<?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 Anamitra Musib on Medium]]></title>
        <description><![CDATA[Stories by Anamitra Musib on Medium]]></description>
        <link>https://medium.com/@anamitramusib?source=rss-7f64badeed95------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*AFdbozaBnMBtY1fTHyQh-w.jpeg</url>
            <title>Stories by Anamitra Musib on Medium</title>
            <link>https://medium.com/@anamitramusib?source=rss-7f64badeed95------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 13:58:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anamitramusib/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[Deep Dive on YOLOv2 and YOLO9000]]></title>
            <link>https://medium.com/analytics-vidhya/deep-dive-on-yolov2-and-yolo9000-2eba212dcf8a?source=rss-7f64badeed95------2</link>
            <guid isPermaLink="false">https://medium.com/p/2eba212dcf8a</guid>
            <category><![CDATA[object-detection]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Anamitra Musib]]></dc:creator>
            <pubDate>Tue, 01 Sep 2020 04:12:16 GMT</pubDate>
            <atom:updated>2020-09-01T04:12:16.295Z</atom:updated>
            <content:encoded><![CDATA[<h4>Better, Faster, and Stronger versions of the YOLO object detection model</h4><p>The <strong>YOLO (<em>You Only Look Once)</em></strong><em> </em>object detection model brought out in 2015, was state-of-the-art among similar models of that time, boasting of astounding real-time prediction speeds (<strong><em>Fast YOLO</em></strong> could predict objects at 155 FPS (frames per second)) and an improved accuracy over previous models (YOLO did much better in detecting objects than models such as <em>DPM</em> and <em>Zeiler-Fergus Faster R-CNN</em>, high-performing models of that time).</p><p>While YOLO fared pretty well due to its faster detections and improved accuracies over previous models, it also had some shortcomings.</p><p>Relative to Fast R-CNN, a top-ranking model, YOLO makes <strong>a significant number of localization errors</strong> (8.6% compared to 19%)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/596/1*AU_fYWtiwpRBdhcogpjrAQ.png" /><figcaption>Error analysis of YOLO compared to Fast R-CNN (<a href="https://arxiv.org/pdf/1506.02640.pdf">source</a>)</figcaption></figure><p>Furthermore, it had a <strong>relatively low recall when contrasted with models which used region proposal-based methods</strong> (e.g. <em>Faster R-CNN</em>)</p><p>The creators of YOLO decided to fix these two main problems, while ensuring that the detection accuracy stays the same.</p><h3>Introduction to YOLOv2 and YOLO9000</h3><p>Redmon and Farhadi developed an improved version of YOLO, named <strong><em>YOLOv2</em></strong>, which had several features not seen in YOLO, such as multi-scale training (training on images of different resolutions), thus providing an easy trade-off between speed and accuracy.</p><p>They also proposed a novel method by which they could train the model simultaneously on both object detection and image classification datasets, thus bridging the gap between these two kinds of datasets (we’ll look at this below).</p><p>It takes two to tango, and in the same paper, the authors also talked about another model, <strong><em>YOLO9000</em></strong>, a real-time object detection model which can detect over 9000 object categories!</p><p>Trained jointly on both object detection and classification datasets, YOLO9000 can predict detections for object classes that don’t even have labeled detection data, and this model gives a nice performance on the <em>ImageNet detection task</em>.</p><p>Sounds interesting? Let’s take a deep dive below.</p><h3>Better than YOLO</h3><p>As we looked earlier, YOLO makes a larger number of localization errors compared to Fast R-CNN, and it’s recall is low when matched with region proposal based methods. In YOLOv2, the main concern was fixing these errors while maintaining the detection speed and accuracies.</p><p>The authors utilized a variety of concepts learned from field knowledge as well as some new techniques to result in a model which does better than YOLO. So, what are those techniques?</p><h4>1. Batch Normalization (BN)</h4><p>In simple terms, BN provides every layer of a neural network the capability to learn independently of other layers.</p><p>It normalizes the output of the previous activation layer by subtracting the mean of the batch and dividing by the standard deviation of the batch, &amp; this output gets multiplied by a standard deviation parameter (<em>gamma</em>), then a mean parameter (<em>beta</em>) is added.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/405/1*AdWaQr18d5h5soPS8T7t9w.png" /><figcaption>Batch Normalization process (<a href="https://arxiv.org/pdf/1502.03167v3.pdf">source</a>)</figcaption></figure><p>This process leads to notable changes in convergence and removes the need for other forms of regularization (such as <em>dropout</em>) without overfitting the model.</p><p>✅ When BN is added to all convolutional layers in YOLO, a <br><strong>2% improvement in mAP</strong> (<em>mean average precision</em>)<em> </em>is observed.</p><h4>2. High Resolution Classifier</h4><p>YOLO trained the classification network at 224 x 224 image size and increased this to 448 x 448 for the detection process.<br>This results in the network switching to learning object detection in a short period of time and adjust itself to the new resolution.</p><p>In YOLOv2, the classification network is first trained on ImageNet at <br>224 x 224 resolution, then fine-tuned at the full 448 x 448 resolution for 10 epochs, on the ImageNet dataset, giving the network time to adjust its filters to work better on higher resolution input.<br>The final network is then fine-tuned on detection.</p><p>✅ This process of developing a high-resolution classification network results in <strong>an increase of around 4% mAP</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MYywjgO_GjPdZOq9Ae-RIg.png" /><figcaption>High Resolution Classifier (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g23eff386b7_0_24">source</a>)</figcaption></figure><h4>3. Convolutional with Anchor Boxes</h4><p>YOLO predicts the bounding box coordinates directly using fully connected (FC) layers at the end of the model, on top of all convolutional layers (feature extractors).</p><p><strong><em>Faster R-CNN</em></strong> has a better technique — using hand-picked <em>priors </em>(anchor boxes), it predicts offsets and confidences for these boxes at every location in a feature map. This simplifies the problem and makes it easier for the network to learn.</p><p>We can create 5 anchor boxes as follows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/504/1*kodho4JjG_9zyPAysWns9A.png" /><figcaption>Anchor boxes</figcaption></figure><p>We won’t predict 5 arbitrary bounding boxes, but predict the offsets constrained to the anchor boxes above. Doing this maintains the diversity of the predictions. The training of the model will be more stable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/691/1*G5uRRAl9qdbVY0G_rxxjKg.png" /><figcaption>Maintaining the diversity of predictions (<a href="https://miro.medium.com/max/864/1*CGWTPTY0sfvQoxsS0X6VFg.jpeg">source</a>)</figcaption></figure><p>If instead, we predict arbitrary bounding boxes for each object (like YOLO does), the training will be unstable, as our guesses might work well for some objects but not for others.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/691/1*3gk2YUgbKrq3RztoKCLDJA.png" /><figcaption>Arbitrary bounding box predictions means unstable training (<a href="https://miro.medium.com/max/864/1*7iwTsezrn-tSndx96twprA.jpeg">source</a>)</figcaption></figure><p>Thus, the changes we make to YOLO are listed below —</p><ol><li>Remove the FC layers at the end which predicted the bounding boxes, and use anchor boxes instead</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*HJXc6AWMG1wfwLgkoN5PZg.png" /><figcaption>Removing the FC layers (<a href="https://miro.medium.com/max/875/1*tavjieD0Bum_uX-svYUTKA.jpeg">source</a>)</figcaption></figure><p>2. Use input size of 416 x 416 instead of 448 x 448. Objects (especially large ones) usually occupy the center of the image, so for predicting these objects, a single location at the center would be better than 4 boxes close by. This step means our feature map will have a size of 13 x 13, because YOLO’s convolutional layers downsample images by a factor of 32.</p><p>Thus, an odd number of locations on our feature map would be better than an even number .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*8-xZhTENWk5AIBREEOf50A.png" /><figcaption>Odd number of locations on feature map (<a href="https://miro.medium.com/max/875/1*89qezEeLKJLpD8_fM_H4qQ.jpeg">source</a>)</figcaption></figure><p>3. Decouple the class prediction mechanism from spatial locations and shift it to bounding box level.