<?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 Ujjawal Poudel on Medium]]></title>
        <description><![CDATA[Stories by Ujjawal Poudel on Medium]]></description>
        <link>https://medium.com/@ujjawalpoudel?source=rss-6f9583b7544b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*-Nhz60O_VtPrvbDWxL_0rA.jpeg</url>
            <title>Stories by Ujjawal Poudel on Medium</title>
            <link>https://medium.com/@ujjawalpoudel?source=rss-6f9583b7544b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 08:38:23 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ujjawalpoudel/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 Netflix Handles Sudden Load Spikes in the Cloud? (What I Learned)]]></title>
            <link>https://medium.com/@ujjawalpoudel/how-netflix-handles-sudden-load-spikes-in-the-cloud-what-i-learned-f058a4d80af3?source=rss-6f9583b7544b------2</link>
            <guid isPermaLink="false">https://medium.com/p/f058a4d80af3</guid>
            <category><![CDATA[netflix-engineering]]></category>
            <category><![CDATA[load-balancing]]></category>
            <category><![CDATA[cloud-architecture]]></category>
            <category><![CDATA[scalability]]></category>
            <category><![CDATA[aws-reinvent-2024]]></category>
            <dc:creator><![CDATA[Ujjawal Poudel]]></dc:creator>
            <pubDate>Mon, 07 Apr 2025 00:22:46 GMT</pubDate>
            <atom:updated>2025-04-07T00:22:46.149Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QI24ey8ijn013jSS5wKAlA.png" /><figcaption>AWS re:Invent 2024 — How Netflix handles sudden load spikes in the cloud (NFX301)</figcaption></figure><p>📺 <strong>Video:</strong> <a href="https://www.youtube.com/watch?v=TkFyZyxFRBM&amp;ab_channel=AWSEvents">Watch here</a><br>📑 <strong>Slides:</strong> <a href="https://reinvent.awsevents.com/content/dam/reinvent/2024/slides/nfx/NFX301_How-Netflix-handles-sudden-load-spikes-in-the-cloud.pdf">View here</a></p><p>As someone passionate about scalable backend systems, I recently watched this really insightful talk from AWS re:Invent 2024 — “How Netflix Handles Sudden Load Spikes in the Cloud.” It was a fascinating deep dive into how a company like Netflix, with massive global traffic, keeps everything running smoothly when traffic suddenly surges. Here’s what I learned and understood from the session in my own words.</p><h3>1. The Challenge of Global Scale</h3><p>Netflix operates on an <strong>active-active architecture</strong>, meaning any node in their system can both read and write data. This design ensures <strong>regional failover</strong> — if one region fails, another one can take over seamlessly. The key goal? <strong>Zero interruption for users.</strong> No matter where the traffic is coming from, users get a smooth experience.</p><p>Since users are scattered across time zones, traffic on Netflix follows a <strong>sinusoidal daily pattern</strong>. For example, when it’s peak hours in North America, it might be off-hours in Asia-Pacific. This rolling wave of usage makes traffic forecasting and load handling even more complex.</p><h3>2. Handling Traffic Surges</h3><p>Netflix tracks traffic using a metric called <strong>Starts Per Second (SPS)</strong> — the number of times users start a video per second. Predictably, SPS rises during evenings and weekends. To tackle this surge, Netflix uses two main buffers:</p><ul><li>✅ <strong>Success Buffer</strong> — to absorb increased demand under normal load</li><li>🛑 <strong>Failover Buffer</strong> — to handle traffic when things go wrong</li></ul><h3>Strategies They Use:</h3><h4>🔥 Pre-Warming (a.k.a. “Pre-heating”)</h4><p>If Netflix is launching a new season of a popular show (like <em>Squid Game</em>), they don’t wait for the traffic spike to happen. They <strong>predict traffic patterns</strong> and <strong>scale up the servers in advance</strong> to stay ready.</p><h4>🌍 Intelligent Traffic Shaping</h4><p>Rather than always routing users to the geographically nearest region, they sometimes <strong>re-route users to less-loaded regions</strong>, balancing the load globally. It’s a smart way to avoid overloading hotspots and offer uninterrupted service.