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5Y2juz1nu0VHw-k3WBBGNQ.png" /><figcaption>Architecture of CNN in the paper</figcaption></figure><p>Similar but less deeper architecture was adopted to train model for data from automobiles industry. Below are the layers used in deep CNN:</p><ul><li><strong>Convolution </strong>— convolves the input image with a set of learnable filters, each producing one feature map in the output image</li><li><strong>ReLU (Rectified Linear Unit) — </strong>The most common model of a neuron’s output is Sigmoid Function, i.e., f(x) = 1/(1+exp(-x)). Replacing it with f(x) = max(0, x) helps neural network train much faster. Also, since they are already normalized, there is no input normalization needed.</li><li><strong>Max Pooling — </strong>It is commonly used in CNNs to reduce the size of the input feature space and thus reduce the model complexity by reducing the number of parameters. This also leads to some sort of high level feature capturing from image and thus adding to robustness of the model.</li><li><strong>LRN (Local Response Normalization) — </strong>It does what it says, i.e., it normalizes in local input regions.</li><li><strong>Inner Product </strong>— this layer is called FC ( Fully Connected ) layer. It’s output is in the form of a single vector.</li></ul><h4><strong>Techniques used to avoid overfitting of data:</strong></h4><ul><li><strong>Data Augmentation</strong> — The training set is expanded by generating image translations and horizontal reflections. Thus for each label many more images are generated adding to robustness of the model.</li><li><strong>Dropout</strong> — In this method, the output from randomly selected neurons are ignored in both forward pass and backward pass of neural network. Thus any kind of local dependency of neurons is removed, making it learn more robust features. Usually 50% of neurons are ignored in each dropout layer.</li></ul><h3><strong>Results (</strong><a href="http://www.spotr.in/context/">website</a>)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*GaXmu8vhXj3cKHfNMDXc4w.jpeg" /></figure><p><strong>Attributes assigned: </strong>Car, Rear View</p><p><strong>Background Scene detected: </strong>Waterbody, Sea</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*wkeepk-QgvJ2b6c_6P4KhA.jpeg" /></figure><p><strong>Attributes assigned: </strong>Car, Vintage car, Front View</p><p><strong>Background Scene detected: </strong>Buildings, structures and road</p><p><strong><em>References</em></strong></p><ul><li>ImageNet Classification with Deep Convolutional Neural Networks by Kirzhevsky, Sutskever and Hinton</li><li>Imagenet Large Scale Visual Recognition Challenge (ILSVRC) [<a href="http://www.image-net.org/challenges/LSVRC/">Link</a>]</li><li>Learn more about different models that are trained on ILSVRC 2012 dataset [<a href="http://caffe.berkeleyvision.org/model_zoo.html">Link</a>]</li><li>Learn more about the different CNN layers in Caffe in [<a href="http://caffe.berkeleyvision.org/tutorial/layers.html">Link</a>]</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=53d63fd210e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>