</p><p>For every bounding box, we now have 25 parameters — 4 for the bounding box, 1 box confidence score (called <em>objectness</em> in the paper), and 20 class probabilities (for the 20 classes in the VOC dataset).</p><p>We have 5 bounding boxes per cell, thus each cell will have 25 x 5, i.e. 125 parameters.<br>Similar to YOLO, the objectness prediction mechanism predicts the IOU (<em>intersection over union</em>) of the ground truth and the proposed box</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*kBCXby1Hn66WQDt1LKikVw.png" /><figcaption>25 parameters per bounding box, 125 parameters per cell (<a href="https://miro.medium.com/max/875/1*UsqjfoW3sLkmyXKQ0Hyo8A.png">source</a>)</figcaption></figure><p>We also remove one pooling layer to make the spatial output of the model <br>13 x 13, compared to 7 x 7 earlier.</p><p>✅Using anchor boxes <strong>decreases the accuracy to 69.2 mAP from 69.5, but results in a significant increase in recall, from 81% to 88%</strong>.</p><h4>4. Dimension Clusters</h4><p>This method suggests running <strong>K-means clustering</strong> on the training set bounding boxes to automatically find good priors, instead of starting with hand-picked anchor box dimensions.</p><p>We want priors that are independent of the bounding box size, so we use the following as our distance metric.</p><pre><strong><em>d(box, centroid) = 1 - IOU(box, centroid)</em></strong></pre><p>The following figures show the result obtained by running K-means for various values of <em>K</em> and plotting the average IOU with the closest centroid. The authors note that K=5 offers a good trade-off between model complexity and high recall.</p><p>Relative centroids are shown for the VOC and COCO datasets. One thing common in both sets of centroids is the preference for tall, thin boxes than short, wide boxes. COCO has greater variation in size than VOC.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/607/1*V3cLKLlLlpuM1IMnsWW-yA.png" /><figcaption>#Clusters vs. Avg. IOU &amp; Relative centroid for VOC and COCO (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><p>✅ The following table compares the average IOU to the closest prior of the clustering strategy and the predefined anchor boxes. <strong>Using K-means to generate the bounding boxes initializes the model with a better representation and makes the task easier to learn</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/404/1*mP0uIKM-71Zsqq_pBXuoAA.png" /><figcaption>Average IOU of boxes to closest priors on VOC 2007 (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><h4>5. Direct Location Prediction</h4><p>A problem that results from directly predicting <em>(x, y)</em> locations of the bounding box is model instability. What region proposal networks do instead is predict values tₓ and tᵧ and the (x, y) coordinates are calculated as follows</p><pre>x = (tₓ * wₐ) - xₐ<br>y = (tᵧ * hₐ) - yₐ</pre><p>Due to the offsets not being constrained here, any anchor box can end up at any point in the image, regardless of what location predicted it — the model takes a long time to stabilize for predicting sensible offsets.</p><p>An alternative is to predict location coordinates relative to the location of the grid cell, just like in YOLO. The network predicts 5 coordinates for every bounding box — tₓ, tᵧ, tཡ , tₕ, and tₒ, and a logistic function is applied to constrain the values of these coordinates to [0, 1]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*HMfUMmsH5d9f7uRGCpG9iw.png" /><figcaption>Network coordinate predictions (<a href="https://miro.medium.com/max/875/1*38-Tdx-wQA7c3TX5hdnwpw.jpeg">source</a>)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*-cj0MmfFv3TBa-CFJgakXA.png" /><figcaption>Blue box is the predicted bounding box, dotted rectangle is the anchor (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><p>This makes the network training more stable.</p><p>✅ The technique of dimension clusters and directly predicting the bounding box center location <strong>improves YOLO by ~5% over the version with anchor boxes</strong>.</p><h4>6. Fine-Grained Features</h4><p>Predicting detections on a 13 x 13 feature map works good for large objects, but for localizing smaller objects, finer-grained features might prove useful.</p><p>Instead of running region proposal networks at various feature maps to many resolutions, like in <em>Faster R-CNN</em> and <em>SSD</em> (Single-Shot Detector), the modified YOLO version adds a <strong><em>passthrough layer</em></strong> which brings features from an earlier layer at 26 x 26 resolution.</p><p>This layer concatenates the higher-resolution features with the low-resolution ones by stacking adjacent features into different channels, turning the <br>26 x 26 x 512 feature map into a 13 x 13 x 2048 one, which is again concatenated with the original 13 x 13 x 1024 features.</p><p>✅ This expanded feature map, having a feature map of 13 x 13 x 3072, provides access to fine-grained features, and <strong>improves the YOLO model <br>by 1%</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*BCv8BFGzzUtj93oZ-M8M7Q.png" /><figcaption>Expanded feature map → fine-grained features (<a href="https://miro.medium.com/max/875/1*RuW-SCIML8SHc5_PrIE9-g.jpeg">source</a>)</figcaption></figure><h4>7. Multi-Scale Training</h4><p>YOLO originally used an input resolution of 448 x 448, which got changed to 416 x 416 by adding anchor boxes. But as the modified network has no FC layers (just convolutional and pooling layers), it can be resized on the fly.</p><p>The network is modified every few iterations — every 10 batches, it randomly chooses a new image dimension size. As the model downsamples by a factor of 32, the network learns the following resolutions, {320 x 320, 352 x 352,… 608 x 608}.</p><p>The fact that the same network can predict detections at various resolutions means we can now fine-tune the sizes for speed and accuracy.</p><ul><li><strong>At low resolutions</strong>, YOLOv2 is a fairly good and accurate detector ideal for smaller GPUs, high framerate video or multiple video streams.<br>At 288 x 288, it runs at 91 FPS and gives <strong>69.0 mAP</strong>, comparable in accuracy to Fast R-CNN (70.0 mAP), but much better than the latter in speed (Fast R-CNN has a speed of just 0.5 FPS) [<em>scores for VOC 2007</em>]</li><li><strong>At high resolutions</strong>, YOLOv2 is a <strong>state-of-the-art detector giving 78.6 mAP</strong> on the VOC2007 dataset, it’s accuracy the highest among the models compared in the paper, while still having a real-time speed of 40 FPS.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/615/1*SczUZGgJN_lNs1cOKoVZVQ.png" /><figcaption>Detection frameworks on VOC 2007 dataset (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/586/1*IyGxJhpilJMXvHMO4Eq8zQ.png" /><figcaption>Accuracy and speed on VOC 2007 (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><h4>Further Experiments</h4><p>On the <strong>VOC 2012 detection dataset</strong>, YOLOv2 obtains <strong>73.4 mAP</strong>, better than the original YOLO, Fast R-CNN, Faster R-CNN, and SSD 300.</p><p>This performance is comparable to ones by <strong>SSD512</strong> (74.9 mAP) and <strong>ResNet</strong> (73.8 mAP), but YOLOv2 is the fastest (2–10 times) among them all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N67y-69DL1KW4umLnkg1QQ.png" /><figcaption>VOC 2012 <em>(test)</em> detection results (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><p>On the <strong>COCO 2015 </strong><strong><em>test-dev</em></strong><strong> dataset</strong>, using the VOC metric, IOU = 0.5, YOLOv2 gets <strong>44.0 mAP</strong>, comparable to the performances of SSD512 and Faster R-CNN.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/934/1*15NJb4AMjo9K0Fho5wlgzw.png" /><figcaption>COCO 2015 (test-dev) detection results (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><p>To summarize, the table below lists the design decisions which mostly led to increases in mAP in YOLOv2, compared to YOLO, except two — <strong>using a fully convolutional network with anchor boxes &amp; using the new network</strong> (Darknet19, we’ll look at this shortly).</p><p>The anchor box approach increased recall from 81% to 88%, but decreased the mAP from 69.5 to 69.2, while the new network reduced computation by 33%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/878/1*wLFGc52PmOu9fsvKgLBUnQ.png" /><figcaption>The path from YOLO to YOLOv2 (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><p>Phew! That were a lot of changes made just to make the model better, next let’s take a peek at how YOLOv2 is faster than YOLO.