</p><h3>3. Why Autoscaling Alone Isn’t Enough</h3><p>Autoscaling sounds great in theory — more traffic means more servers spin up. But in real life? It takes time:</p><ul><li>4 minutes to detect the spike</li><li>6 minutes to boot new servers</li><li>2 more minutes for another detection round</li></ul><p>That’s <strong>10+ minutes of delay</strong>, and Netflix can’t afford that during a sudden traffic flood.</p><p>So they did <strong>synthetic load testing</strong> — simulating traffic spikes to test how the system behaves in high-pressure moments. From that, they identified four main bottlenecks:</p><ul><li><strong>Detection lag</strong></li><li><strong>App startup time</strong></li><li><strong>System startup</strong></li><li><strong>Hardware provisioning</strong></li></ul><h3>4. How Netflix Fixed It</h3><h3>✅ Faster Detection: From CPU to RPS</h3><p>Previously, Netflix scaled based on <strong>CPU usage</strong>, but they realized it wasn’t reliable. For example, a 2x spike in RPS and a 10x spike both pushed CPU to 100%, giving no clear picture of the <em>actual</em> demand.</p><p>So they switched to <strong>RPS-based scaling</strong>, which is a much more accurate indicator of user activity. To support this:</p><ul><li>They <strong>reduced CloudWatch metric intervals</strong> from 5 minutes to 1 minute.</li><li>They built <strong>Atlas</strong>, a custom metrics tool that reports every <strong>5 seconds</strong>. This saved detection time down <strong>3x</strong>!</li></ul><h3>⚙️ Parallel System Startup</h3><p>Instead of starting services one after another (A → B → C), they now <strong>start everything in parallel</strong> and then merge. This cut system startup time by half.</p><h3>⏱️ Time to Recovery: Under 3 Minutes!</h3><p>By optimizing detection and startup, they brought the <strong>total time to recover (TTR)</strong> down to <strong>less than 3 minutes</strong> — a huge win. But what about <strong>during</strong> those 3 minutes?</p><h3>5. Keeping the System Alive During the TTR</h3><p>This part really impressed me. During those critical 3 minutes, Netflix prioritizes services using the following tags:</p><ul><li><strong>Critical</strong> — Video playback must work</li><li><strong>Degraded</strong> — Reduced personalization</li><li><strong>Bulk</strong> — Batch jobs can wait</li><li><strong>Best Effort</strong> — Nice-to-have stuff only if resources allow</li></ul><p>So, if a region hits the <strong>success buffer limit</strong>, non-essential APIs (like personalization) are paused. This prevents <strong>congestive failure</strong> — where everything just freezes, like traffic stuck on a jammed highway.</p><p>They also <strong>re-route traffic across regions</strong> with fully replicated Cassandra databases. If one region is overloaded, another can step in and handle only <strong>critical requests</strong>. If it can’t, it drops the rest — <strong>better to degrade gracefully than crash hard.</strong></p><h3>Final Thoughts</h3><p>This talk was an eye-opener. As someone interested in backend systems and real-world scalability challenges, I learned so much from Netflix’s approach. It’s not just about having more servers — it’s about <strong>smart prediction, fast detection, regional flexibility, and graceful degradation</strong>.</p><p>📚 If you’re into cloud architecture, I definitely recommend watching the full session.<br>Again, here’s the link: <a href="https://www.youtube.com/watch?v=TkFyZyxFRBM">https://www.youtube.com/watch?v=TkFyZyxFRBM</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f058a4d80af3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pagination in MongoDB]]></title>
            <link>https://medium.com/@ujjawalpoudel/pagination-in-mongodb-1c20596d5298?source=rss-6f9583b7544b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c20596d5298</guid>
            <category><![CDATA[pagination]]></category>
            <category><![CDATA[mongodb]]></category>
            <category><![CDATA[mongodb-tutorial]]></category>
            <dc:creator><![CDATA[Ujjawal Poudel]]></dc:creator>
            <pubDate>Fri, 14 Mar 2025 20:54:07 GMT</pubDate>
            <atom:updated>2025-03-14T20:54:07.713Z</atom:updated>