</p><h3>Faster than YOLO</h3><p><strong>VGG-16</strong> is the classification feature extractor used by most detection architectures (<a href="https://arxiv.org/pdf/1409.1556.pdf">source</a>); though it gives state-of-the-art performance, it’s unnecessarily complex (it uses ~31 billion floating point operations for a single pass over a single image at 224 x 224, thus requiring heavy computing power).</p><p>Based on the <strong>Googlenet architecture</strong>, YOLOv2 uses a custom framework named <strong><em>Darknet19</em></strong> (it has 19 convolutional layers, 5 max pooling layers).</p><p>✅It is better than VGG-16 because —</p><ul><li>It requires just 5.58 billion f.p. operations compared to ~31 billion in VGG-16, so it’s much faster</li><li>Though it uses relatively low f.p. operations, it still scores a higher top-5 accuracy, getting 91.2% on ImageNet, compared to 90% by VGG-16 (YOLO used 8.52 billion f.p. operations and got a top-5 accuracy of 88%)</li><li>They used the then-novel <strong><em>global average pooling</em></strong> to make predictions, compared to max pooling layers in VGG-16</li><li><strong><em>Batch normalization</em></strong> was used for stabilizing training and model regularization, it also made convergence faster</li><li>1 x 1 convolutional filters were used between 3 x 3 layers to compress the features (<a href="https://arxiv.org/pdf/1312.4400.pdf">source</a>)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wSJ6CEuwfVmQVRgrKUIexQ.png" /><figcaption>Darknet19 compared in accuracy, speed and #parameters to other models (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g23eff386b7_0_69">source</a>)</figcaption></figure><p><strong>Global average pooling</strong> layers are used to cut down overfitting by reducing the number of parameters in the network.</p><p>It’s a type of dimensionality reduction where tensors having dimensions of <br><strong><em>(h x w x d) </em></strong>are converted to have a dimension of <strong><em>(1 x 1 x d)</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/788/1*QrfIlcfdxzPszRTj4hSENA.png" /><figcaption>Global Average Pooling (<a href="https://alexisbcook.github.io/2017/global-average-pooling-layers-for-object-localization/">source</a>)</figcaption></figure><p>Given below is the model structure of Darknet19.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/464/1*XGEqb70EKPTo8DjKB9vijQ.png" /><figcaption>Darknet19 model structure (<a href="https://miro.medium.com/max/580/1*8FiQUakp9i4MneU4VXk4Ww.png">source</a>)</figcaption></figure><p>The portion striked through in the image above is replaced with three 3 x 3 conv layers (1024 filters each) followed by a 1 x 1 conv layer with the number of outputs needed for detection, the final output is 7 x 7 x 125.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M6plQ2kQyNYF7-7oUuztzg.png" /><figcaption>YOLOv2 with Darknet19 (<a href="https://media.geeksforgeeks.org/wp-content/uploads/20200401004021/darknet-19-simplified.jpg">source</a>)</figcaption></figure><p>The YOLOv2 paper explains in depth, the hyperparameters (and their values) the authors used for best performance. I won’t go into that detail and bore you, but just know it’s all out there if you’re interested.</p><h3>Stronger than YOLO</h3><p>Finally, let’s look at the new features in YOLOv2 and YOLO9000 which make them better than their previous counterpart.</p><p>One feature specific to object detection datasets when compared to image classification datasets is that, the former have limited labeled examples. Vanilla <strong>detection datasets</strong> contain images in the range 10³-10⁵ with dozens to hundreds of tags. (e.g. the <strong>COCO dataset</strong> has 330K images with just 80 object categories)</p><p><strong>Classification datasets</strong> are huge in that, they contain millions of images with thousands to tens of thousands of classes (e.g. <strong>ImageNet</strong> contains around 14 million images with ~22,000 distinct object classes).</p><p>Because labeling images for detection (predicting bounding boxes, performing non-maximal suppression, etc.) is expensive compared to labeling images for classification (simple <em>tagging</em>), detection datasets are difficult to scale to the level of classification datasets.</p><p>What if we could combine the two types of datasets together and build our network such that it could perform both classification and detection, &amp; even better, detect objects in images which have classes that it hasn’t seen any labeled detection data for?</p><h4>Hierarchical Classification — combining detection and classification datasets</h4><p>Turns out, the <a href="https://arxiv.org/pdf/1612.08242.pdf">paper</a> presented a solution to the very same question we posed above.</p><p>We can combine images from both types of datasets during training, but we have a problem — while detection datasets have simple object labels like <em>“bottle”</em>, <em>“microwave”</em> and <em>“bus”</em>, classification datasets are more diverse and detailed in their classes (I spotted 15 different kinds of snakes <a href="https://gist.github.com/aaronpolhamus/964a4411c0906315deb9f4a3723aac57">here</a> on a cursory glance, like <em>“boa constrictor”</em>, <em>“ringneck snake”</em>, and <em>“green mamba”</em>). So how do we merge labels from both datasets?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UxkHVzWJrH-L-Fy-C8R8cw.png" /><figcaption>How do we merge classes from the two dataset types? (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g23f389e290_1_78">source</a>)</figcaption></figure><p>Typically, we’d use <strong><em>softmax classification</em></strong> but it’s not an option here because the classes are not mutually exclusive. What I mean is, softmax only works when no class is related to each other, thus we can easily have the probability of an object belonging to a class, and all probabilities will <br>add to 1. <br>In our case, <em>“tarantula”</em>, <em>“garden spider”</em> and <em>“wolf spider”</em> are related to each other, as an example.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YwllMG2T-CfwkHEFwMM0_w.png" /><figcaption>Can’t use softmax classification (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g23f389e290_1_46">source</a>)</figcaption></figure><p>Instead, we can start from scratch, forgetting all the knowledge we have about the labels being related to each other.</p><p>ImageNet labels are extracted from <strong><em>WordNet</em>, an online lexical database</strong> that structures concepts and how they relate to each other.</p><p>Structured as a directed graph, different object classes can belong to the same parent class (hyponym), which can then be among various types of objects, etc.</p><p>For example, in WordNet, <em>“laughing owl”</em> and <em>“great grey owl”</em> are both hyponyms of “owl”, which is a type of “bird”, which is a type of “mammal”, etc.</p><p>To build the tree, we follow these steps —</p><ol><li>Observe the visual nouns in ImageNet and follow their paths through the WordNet graph to the root node (<em>“physical object”</em>)</li><li>Because many synsets just have a single path, we add all of them to our tree</li><li>For the concepts left, we iteratively examine them and add those paths that grow the tree by as little as possible (if a concept has 2 paths to the root, one having 5 edges and the other having 3 edges, we’ll choose the latter)</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/999/1*A85uZrmOZ-4irNm9qPs2ZQ.png" /><figcaption>WordNet’s structure can be difficult to grasp (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g23f389e290_1_122">source</a>)</figcaption></figure><p>We thus obtain the <strong>WordTree, a hierarchical model of visual concepts</strong>.</p><p>Instead of performing a softmax on 1000 classes (of ImageNet), the WordTree will have 1000 leaf nodes corresponding to the original labels plus 369 nodes for their parent classes.</p><p>At every node, we predict the probability of each hyponym of a synset <em>given</em> that synset.<br>Say I want to find the probability of the node <em>“terrier”.<br></em>For this, we’ll predict the following probabilities (probability of a <em>Norfolk terrier</em> given that the object is a <em>terrier</em> and so on)</p><pre>Pr(Norfolk terrier | terrier)<br>Pr(Airedale terrier | terrier)<br>Pr(Sealyham terrier | terrier)<br>Pr(Lakeland terrier | terrier)<br>...</pre><p>If we’d like to compute the absolute probability for some node, we follow the path through the tree to the root node and take a product of all conditional probabilities. <br>The follow example is for computing the probability that a picture is of an <em>Airedale terrier</em>.</p><pre>Pr(Airedale terrier) = Pr(Airedale terrier | terrier)<br>                     * Pr(terrier | hunting dog)<br>                     * ....