            <content:encoded><![CDATA[<p>MongoDB supports two main types of pagination: <strong>Offset-based pagination</strong> and <strong>Cursor-based pagination</strong>.</p><h4>Offset-Based Pagination</h4><p>In MongoDB, offset-based pagination is typically implemented using the skip() and limit() methods. The skip()method scans records up to the specified point, while limit() defines the number of records to be retrieved.</p><p>For example:</p><ul><li>If skip = 5 and limit = 6, the database cursor will skip the first five records and return rows 6 to 11 (both inclusive), totaling six records.</li></ul><p><strong>Limitations:</strong><br>The major drawback of this approach is performance degradation with large datasets. If a high skip value is used (e.g., skip = 100,000), MongoDB must scan all rows up to 100,000 before retrieving the required records. This makes the query inefficient and slow, making offset-based pagination impractical for large-scale applications.</p><h4>Cursor-Based Pagination</h4><p>Cursor-based pagination addresses the inefficiencies of offset-based pagination by leveraging MongoDB’s <strong>ObjectId</strong> as a pointer to the documents in a collection.</p><p><strong>Key Concepts:</strong></p><ul><li>This approach does not support direct jumps (e.g., from page 1 to page 10). It only allows navigating forward (“Next”) and backward (“Previous”).</li><li>When the “Next” button is pressed, the application retrieves the <strong>ObjectId</strong> of the last record in the current result set and queries for records where ObjectId &gt; last_record_id, using the limit() method to fetch the next set of results.</li><li>To paginate backward, the <strong>first ObjectId</strong> from the current dataset is used to retrieve previous records.</li><li>Since querying by <strong>ObjectId</strong> is an <strong>O(1) operation</strong>, cursor-based pagination is significantly faster than offset-based pagination.</li></ul><p><strong>Considerations:</strong></p><ul><li>If multiple records are inserted within the same second, <strong>ObjectId</strong> values may be very close, potentially causing issues with pagination.</li><li>Cursor-based pagination is ideal when timestamps are carefully managed and when all records are inserted from a single system. However, in distributed systems where system clocks may vary, ObjectId values may not always follow a strict order.</li></ul><h3>References</h3><ul><li><a href="https://www.mongodb.com/docs/manual/reference/method/cursor.skip/#using-skip--">skip() Method</a></li><li><a href="https://www.mongodb.com/docs/manual/reference/bson-types/#objectid">ObjectId</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c20596d5298" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vector Modeling and Text Pre Processing]]></title>
            <link>https://medium.com/@ujjawalpoudel/vector-modeling-and-text-pre-processing-af744372cb6e?source=rss-6f9583b7544b------2</link>
            <guid isPermaLink="false">https://medium.com/p/af744372cb6e</guid>
            <category><![CDATA[tokenization]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[stemming]]></category>
            <category><![CDATA[lemmatization]]></category>
            <category><![CDATA[vectorization]]></category>
            <dc:creator><![CDATA[Ujjawal Poudel]]></dc:creator>
            <pubDate>Sat, 27 May 2023 18:12:29 GMT</pubDate>
            <atom:updated>2023-05-27T18:12:29.553Z</atom:updated>
            <content:encoded><![CDATA[<p>Natural Language Processing (NLP) refers to the technology that enables machines to comprehend and respond to human language in a manner that is indistinguishable from human understanding, thus passing the Turing Test. Various applications exemplify the capabilities of NLP, such as email spam detection, grammar correction in Grammarly, and voice assistance provided by Google Assistant.</p><h3><strong>Terms used in NLP:</strong></h3><ol><li><strong>Corpus:-</strong> In NLP, a corpus refers to a collection of words or tokens centered around a specific topic or case. This corpus serves as the dataset used for training machine learning models. One can liken a corpus to a data frame employed in linear regression to draw a parallel, as both represent essential data structures for analysis and modeling.</li><li><strong>N-Gram:-</strong> In natural language processing, N-Gram refers to a sequence of N consecutive items, such as words, subwords, or characters. Depending on the value of N, different terms are used to describe these sequences. When N is equal to one, it is called a unigram. N equal to two is referred to as a bigram; for N equal to three, it is known as a trigram. N-Grams find applications in various language processing tasks. For instance, the word2vec algorithm utilizes N-Grams, particularly bigrams or “skip grams,” to construct a neural network to analyze language patterns. Additionally, Markov models rely on bigram probabilities to make predictions and generate coherent sequences of words or characters. By considering the frequency and relationships between consecutive items, N-Grams contribute to the understanding and generation of text in natural language processing.