<br>                     * Pr(mammal | animal)<br>                     * Pr(animal | physical object)</pre><p>We always assume that the picture contains an object, so <em>Pr(physical object) = </em>1.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gs4ZvI4aCJPdMbyfoc9mxw.png" /><figcaption>Conditional probabilities in Wordnet (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g1f9fb98e4b_0_147">source</a>)</figcaption></figure><p>For computing the conditional probabilities, Darknet19 predicts a vector of 1369 values and the softmax over all synsets that are hyponyms of the same concept is calculated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/808/1*8_XhvbI9RdQIMQ-iHt2O0Q.png" /><figcaption>Softmax in WordNet (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g1f9fb98e4b_0_103">source</a>)</figcaption></figure><p>As you can see, all hyponyms of the synset <em>“body” (head, hair, vein, mouth</em>, etc.) come under one softmax calculation.</p><p>Thanks to WordNet, DarkNet19 achieves <strong>90.4% top-5 accuracy on ImageNet</strong>, keeping all the parameters same as before.</p><p>Following is a representative image, showing how classes from ImageNet and COCO datasets are merged using the WordNet structure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hozb39-BCOhTLyVv-FvPlQ.png" /><figcaption>Combining datasets using WordNet (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><h4>YOLO9000 — Simultaneous Classification and Detection</h4><p>Now we have a technique to combine object detection and image classification datasets together to get one huge datasets, and the best part is, we can train our model to perform both classification and detection.</p><p>Using the technique above, we combine the COCO dataset (for detection) and the top 9000 classes from the full ImageNet dataset (for classification) — the combined dataset contains <strong>9418 classes</strong>.</p><p>You might observe that the size of ImageNet is very huge compared to COCO, it has multiple orders of magnitude higher number of images and classes.<br>To tackle this unbalanced dataset problem, we oversample COCO so that the ratio of images in ImageNet and COCO is 4:1</p><p>On this dataset we train our <strong>YOLO9000 model</strong>, using the architecture of YOLOv2 but 3 anchor boxes instead of 5 to limit the output size.</p><ul><li><strong>On a detection image</strong>, the network backpropagates loss as usual — <em>objectness</em> (box confidence score, i.e. confidence that the image contains an object), <em>bounding box</em>, and <em>classification error</em></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dwdYwG_xmjxo9tgWre0cRg.png" /><figcaption>Backpropagating loss on a detection image (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g23f389e290_1_128">source</a>)</figcaption></figure><ul><li><strong>On a classification image</strong>, the network backpropagates <em>objectness loss</em> and <em>classification loss</em>.</li></ul><p>For objectness loss, we backpropagate using the assumption that the predicted box overlaps the ground truth label by ≥ 0.3 IOU.</p><p>For classification loss, we find the bounding box that predicts the highest probability for the class and the loss on just its predicted tree is computed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i6Hmz6mfin8Wwyasg9IoKw.png" /><figcaption>Backpropagating loss on a classification image (<a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g1f9fb98e4b_0_40">source</a>)</figcaption></figure><p>YOLO9000 is evaluated on the <a href="http://image-net.org/challenges/LSVRC/2017/"><strong>ImageNet detection task</strong></a>, which has 200 fully labeled categories.</p><p>44 of these 200 categories are there in COCO as well, so the model is expected to do better on classification than detection (because it has lots of classification data thanks to ImageNet, but relatively less detection data).</p><p>Though it hasn’t seen any labeled detection data for the remaining 156 disjoint classes, YOLO9000 still gets <strong>19.7 mAP overall</strong>, 16.0 mAP on those 156 classes.</p><p>This performance is better than the one by DPM (<em>Detectable Parts Models</em>, which use a sliding-window approach to detection), but then YOLO9000 is trained on a combined dataset having both classification and detection images, &amp; it is simultaneously detecting 9000+ object categories.</p><p><strong>All in real time.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T8NJQrW_ZLNgTqYsGunX-g.png" /><figcaption>Accuracy vs. speed, for various architectures and feature extractors (<a href="https://arxiv.org/pdf/1611.10012.pdf">source</a>)</figcaption></figure><p>On ImageNet, YOLO9000 does well in detecting animals but <strong>performs poorly</strong> when it comes to learning classes such as clothing and equipment.</p><p>This can be attributed to the fact that COCO has lots of labeled data for animals, thus YOLO9000 generalizes well from the animals there; but COCO has no bounding box labels for objects such as <em>“swimming trunks”, “rubber eraser”</em>, etc. so our model doesn’t do well on such categories.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/272/1*hns6NaU9XMg1Zrq86KBCvw.png" /><figcaption>YOLO9000 Best and Worst Classes on ImageNet (<a href="https://arxiv.org/pdf/1612.08242.pdf">source</a>)</figcaption></figure><p>Kudos to you if you read till here. Now you are aware of how YOLOv2 and YOLO9000 are <em>better, faster</em> and <em>stronger</em> than the original YOLO model.</p><p>We saw how we can combine object detection and image classification datasets using <strong>WordNet (hierarchical classification)</strong> and get great performance. Using this technique, YOLO9000 can detect over 9000 object categories, and simultaneously perform classification and detection, that too in real time!</p><p>Using multi-scale training (training the model on images of different resolutions), we saw how we can make the YOLOv2 model <strong>more robust</strong> and <strong>make training more stable</strong>.</p><p>In future articles, we’ll see how we can implement YOLOv2 ourselves using the Darknet library, and talk about even better versions of YOLO, like <strong>YOLOv3</strong> and <strong>v4</strong>.</p><p>Hope this article enlightened you about the YOLOv2 and YOLO9000 architectures.<br>Follow me on <a href="https://www.linkedin.com/in/anamitra-musib-69694a63/">LinkedIn</a> and <a href="https://medium.com/@anamitramusib">Medium</a> for upcoming articles :)</p><h4>References</h4><ol><li>BatchNorm (<a href="https://towardsdatascience.com/batch-normalization-in-neural-networks-1ac91516821c">TDS article</a>), (<a href="https://arxiv.org/pdf/1502.03167v3.pdf">research paper</a>)</li><li><a href="https://arxiv.org/pdf/1312.4400.pdf">Network in Network</a> research paper</li><li><a href="https://alexisbcook.github.io/2017/global-average-pooling-layers-for-object-localization/">Global Average Pooling</a></li><li><a href="https://cocodataset.org/#explore">COCO Dataset</a> (80 classes)</li><li><a href="https://wordnetcode.princeton.edu/5papers.pdf">WordNet</a> — An online lexical database</li><li><a href="https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g23f389e290_1_13">YOLO9000 slides</a></li><li><a href="https://arxiv.org/pdf/1506.01497.pdf">Anchor boxes</a> in <em>Faster R-CNN </em>research paper (Section 3.1.1)</li><li>Object detection with YOLO, YOLOv2 and YOLOv3 (<a href="https://medium.com/@jonathan_hui/real-time-object-detection-with-yolo-yolov2-28b1b93e2088">TDS article</a>)</li><li><a href="https://gist.github.com/aaronpolhamus/964a4411c0906315deb9f4a3723aac57">1000 classes of ImageNet</a></li><li><a href="http://wordnetweb.princeton.edu/perl/webwn?o2=&amp;o0=1&amp;o8=1&amp;o1=1&amp;o7=&amp;o5=&amp;o9=&amp;o6=&amp;o3=&amp;o4=&amp;r=1&amp;s=owl&amp;i=1&amp;h=10000#c">WordNet owl example</a></li><li><a href="http://wordnetweb.princeton.edu/perl/webwn?o2=&amp;o0=1&amp;o8=1&amp;o1=1&amp;o7=&amp;o5=&amp;o9=&amp;o6=&amp;o3=&amp;o4=&amp;r=1&amp;s=terrier&amp;i=1&amp;h=100#c">WordNet terrier example</a></li><li><a href="http://image-net.org/challenges/LSVRC/2017/">ImageNet detection task</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2eba212dcf8a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/analytics-vidhya/deep-dive-on-yolov2-and-yolo9000-2eba212dcf8a">Deep Dive on YOLOv2 and YOLO9000</a> was originally published in <a href="https://medium.com/analytics-vidhya">Analytics Vidhya</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Notion is now free, forever!]]></title>
            <link>https://medium.com/@anamitramusib/notion-is-now-free-forever-25027f87dec1?source=rss-7f64badeed95------2</link>
            <guid isPermaLink="false">https://medium.com/p/25027f87dec1</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[creativity]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[life-hacking]]></category>