</li></ol><p>In machine learning, most operations are conducted through mathematical computations. However, we encounter speech or text data that cannot be directly processed as numerical values in natural language processing. Consequently, it becomes necessary to convert this textual information into a numerical representation. This article aims to elucidate the approaches to converting alphabetic characters into numerical values suitable for machine learning algorithms.</p><h3><strong>The Vector in NLP:</strong></h3><p>In mathematical terms, a vector can be defined as a quantity representing direction and magnitude. It serves as a fundamental unit in mathematics and physics. For instance, when considering velocity, we can observe that it encompasses both the speed of a moving object and its direction. Thus, velocity is an example of a vector quantity as it encapsulates magnitude and direction within its representation.</p><p>In machine learning, particularly in high-dimensional spaces, visualizing the direction of a vector can be challenging or even impossible. Therefore, instead of focusing on the direction directly, we often rely on the scalar value of the vector, which represents its magnitude. To capture the direction information, we can express it in terms of radians or angles. For example, if a vector is defined as (3,4) in a two-dimensional graph, its equivalent representation would include a magnitude of 5 and an angle of 0.93 radians. The first representation is based on the Cartesian plane, while the latter is expressed using polar coordinates, where the angle provides insight into the direction of the vector.</p><h3><strong>Bag of Words:</strong></h3><p>Text is composed of combinations of words. When we arrange the words correctly, they can convey meaningful information. For instance, consider the sentence, “I love to learn the mathematics behind different machine learning algorithms.” Randomizing this sentence makes it meaningless and fails to convey the intended information accurately. Surprisingly, many NLP approaches overlook the importance of textual sequentiality in information generation. However, a concept called the “bag of words” addresses this issue. It involves converting words into numerical representations and analyzing those numerical values to derive meaning.</p><p>Using the bag of words concept can sometimes result in losing the information within a sentence. Consider the example of the phrase “dog toy,” which typically refers to a toy intended for dogs to play with or enjoy. However, if we change the order of the words to “toy dog,” it now refers to a toy that resembles a dog (an inanimate object, not a natural dog) and is meant for children to play with. This demonstrates that relying solely on the bag-of-words approach may only sometimes be helpful for accurate analysis.</p><p>The bag of words concept is primarily utilized in vector models and classic machine learning algorithms. However, it is less commonly used in probabilistic models and deep learning techniques. Although this represents a general categorization of its usage, there are instances where the bag of words approach can still be applied in probabilistic and neural models.</p><h3><strong>Count Vectorizer:</strong></h3><p>A count vectorizer is a straightforward method for converting text into vectors. It is based on the concept of the bag of words approach.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vbar25Rf_tsDUeznMVCQDg.png" /><figcaption>Figure:- A example of a Count Vectorizer</figcaption></figure><p>This figure illustrates how a count vectorizer converts a document or data into a vector in NLP. It serves as a simple example to explain the concept of count vectorizer.</p><p>The following steps are taken to obtain the vector representation:<br><strong>Step 1:</strong> Identify the unique words present in the document. In this example, there are six unique words: “and,” “cats,” “eggs,” “hate,” “I,” and “like.”</p><p><strong>Step 2:</strong> Count the occurrences of the unique words within each sentence. This counting is then represented in a tabular format, making it easy to visualize and create a vector representation based on the counts.</p><p>Count vectorizer aids in segmenting or clustering documents into different groups. By counting the occurrences of specific words in the document, we can cluster them based on the frequency of certain unique words. This clustering can be represented graphically, where hyperplanes can separate the points corresponding to different documents. For instance, if we attempt to separate documents into Biology and Physics categories, we can cluster them based on the highest occurrence of specific words. The document is clustered as biology if the word “blood” has the highest frequency. Conversely, the document is clustered as physics if “force” has the highest frequency. Count vectorizer allows for identifying and clustering new documents based on this approach, providing a means to assign them to appropriate clusters.</p><h3><strong>Tokenization:</strong></h3><p>In simple terms, tokenization in NLP refers to splitting a sentence into individual words or tokens. While using the split(), function in Python can split a sentence by whitespace, proper tokenization also requires considering punctuation. For instance, “I hate cats.” and “I hate cats?” convey different meanings. The first example with a period indicates a statement, whereas the second with a question mark suggests a question. Therefore, tokenization in NLP involves splitting the sentence while considering all possibilities and variations in the punctuation.</p><p>However, it is worth noting that CountVectorizer in the sklearn library completely disregards punctuation during tokenization. While considering punctuation in sentence splitting is a good idea, it is essential to consider that doing so will increase the dimensionality of our data.</p><h3>Stopwords:</h3><p>First, let us understand what stop words are in the context of NLP. Stop words are commonly occurring words in a language that often do not carry significant meaning in a specific scenario. Examples of stop words include prepositions, articles, and common pronouns. In many sentences, we frequently encounter words like “a,” “an,” “the,” “in,” “on,” “that,” and “this.”</p><p>There are several reasons to remove stop words:</p><ol><li><strong>Limited Information:</strong> Stop words generally do not provide essential information in model building. Removing them allows the model to focus on more meaningful words that can contribute to accurate predictions or analyses.</li><li><strong>Dimensionality and Computational Cost:</strong> Including stop words in the model increases its dimensionality, resulting in higher computational costs. More dimensions mean more computations and greater consumption of time and memory resources.</li><li><strong>Distortion of Similar Words:</strong> Stop words can interfere with accurately representing similar words. In tasks like clustering, frequent occurrence of certain stop words can cause sentences or words to be grouped, overshadowing the text’s true meaning. For example, words like “a,” “and,” or “the” can be clustered together, leading to potential confusion or loss of context.</li></ol><p>By removing stop words, we can enhance the model’s ability to focus on meaningful and informative words, leading to more efficient processing and improved accuracy in NLP tasks.</p><h3>Stemming and Lemmatization:</h3><p>Stemming and lemmatization are techniques used to obtain the root forms of words. Stemming involves reducing words to their base or root form, while lemmatization focuses on transforming words to their canonical or dictionary form. For instance, the root form of “walking” is “walk,” and the root form of “ate” is “eat.”</p><p>So, why do we need these techniques? When we consider each word in its specific form (e.g., past, present, future tense), the dimensional space of our vectors increases. This means that the system treats words like “walk” and “walking” as separate entities, even though they convey the same meaning in real-world contexts. Stemming helps by truncating words to their root forms, while lemmatization ensures that we obtain the root forms of words even when they appear in different grammatical forms. For example, lemmatization would transform both “ate” and “eat” to their common root form, “eat.”