            <dc:creator><![CDATA[Anamitra Musib]]></dc:creator>
            <pubDate>Wed, 20 May 2020 15:48:13 GMT</pubDate>
            <atom:updated>2020-05-20T15:50:08.333Z</atom:updated>
            <content:encoded><![CDATA[<h4>What this means for creators like you.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/953/1*X3z4KrVm4bIkXWvYqNDKbA.png" /></figure><p><a href="http://notion.so/"><strong>Notion</strong></a> is a recent addition to a horde of online applications aimed at making life easier, and helping people to be more productive. With Notion, there is no limit to how and what you can use your own personal space online for — journaling, habit tracking, to-do lists, maintaining meeting notes, creating budget planning, charting travel ideas and so much more.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*turi3AB_OG5Qgl-29Y-Rbw.jpeg" /><figcaption><a href="https://pbs.twimg.com/media/DetOtbSVMAcQRUU.jpg">Source</a></figcaption></figure><p>I discovered Notion in August last year and continued to use it for a few weeks as my go-to site for journaling, tracking my goals, planning my week ahead <strong><em>until </em></strong>one sunny day I found that I had reached my limit of <strong>1,000 free “blocks”</strong>.</p><p>See, Notion has thing called <strong>“block”</strong> which can be anything really.</p><p>You want to create a new untitled page for your diary entry today? Create a page block.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vULDFxer471Iqzo3OY5qrA.png" /><figcaption><a href="https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Facd9d70e-67f1-44b9-8cf1-9623d17624be%2Fhome2.png?table=block&amp;id=f6c1f8f9-6159-49d4-9ad7-e0f6bb3b7333&amp;width=3070&amp;cache=v2">Source</a></figcaption></figure><p>Wanna set up a database to manage your expenses? Sure, go right ahead.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/876/1*N4dlz5o_YyhuiVSSMkwN_g.png" /><figcaption><a href="https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F612cf6f6-795f-4f4b-8838-04173f46c5df%2Fweekly-agenda-2-4.png?table=block&amp;id=824c8c5a-c848-4dfb-bc50-8b65f5f2a185&amp;width=4000&amp;cache=v2">Source</a></figcaption></figure><p>Feel like creating a new blog but don’t know where to get started? You can even <strong>create your personal blog</strong> using Notion!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/633/1*bfYsk7OeWOfeIRzt1zrV-g.png" /><figcaption><a href="https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fca316019-7753-49d7-a93b-984b1cb316d2%2Fblog-post-3.png?table=block&amp;id=025fc965-1edf-4b3e-a9aa-86fefea9103a&amp;width=4000&amp;cache=v2">Source</a></figcaption></figure><p>There is practically no limit to how creative you can be using Notion. (Check out <a href="https://www.notion.so/personal">this site</a> for awesome ideas.)<br>With the restricting frontier of <strong><em>1,000 blocks now lifted</em></strong>, you can use Notion to be insanely productive, and change your life for the better.</p><p><strong>All at no cost. Everything online. Easily accessible from your browser and phone.</strong></p><p>Let your imagination run wild.</p><p>I’m waiting to hear how you’ll use Notion to your benefit. Share your ideas with me below :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25027f87dec1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gajar ka Halwa]]></title>
            <link>https://medium.com/@anamitramusib/gajar-ka-halwa-e40b1f485ffb?source=rss-7f64badeed95------2</link>
            <guid isPermaLink="false">https://medium.com/p/e40b1f485ffb</guid>
            <category><![CDATA[love]]></category>
            <category><![CDATA[jobs]]></category>
            <category><![CDATA[life]]></category>
            <category><![CDATA[moms]]></category>
            <category><![CDATA[food]]></category>
            <dc:creator><![CDATA[Anamitra Musib]]></dc:creator>
            <pubDate>Sun, 16 Dec 2018 10:11:58 GMT</pubDate>
            <atom:updated>2018-12-16T10:11:58.266Z</atom:updated>
            <content:encoded><![CDATA[<p>Last night had been hard, some might even call it tough, but for a person who’s been waiting all day to reach home and be lost in the comfort of a fluffy and warm blanket, it hardly matters.</p><p>People are true, maybe, 15–16 hour shifts are tough, real tough. But, for Adi, it had become a routine.</p><p>It had to.</p><p>When you become the sole bread-winner in a family of 5, you magically develop high levels of endurance, and motivation (if it even exists). <br>Anyway, quiet now, let him sleep, it’s only a few hours before he faces the world again.</p><p>Morning, sunshine. Ugh, it’s morning this early?<br>He was grumpy, anyone is, after just a 4-hour night. Took a quick shower, freshened up and donned his favorite shirt, anything to make the mood better. But wait, what’s that aroma, and why is it so familiar?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/520/1*PaM53E9bQ4K1EMzxanDHiw.jpeg" /><figcaption><em>Gajar ka Halwa</em> (an Indian dessert)</figcaption></figure><p>It was like flashbacks to when he was 10 years old, and 11, and 12. <br>Every year on that particular day, he remembers inhaling in that fragrance, which tantalized his senses.<br>It takes him somewhere nothing else tangible in the world can, and it hits him close to home.</p><p>Anyway, no time to cater to old memories. Ma was saying something, but he didn’t hear. <br>Obviously. There are things of higher importance to a tired mind.</p><p>Back-to-back meetings, candidate appraisals and an internal audit — all on the same effing day. It’s like God was playing dice in the firmament, and consecutive ones were the only thing his luck could afford. But, it all ended and good heavens did you look at the time?</p><p>It was 10.30 PM.</p><p>Everybody in the office had left, even the lone security guard had thought it a good idea to go out and have a smoke. Just 1 table lamp flickering in the 14-storey office.<br>The radiance was like a drug to his tired eyes — he could see figures dancing when he closed his drooping eyelids.</p><p>Alright, time for Mr. Stomach to demonstrate a whale call.<br>He was ravenous, he had taken care to almost every task he was assigned on that day — except one. To feed himself.<br>But how could he do it now? The canteens were closed, the vending machines were shut down, and he was all alone by himself.</p><p>Call it serendipity or great timing, he now recalled the aroma from the morning. Showering his mom with a thousand blessings all in his mind, he rummaged his satchel to find a tiffin box. He opened it in a jiffy, to find a soaked note inside.<br>It read “Happy 35th, Adi <em>beta</em>”, and that’s when it all came rushing back.</p><p>It was his birthday.</p><p>And he didn’t even remember, but mothers don’t forget, do they?<br>That’s why the smell of sizzling <em>ghee</em> doled on <em>gajar ka halwa</em> seemed so well known.</p><p>Without a care in the world, tired Adi sat on the floor and relished the delicacy till the last morsel, and when he was done, he knew exactly what to do next.</p><p>Drive home and hug Ma for everything good in this world.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/511/1*FBETZgtNMvv0Uzmymkku3Q.gif" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e40b1f485ffb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Curated resources to study Data Science]]></title>
            <link>https://medium.com/good-audience/curated-resources-to-study-data-science-83afe554bd14?source=rss-7f64badeed95------2</link>
            <guid isPermaLink="false">https://medium.com/p/83afe554bd14</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Anamitra Musib]]></dc:creator>
            <pubDate>Sun, 29 Jul 2018 18:06:12 GMT</pubDate>
            <atom:updated>2018-08-01T11:58:37.814Z</atom:updated>