</p><p>By applying stemming and lemmatization, we can reduce the complexity and dimensionality of our data while preserving the semantic meaning of words. This allows for more efficient processing and improved accuracy in various natural language processing tasks.</p><h3>Reference:</h3><p>Lazy Programmer Team. (2023). Machine Learning: Natural Language Processing in Python (V2). In Udemy. <a href="https://www.udemy.com/course/natural-language-processing-in-python/learn/lecture/28889570?start=180#overview">https://www.udemy.com/course/natural-language-processing-in-python/learn/lecture/28889570?start=180#overview</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=af744372cb6e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Market Dominance and Strategies: A Comparative Analysis of Qualcomm and Apple in the Mobile…]]></title>
            <link>https://medium.com/@ujjawalpoudel/market-dominance-and-strategies-a-comparative-analysis-of-qualcomm-and-apple-in-the-mobile-b5ae7a9b54e6?source=rss-6f9583b7544b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b5ae7a9b54e6</guid>
            <dc:creator><![CDATA[Ujjawal Poudel]]></dc:creator>
            <pubDate>Thu, 18 May 2023 23:19:30 GMT</pubDate>
            <atom:updated>2023-05-18T23:21:09.158Z</atom:updated>
            <content:encoded><![CDATA[<h3>Market Dominance and Strategies: A Comparative Analysis of Qualcomm and Apple in the Mobile Chipset Industry</h3><h3>Summary:</h3><p>The mobile chipset market is dominated by Qualcomm, which has revenues nearly double those of Apple. Due to its commercial strategies, market positioning, and partnerships with essential smartphone manufacturers, Qualcomm has managed to preserve its market share despite Apple’s superior chipset performance. Apple, in contrast, only provides chipsets for its products, and its market share is far lower. The analysis suggests that Qualcomm increase its corporate partnerships and alliances, increase its R&amp;D spending, and create specialized chipsets for specialized applications. It is advised that Apple use technology like machine learning algorithms to grow its client base and create unique value propositions for its chipsets.</p><h3>Introduction:</h3><p>Qualcomm has become a dominating participant in the chipset industry in the world of mobile technology, even though its products perform less well than those of Apple (How Is Qualcomm Dominating the Mobile Semiconductor Industry? — VYRIAN, 2022). This begs the question of what variables have helped Qualcomm maintain its success despite Apple’s superior results. The causes for Qualcomm’s dominance, including its business plans, market positioning, and collaborations with significant smartphone producers, will be examined in this thesis. By investigating these characteristics, we may better understand why Qualcomm has held onto its position as a market leader in the mobile chipset market.</p><p>The target audience for this analysis would be experts in the mobile and technology industries and tech enthusiasts curious about the market competition and performance differences between Qualcomm and Apple mobile chipsets. Consumers and tech enthusiasts who want to learn how various mobile processors perform and the elements that affect that performance may find this information interesting.</p><h3>Background:</h3><p>Qualcomm is a well-known American multinational headquartered in San Diego, California. It is renowned for its skill in creating semiconductors, software, and services connected to wireless technology, according to Qualcomm (n.d.). The company, founded in 1985 by Irwin M. Jacobs and six other people, had a fundamental impact on the development of CDMA, a wireless communications technology that went on sale in 1995. Qualcomm has established a reputation as a top producer of mobile chipsets and a supplier of wireless communications technologies.</p><p>A technology business called Apple was founded in 1976 by Steve Jobs, Steve Wozniak, and Ronald Wayne. Its headquarters are in Cupertino, California (Brashares, 2001, pp. 15–16). The Samsung S3C6400 chipset, based on ARM, provided power to the first iPhone when it was introduced in 2007 (The Evolution of the Apple iPhone and its multiple CPUs — Even within a Model, 2010). The A16 Bionic, Apple’s most recent mobile processor, is intended to provide a better performance, faster processing speeds, and longer battery life than earlier versions (iPhone 14 pro, n.d.).