            <content:encoded><![CDATA[<p>Let’s face it — the Internet has opened new doors for learning and taking in knowledge than was possible just a few decades ago. Anyone, from any corner of the world can learn practically anything, thanks to the huge collection of resources, articles, courses and so much more online.<br>All you require for learning is dedication, a huge load of patience and an internet connection.</p><p>Don’t know how to create the perfect resume for your interviews starting next week? No problem. Just Google your queries and you’ll most probably find a helpful link on the first page itself.</p><p>Want to create a poster for your school’s annual day but not sure where to learn designing from? YouTube to the rescue! A huge number of tutorials and recipes are available there at your disposition.</p><p>See what I mean? You can learn almost anything you want with the Internet, <strong><em>you just need where to look</em>.</strong></p><p>Now, this is the difficult part. More than often we’re filled with enthusiasm and zeal when we’re interested to learn something from scratch. So, ideally, you go to Google and search for good sites, or videos, or books etc. which can aid your learning. But the problem is, Google can only show you all the available resources online, not which one is a great place to begin, or which one is ideal for a beginner. More so, you are so confused looking at all the links and tools online that you shut the door to learning what you wanted to altogether.<br><strong>More than 90% of the people lose interest in learning something after they can’t find a site, video or tutorial suitable for their needs</strong>. Now that is <em>a lot </em>of wasted potential if you ask me.</p><p>I have gone through the aforementioned problem and wasted a lot of time in this process, leading to no fruit. <strong>I don’t want anyone who is really interested to learn anything to lose that interest just because they couldn’t decide which path to walk on, or there was no one to lead the way.</strong></p><p>Data Science novices, and those who are new to the field of Data Science — you are in the right place. In this detailed article, I will list some of the sites, online courses, books and other study materials which has helped me advance in this amazing field, and which will prove very helpful to you <em>if </em>you have the passion to learn. So hold on to your seat belts everyone and let this ride begin!</p><p><strong>NOTE —</strong> <br><strong>The lists below on Books, MOOCs &amp; Online Courses, etc. are not exhaustive, or the only material you should refer to. They are just some of the best in my opinion. You are always free to check out other sites and resources on Google.</strong></p><h3>Websites</h3><p>In this section, I will list some of the great websites and links that you should become a regular visitor of, if you are someone who’s very into Data Science, Machine Learning and similar cool stuff.</p><ol><li><a href="https://www.analyticsvidhya.com/"><strong>Analytics Vidhya</strong></a></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/569/1*rEtHWwdXEeUCCsruQg22yw.png" /><figcaption>Analytics Vidhya</figcaption></figure><p>Analytics Vidhya (AV) is a great and highly informative site containing numerous <strong>blogs</strong> on various topics such as machine learning, data visualization, etc., <strong>infographics </strong>— which convey knowledge quickly via beautiful diagrams and graphs, <strong>learning paths</strong> — curated and handcrafted articles guiding you in becoming a good data science programmer, and so much more.<br>They also regularly conduct hackathons and coding competitions, which can be entered by anyone to test their coding and domain-specific mastery.</p><p>2. <a href="https://www.datacamp.com/community"><strong>DataCamp Community</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/372/1*tQpnwSl7EDg7AaSI5mzxNg.png" /><figcaption>DataCamp</figcaption></figure><p>DataCamp is a wonderful site developed for those who want to enhance their grasp and programming skills in the R, Python and SQL languages. I will talk about DataCamp more in detail later.<br>The DataCamp community is a vibrant one, including field experts, data scientists, researchers etc. who regularly publish helpful articles on a variety of topics — NLP, Deep Learning, Bayesian Statistics, etc. to name a few. This site is very helpful if you quickly want to learn about something new and happening, in the field of Data Science.</p><p>3. <a href="https://www.kdnuggets.com/"><strong>KDnuggets</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/663/1*WzCiYUo1jFWYbPFbMg3-fQ.png" /><figcaption>KDnuggets</figcaption></figure><p>This site easily comes in the Top 5 best websites to learn about data science and all cool things related to data, in almost every data science veteran out there. <br>This site has it all — a <strong>News/Blog section</strong> where you can find articles on diverse topics of ML, AI, Reinforcement Learning etc., <strong>Top Stories </strong>— collection of best articles related to data science and publications of every month, <strong>Opinions section</strong> — where experts share their advice and basic know-how of what to do, where and when, <strong>Tutorials</strong>, <strong>Courses</strong>, <strong>Datasets</strong>, <strong>Webinars</strong> and the list goes on. You can afford to miss the daily 10PM movie shown on your television, but not reading and learning from this awesome website.</p><p>4. <a href="https://www.kaggle.com/"><strong>Kaggle</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/429/1*gWd-BwsCnpCbK4-P1batqw.png" /><figcaption>Kaggle — a competition platform for data science lovers</figcaption></figure><p>You probably have heard of Kaggle already, if you are venturing in the data science field. This site is the best one, in my opinion, for anyone who believes in the concept of <em>learning by doing</em>. <br>In the<strong> Datasets</strong> section, you can find data related to almost every field — medical data, financial data, data on demographics etc., the <strong>Kernels</strong> section features interactive notebooks or scripts (in R or Python) referred to as <em>kernels, </em>where people share their work on a particular problem or submit an entry to a coding competition, <strong>Discussions</strong> features articles and blogs describing different topics, walk-throughs and Q-A and finally, in <strong>Learn</strong>, you can take the tailor-made courses for beginners or intermediate learners on R and Python.<br>I learnt a lot of subtle nuances of R and Python from Kaggle Kernels itself, and recommend it to anyone, it’s a great way to learn new methods of doing something you probably didn’t know.</p><p>5. Other honorable mentions</p><p>a) <a href="https://www.topcoder.com/"><strong>Topcoder</strong></a></p><p>b) <a href="https://www.r-bloggers.com/"><strong>R-bloggers</strong></a></p><p>c) <a href="https://flowingdata.com/"><strong>FlowingData</strong></a></p><p>d) <a href="https://www.reddit.com/r/datascience/"><strong>Reddit — DataScience</strong></a></p><p>e) <a href="https://www.codementor.io/"><strong>CodeMentor</strong></a></p><h3>Books</h3><h4>A list of books on R, Python and other topics which I learned a lot from, and promise that you will too.</h4><h4>Mind you, I have provided a direct download link for every book in the lists below.</h4><h4>Books on R</h4><p>1. <a href="https://drive.google.com/file/d/1WKaD5Wp46UsqY1_5n237TeHbrLM9y6iO/view?usp=sharing"><em>R Object-Oriented Programming</em></a>, by Kelly Black</p><p>2. <a href="https://drive.google.com/file/d/1zDTM-0NUJk8n7H1FYrXXg9dx0xQoNxAF/view?usp=sharing">R for Data Science</a>, by Dan Toomey</p><p>3. <a href="https://drive.google.com/file/d/1nQ7kxVKJBVRLiaGY_sL0jAuuiXGBRIaz/view?usp=sharing">Big Data Analytics with R and Hadoop</a>, by Vignesh Prajapati</p><p>4. <a href="https://drive.google.com/file/d/1oMKWwQcAjRVAHgx58o1Sm-wZUIxCMO_a/view?usp=sharing">Building a Recommendation System with R</a>, by Suresh K. Gorakala and Michele Usuelli</p><p>5. <a href="https://drive.google.com/file/d/1sfesQgIeDnjhNNsFC-ynpsf8bkOITfpH/view?usp=sharing">Data Analysis with R</a>, by Tony Fischetti</p><p>6. <a href="https://drive.google.com/file/d/1B-No_d8vPoyFo4ZJW3FkH7Jp7xtFCnlk/view?usp=sharing">Data Manipulation with R</a>, by Jaynal Abedin and Kishor Kumar Das</p><p>7. <a href="https://drive.google.com/file/d/1yEB05tc5TbSP67qCl7ewU5nF4XUo2a85/view?usp=sharing">ggplot2 Essentials</a>, by Donato Teutonico</p><p>8. <a href="https://drive.google.com/file/d/1LYwv1iCAnhqlk8UoLg5LBP6vAxyk7-I4/view?usp=sharing">Introduction to R for Quantitative Finance</a>, by Puhle, Berlinger et al.</p><p>9. <a href="https://drive.google.com/file/d/14krYlRGec3xfQgOZuMmPnmDp1fRbpt4A/view?usp=sharing">Learning Bayesian Models with R</a>, by Dr. Hari M. Koduvely</p><p>10. <a href="https://drive.google.com/file/d/14EVwg5_ZDym2RIcgoPnqaeOAXICpppNS/view?usp=sharing">Learning Predictive Analytics with R</a>, by Eric Mayor</p><p>11. <a href="https://drive.google.com/file/d/1nTFsAGxX1D8OUfRKagwIs1887-l-iXIY/view?usp=sharing">Learning Probabilistic Graphical Models in R</a>, by David Bellot</p><p>12. <a href="https://drive.google.com/file/d/1eIGyd9-b1otwlPvrUWTDOnGXHbIQjPo7/view?usp=sharing">Learning R for Geospatial Analysis</a>, by Michael Dorman</p><p>13.<a href="https://drive.google.com/file/d/1uD9dDsiCX9zH-R2FdNkIBBUoqguxloIB/view?usp=sharing"> Machine Learning with R Cookbook (2nd ed.)</a>, by Ashish Singh Bhatia, Yu-Wei and David Chiu</p><p>14. <a href="https://drive.google.com/file/d/1rINIQv7LCNFcxi03wNZ8ym0nxe5oy1Nr/view?usp=sharing">Mastering Data Analysis with R</a>, by Gergely Daróczi</p><p>15. <a href="https://drive.google.com/file/d/1887yM-R-qqxCMpJ_V-I1hPPaVGMhlaaq/view?usp=sharing">Mastering Predictive Analytics with R</a>, by Rui Miguel Forte</p><p>16. <a href="https://drive.google.com/file/d/1StmytVgle3wDVPRXTy-xsBFIYimW0a9x/view?usp=sharing">Mastering R for Quantitative Finance</a>, by Tuza, Berlinger et al.</p><p>17. <a href="https://drive.google.com/file/d/1L3TZcaTTksQuKxN0xYtSb8LT3ffyaJ3J/view?usp=sharing">Mastering Scientific Computing with R</a>, by Paul Gerrard and Radia M. Johnson</p><p>18. <a href="https://drive.google.com/file/d/13SJYo35JCbqX0y3XAlqAYLlP0epToe6Q/view?usp=sharing">Mastering Social Media Mining with R</a>, by Sharan Kumar Ravindran and Vikram Garg</p><p>19. <a href="https://drive.google.com/file/d/1NXMzkI4SG6E8wpGLTGNpiB2CxbyzeYb6/view?usp=sharing">R Data Mining Blueprints</a>, by Pradeepta Mishra</p><p>20. <a href="https://drive.google.com/file/d/1SDqMz0evLTNi2JmnX3D1OTcwv9MsW0Eu/view?usp=sharing">R Data Visualization Cookbook</a>, by Atmajitsinh Gohil</p><p>21. <a href="https://drive.google.com/file/d/1o5na8aqpm1sVx8oeA-cX2R6gT63H4DlD/view?usp=sharing">R Deep Learning Essentials</a>, by Dr. Joshua F. Wiley</p><p>22. <a href="https://drive.google.com/file/d/1jaT2B79rwbc0hqizHcXG6kdP_4ob6yBy/view?usp=sharing">R Graph Essentials</a>, by David Alexander Lillis</p><p>23. <a href="https://drive.google.com/file/d/1wksW09uMrPVuAcRdkr7NbOdpY2oPIkPK/view?usp=sharing">R Graphs Cookbook (2nd ed.)</a>, by Jaynal Abedin and Hrishi V. Mittal</p><p>24. <a href="https://drive.google.com/file/d/1lT6gU-KsBonhbPsZRDm3PmF7A8ferhBl/view?usp=sharing">R Machine Learning Essentials</a>, by Michael Usuelli</p><p>25. <a href="https://drive.google.com/file/d/1-YXRkuTXAiGhFbO7KA5-t3ullZuaDGC_/view?usp=sharing">R Statistical Application Development by Example</a>, by Prabhanjan Narayanachar Tattar</p><p>26. <a href="https://drive.google.com/file/d/15cw3eGHF7UxtyiDBCjocjPaqfSryL4To/view?usp=sharing">RStudio for R Statistical Computing Cookbook</a>, by Andrea Cirillo</p><p>27. <a href="https://drive.google.com/file/d/1axF0fmu4IhgYA60YY-W3r_oP5JgQIpBt/view?usp=sharing">Statistical Analysis with R, Beginner’s Guide</a>, by John M. Quick</p><p>28. <a href="https://drive.google.com/file/d/1x3cRSMvN_pl3lDbt4drDRMsxYdV6n47v/view?usp=sharing">Unsupervised Learning with R</a>, by Erik Rodriguez Pacheco</p><p>29. <a href="https://drive.google.com/file/d/1-A4OGcoT1tKjmlETAwv8n74X-6roR5Ej/view?usp=sharing">Advanced R — Data Programming and the Cloud</a>, by Matt Wiley and Joshua F. Wiley</p><p>30.<a href="https://drive.google.com/file/d/1MzEY-9Rq_Hl-znlqkk4PLDVtD13zk2FL/view?usp=sharing"> An Introduction to Spatial Data Analysis and Visualisation in R</a>, by Guy Lansley and James Cheshire</p><p>31. <a href="https://drive.google.com/file/d/1_qrm0XRIUMxpqKxzqR6SdmweR8MIAdQH/view?usp=sharing">Beginning R — The Statistical Programming Language</a>, by Dr. Mark Gardener</p><p>32. <a href="https://drive.google.com/file/d/1gCaRsY9-X3OnKacwedI22GBi4Uq7k0uY/view?usp=sharing">Data Mining and Predictive Analytics</a>, by Daniel T. Larose and Chantal D. Larose</p><p>33. <a href="https://drive.google.com/file/d/1wsG3hdI5bfGvH2uPgC0-DGaaeGrqjQbs/view?usp=sharing">Data Visualisation with R — 100 Examples</a>, by Thomas Rahlf</p><p>34. <a href="https://drive.google.com/file/d/1HrKM5_nwFkrGHJSGPXatnRxo7WdKgiYJ/view?usp=sharing">Deep Learning with R</a>, by François Chollet and J. J. Allaire</p><p>35. <a href="https://drive.google.com/file/d/1BFEWrn7EKYziMNyM2IsFrYYNy790q_YW/view?usp=sharing">Efficient R Programming</a>, by Colin Gillespie and Robin Lovelace</p><p>36. <a href="https://drive.google.com/file/d/1ym7BbAWFPtYiq3d5gi8GLPOlC2BuOLVV/view?usp=sharing">Hands-On Programming with R</a>, by Garrett Grolemund</p><p>37. <a href="https://drive.google.com/file/d/1O9OzfacFeRli8xruhfvV4PGxFdHrD_ws/view?usp=sharing">Introduction to Deep Learning using R</a>, by Taweh Beysolow II</p><p>38.<a href="https://drive.google.com/file/d/1FH6AFfy_GKDj9FUel_L7h1FIvJevNooA/view?usp=sharing"> An Introduction to R</a>, by W. N. Venables, D. M. Smith and the R Core Team</p><p>39. <a href="https://drive.google.com/file/d/1_mfLU3Q5jZBX0Re7W9nw3Zo_FG4fYrtb/view?usp=sharing">Springer Texts in Statistics (6th printing)</a>, by G. Casella, S. Fienberg and I. Olkin</p><p>40. <a href="https://drive.google.com/file/d/1lfScKFtsbLoDvz4veAj0UoTpFFgNSIu6/view?usp=sharing">Machine Learning for Hackers</a>, by Drew Conway and John Myles White</p><p>41. <a href="https://drive.google.com/file/d/1IkrJxY1p_FJeXIYoZ1Pu-MdbsShr5Azk/view?usp=sharing">Practical Data Science with R</a>, by Nina Zumel and John Mount</p><p>42. <a href="https://drive.google.com/file/d/1L1cGIBBiDA__0aXlqtgAC_Z66R1OW4p8/view?usp=sharing">Practical Graph Mining with R</a>, by Samatova, Hendrix et al.</p><p>43. <a href="https://drive.google.com/file/d/1Ea6V3OeFCU82KephuZQtUfic2mPPTmhe/view?usp=sharing">Practical Statistics for Data Scientists</a>, by Peter Bruce and Andrew Bruce</p><p>44. <a href="https://drive.google.com/file/d/1KV-9AQYInPMPqm3CS4GF8gOTb7kQp8-x/view?usp=sharing">R Cookbook</a>, by Paul Teetor</p><p>45. <a href="https://drive.google.com/file/d/1JZLaNBCvu1tNdcAOCR0oGlLzgPlp8Gb-/view?usp=sharing">R for Data Science</a>, by Hadley Wickham and Garrett Grolemund</p><p>46. <a href="https://drive.google.com/file/d/1Tv2B_jw5NxmU0upt8BFlSD5ugNu9KHNv/view?usp=sharing">R for Programmers — Mastering the Tools</a>, by Dan Zhang</p><p>47. <a href="https://drive.google.com/file/d/1DDz3U6xaFOrSmt5T28E3cdkVjava4JLr/view?usp=sharing">R Graphics Cookbook</a>, by Winston Chang</p><p>48. <a href="https://drive.google.com/file/d/1M23KrlINl0lm6Nkn8d6heBKldYyNt-EZ/view?usp=sharing">R in Action — Data Analysis and graphics with R</a>, by Robert I. Kabacoff</p><p>49. <a href="https://drive.google.com/file/d/14UXqVUvwpoWi2rGOAa-Ucq9s_dKI_8nK/view?usp=sharing">R Packages</a>, by Hadley Wickham</p><p>50. <a href="https://drive.google.com/file/d/1J4uBFmsrg00CvAPD9mYsSkpK098-ztBu/view?usp=sharing">R Projects for Dummies</a>, by Joseph Schmuller</p><p>51. <a href="https://drive.google.com/file/d/1JpLTzqAQ7LOkziJsMaahPtwGjkzrSf9X/view?usp=sharing">R and Data Mining: Examples and Case Studies</a>, by Yanchang Zhao</p><p>52. <a href="https://drive.google.com/file/d/1Qwh0Vn1tZudKPYbjbBcaU-8qLyCZB4nd/view?usp=sharing">Text Mining with R — A Tidy Approach</a>, by Julia Silge and David Robinson</p><p>53. <a href="https://drive.google.com/file/d/1zF91-DKB_qQV_kc0-Uz6wJMComvAcYPl/view?usp=sharing">The Art of R Programming</a>, by Norman Matloff</p><p>54. <a href="https://drive.google.com/file/d/1Wvd97LCWr0kQVX4Spg67sq4tIffxsZjF/view?usp=sharing">The Book of R</a>, by Tilman M. Davies</p><p>55. <a href="https://drive.google.com/file/d/1CauGyJwCSo3LnroRgVDWsnb-ZVY-hm8s/view?usp=sharing">The R Book</a>, by Michael J. Crawley</p><h4>Books on Python</h4><p>1. <a href="https://drive.google.com/file/d/1AwGrMVV1keP0n4SIrBaAwMsc8HF4Q3q8/view?usp=sharing">Building Machine Learning Systems with Python</a>, by Luis Pedro Coelho and Willi Richert</p><p>2. <a href="https://drive.google.com/file/d/1lWsVauF85dTVjdJ0zDnKgR6ikenWmF-P/view?usp=sharing">Deep Learning for Natural Language Processing</a>, by Palash Goyal, Sumit Pandey and Karan Jain</p><p>3. <a href="https://drive.google.com/file/d/1uPTznQxWX7eqBvvgKdEQ8KjVEZTvcEao/view?usp=sharing">Deep Learning with Python</a>, by François Chollet</p><p>4. <a href="https://drive.google.com/file/d/1U2L4sEp-T4-m3uI2g5hxabS6hutPnmIG/view?usp=sharing">Fundamentals of Deep Learning</a>, by Nikhil Buduma</p><p>5. <a href="https://drive.google.com/file/d/1lv7KKcFf1MsUE71Qt7FGP27sowrO_dFK/view?usp=sharing">Hands-On Machine Learning with Scikit-Learn and TensorFlow</a>, by Aurélien Géron</p><p>6. <a href="https://drive.google.com/file/d/1ln8cCKHzdOsr0-KdxE5D7oPDwpkcFbzO/view?usp=sharing">Introduction to Machine Learning with Python</a>, by Andreas C. Mueller and Sarah Guido</p><p>7. <a href="https://drive.google.com/file/d/1uDgExJnJLe3UmE4cc6qobjFhsEwU9Q-D/view?usp=sharing">Learning TensorFlow</a>, by Tom Hope, Yehezkel S. Resheff and Itay Lieder</p><p>8. <a href="https://drive.google.com/file/d/1XgaDa2PvMAUVPR4TrOeqqghDuMXKnGPJ/view?usp=sharing">Machine Learning with TensorFlow</a>, by Nishant Shukla</p><p>9. <a href="https://drive.google.com/file/d/160mvPTZlKsczT_nDHUWp1NEszq-pTx9M/view?usp=sharing">Mastering Machine Learning with Python in Six Steps</a>, by Manohar Swamynathan</p><p>10. <a href="https://drive.google.com/file/d/1StoMc4l56JQNCPtGD1mncjHSH-Ifzc4Y/view?usp=sharing">Mastering Machine Learning with scikit-learn</a>, by Gavin Hackeling</p><p>11. <a href="https://drive.google.com/file/d/1yAgbIjBsml1pBlJMxoscRmI3IKVTef2D/view?usp=sharing">Programming Collective Intelligence</a>, by Toby Segaran</p><p>12. <a href="https://drive.google.com/file/d/1h5QWKhcvOVFUKtXsdChaUjLOlou3g-vQ/view?usp=sharing">Python Data Science Handbook</a>, by Jake VanderPlas</p><p>13. <a href="https://drive.google.com/file/d/1RJtFqm4icmG3T-X3PasEwJCD87EjHyzd/view?usp=sharing">Python Deep Learning Cookbook</a>, by Indra den Bakker</p><p>14.<a href="https://drive.google.com/file/d/1nufsA4i8bNoKxD7FlcPmWtCjm-APU1W7/view?usp=sharing"> Python for Data Analysis</a>, by Wes McKinney</p><p>15. <a href="https://drive.google.com/file/d/1uj0qHVVpeTwE9ELVFe2ky1eSnSJ8suFY/view?usp=sharing">scikit-learn Cookbook</a>, by Julian Avila, Trent Hauck</p><h4>Books on other Data Science topics</h4><p>1. <a href="https://drive.google.com/file/d/18KmXscVFmDJZx1ekZPSoaCKqUD8iJVuq/view?usp=sharing">Social Media Mining — An Introduction</a>, by Reza Zafarani, Mohammad Ali Abbasi and Huan Liu</p><p>2. <a href="https://drive.google.com/file/d/1u9HHyhGEa1c07jCzHZ7Nz115FRIl_2gD/view?usp=sharing">The LION Way — Machine Learning plus Intelligent Optimization</a>, by Roberto Battiti and Mauro Brunato</p><h3>Online Courses and MOOCs</h3><p>Online courses and Massive Open Online Courses (MOOCs) have risen in popularity just recently as they are a great medium by which one can accomplish 2 things:-<br>1) <strong>Gain impressive knowledge</strong> about a specific domain or field, including both theory and programming skills<br>2) <strong>Obtain certifications</strong> in some chosen topics which helps them stand out from the rest, and by showcasing these certificates on professional networking sites like LinkedIn, they can let recruiters know that they have proven expertise in a field, which again, differentiates them easily from the enormous crowd and ultimately, may also help them to get a job.</p><p>1.<strong> </strong><a href="https://www.coursera.org/learn/machine-learning"><strong>Coursera — Machine Learning, by Stanford University</strong></a><strong>, taught by Andrew Ng.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/355/1*rDOWnlFOvl6nxTyWlOsTZQ.png" /></figure><p>This course has become the most famous and widely known MOOC in the fields of Data Science and Machine Learning. Taught by <strong>Andrew Ng</strong>, an industry expert (he’s worked in Google Brain, Baidu and teaches as an adjunct professor at Stanford University) and highly knowledgeable person, this <strong>11 week course</strong> is assured to provide you with extensive theoretical knowledge and programming skills, which is priceless nowadays in these fields.</p><p>2.<strong> </strong><a href="https://www.coursera.org/specializations/deep-learning"><strong>Coursera — Deep Learning Specialization</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/411/1*sqk3U6tN7HWtKNVDwM-9Dw.png" /></figure><p>Coming second after the Machine Learning MOOC taught by Andrew Ng, this 5 course specialization, again taught by Andrew Ng, Younes Mourri and Kian Katanforoosh is a promising investment of time. Throughout the courses, one gains knowledge of concepts including Neural Networks, Shallow and Deep Neural Networks, Hyperparameter Tuning, Convolutional and Recurrent Neural Networks and NLP. After completing this course, anyone can call themselves to be ‘very well versed in Deep Learning and AI’.</p><p>3. <a href="https://www.udemy.com/machinelearning/"><strong>Udemy — Machine Learning A-Z™: Hands-On Python &amp; R In Data Science</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/611/1*uTg4DUbGzXwXVryYgP5n2Q.png" /></figure><p>If you want to learn the concepts of Machine Learning from scratch, also gaining programming skills in both R and Python at the same time, this course is just the right one for you.<br>Kirill Eremenko, Hadelin de Ponteves and the SuperDataScience Team come together to impart the theoretical workings of ML algorithms, and also their codes in R and Python, in this 41 hour course consisting of around 285 videos.</p><p>4. <a href="https://www.datacamp.com/"><strong>DataCamp</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*zTCdAHMd1evwGVrjAb8TiA.png" /></figure><p><em>Learning by doing</em> is one of the core ideas behind the running of Datacamp. This site contains around <strong>150 courses on Python, R and SQL</strong> each dealing with diverse concepts such as Machine Learning, Deep Learning, data visualization, NLP etc. There are also specialized <strong>‘<em>skill tracks</em>’</strong> and <strong>‘<em>career tracks</em>’</strong> designed to enhance one particular concept among many others, some being <em>Python Programmer Career Track, Machine Learning in R Skill Track</em> etc.<br>I would recommend every novice in the Data Science field to get a paid subscription of Datacamp and make use of its features such as <strong><em>Projects, Tutorials and Open Courses</em></strong>, it’s absolutely worth it.</p><p>5. <a href="https://www.dataquest.io/"><strong>Dataquest</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/187/1*H7XAzDJ5aDl3Oxwke2vJlg.png" /></figure><p>Similarly to DataCamp, Dataquest is another great site to learn the basics of R and Python and then gradually working your way up to advanced concepts like Machine Learning and Deep Learning. <br>The courses here are reputed to be very detailed and informative, with a lot of information being conveyed through a single course itself. As of now, only <strong>5 courses are available for free</strong>, the rest can be accessed through a <strong>paid subscription</strong> which again, like DataCamp, is worth it, not kidding.</p><h3>Other Resources</h3><p>1. <a href="https://www.linkedin.com/"><strong>LinkedIn</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*WHjZeRs4d_Ax4bBmXihaOw.png" /></figure><p>LinkedIn is currently the<strong> most popular and used professional networking service online</strong>. It’s a great place to gain knowledge specific to your fields of interest and also at the same time, to connect with other people who share the same interests and interact with them.</p><p>Some people you should definitely follow on LinkedIn (assuming you’re interested in the Data Science domain) are —</p><ol><li><a href="https://www.linkedin.com/in/randylaosat">Randy Lao</a></li><li><a href="https://mx.linkedin.com/in/faviovazquez">Favio Vázquez</a></li><li><a href="https://www.linkedin.com/in/elizabethjcomstock">Beth Comstock</a></li><li><a href="https://www.linkedin.com/in/akshaybahadur21">Akshay Bahadur</a></li><li><a href="https://www.linkedin.com/in/sirajraval">Siraj Raval</a></li><li><a href="https://www.linkedin.com/in/lillianpierson">Lillian Pierson</a></li><li><a href="https://www.linkedin.com/in/kate-strachnyi-data">Kate Strachnyi</a></li><li><a href="https://www.linkedin.com/in/tarrysingh/">Tarry Singh</a></li><li><a href="https://www.linkedin.com/in/snooravi">Sarah N.</a></li><li><a href="https://in.linkedin.com/in/jaink">Kunal Jain</a></li></ol><p>2. <a href="https://medium.com/"><strong>Medium</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/176/1*Pm_ek8_5ZXAVozAZcvDHyQ.png" /></figure><p>For Medium, I can only say one thing right now — <strong>Thank God, I discovered this early in my life.</strong><br>Medium’s a terrific site to read great, amazing articles ranging various topics such as Life, Art, Music, Data Science, etc. In my opinion, Medium will quickly (or has it already?) cross the popularity and outreach of other knowledge imparting sites like Quora.<br>I recently (a week ago) discovered Medium and instantly fell in love with it. I suggest you to check it out if you haven’t yet, and increase your technical and reading skills manyfold, very quickly.</p><p>As I have only just discovered Medium, some of the publications and people here (related to the field of Data Science) that have caught my eye are —</p><ol><li><a href="https://medium.com/@faviovazquez">Favio Vázquez</a></li><li><a href="https://medium.com/towards-data-science/data-science/home">Towards Data Science</a></li></ol><p>So people, that was pretty much it from my side. I hope I have provided you with enough helpful information. The resources and material I provided should be more than enough to get you started, the rest depends on how much effort you put and how passionate in this field you are.</p><p>I wish you the very best. :)</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fupscri.be%2F7b0181%3Fas_embed%3Dtrue&amp;dntp=1&amp;url=https%3A%2F%2Fupscri.be%2F7b0181%2F&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=upscri" width="800" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/748a06e9019942f2677bba917ef55804/href">https://medium.com/media/748a06e9019942f2677bba917ef55804/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=83afe554bd14" width="1" height="1" alt=""><hr><p><a href="https://medium.com/good-audience/curated-resources-to-study-data-science-83afe554bd14">Curated resources to study Data Science</a> was originally published in <a href="https://medium.com/good-audience">Good Audience</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>