</p><h3>Supporting Evidence:</h3><p>The graphical representation below shows the market share of Smartphone AP (chipset) among the various companies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*PYtYohPYBf70ST5PzaeoCQ.png" /><figcaption>Figure 1: Global Smartphone AP Market Share by Shipments (Infographic: Q2–2022 | Global Smartphone AP Market Share, 2022)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*MN-YRWyUnlOTmP3jJw2H3g.png" /><figcaption>Figure 2: Global Smartphone AP Market Share by Revenue (Infographic: Q2–2022 | Global Smartphone AP Market Share, 2022)</figcaption></figure><h3>Explanation and Analysis:</h3><p>Qualcomm and MediaTek are the most prominent players in this sector. Our research shows that Qualcomm has close to double Apple’s market share.</p><p>Revenue generated by Qualcomm is also close to twice that of Apple. Throughout the quarter, Qualcomm gained 29% of the market. Despite difficult macroeconomic conditions and the contracting smartphone industry, Qualcomm retained its dominance in the premium class. (Counterpoint et. el.)</p><p>In terms of sales, Qualcomm led the AP market with 44% in Q2 2022. The higher premium mix, which has increased ASPs, has propelled Qualcomm’s share growth of 56% YoY(Year Over Year) in Q2 2022. Also, the earnings for the premium tier will be supported by the recently announced cooperation between Qualcomm and Samsung for the galaxy s22 series. In the AP market, MediaTek now controls 22% of the market. Baseband that is sold to Apple generates additional income for Qualcomm. The increased 5G ASP and entry into the premium market with the Dimensity 9000 series were the main drivers of MediaTek’s revenue. (Counterpoint et. el.)</p><h3>Conclusion:</h3><p>In conclusion, Qualcomm now has a larger market share than Apple, even though both companies are significant competitors in the mobile chipset industry. This is likely owing to Qualcomm’s more expansive product offerings that appeal to a broader range of devices, including inexpensive to high-end smartphones. At the same time, Apple exclusively supplies chipsets for its products. In addition, Qualcomm’s chipsets are typically less expensive than Apple’s high-end A-series chipsets. Also, Qualcomm has made a name for itself as a pioneer in 5G technology, which is becoming increasingly significant in the mobile market. Overall, even though both businesses are renowned for their innovation and high-performance chipsets, Qualcomm likely has a more significant market share in the industry for mobile chipsets due to its more expansive product line-up and lower price point.</p><h3>Recommendations:</h3><h3>For Qualcomm:</h3><p>1. To stay at the forefront of cutting-edge technology like 6G and AI, continue to invest in research and development.</p><p>2. Develop specialized chipsets for niche applications like gaming or artificial intelligence to broaden its product line.</p><p>3. Develop closer ties with smartphone makers to increase the number of devices that feature their chipsets.</p><p>4. Emphasising transparency and moral conduct in its commercial dealings and relationships will help the organization improve its reputation in the marketplace.</p><h3>For Apple:</h3><p>1. Increase the number of customers by making its chipsets available to other smartphone makers and its own.</p><p>2. Enhance its pricing strategy by lowering its chipset’s cost or providing cheaper solutions for low-cost devices.</p><p>3. Develop distinctive value propositions for its chipsets using its technology, such as machine learning algorithms.</p><p>4. Look for ways to enter new industries where its chipsets could be incorporated, such as those for smart home appliances or automotive technologies.</p><h3>References:</h3><p>Brashares, A. (2001). Steve Jobs: thinks different (pp. 15–16). <em>Twenty-First Century Books</em>.</p><p>Cheng, R. (2022, June 10). Apple’s Chip Success Has Qualcomm Fired Up to Show Who’s Boss. CNET. <a href="https://www.cnet.com/tech/mobile/apples-chip-success-has-qualcomm-fired-up-to-show-whos-boss/">https://www.cnet.com/tech/mobile/apples-chip-success-has-qualcomm-fired-up-to-show-whos-boss/</a></p><p>Company. (n.d.). Qualcomm. <a href="https://www.qualcomm.com/company">https://www.qualcomm.com/company</a></p><p>How is Qualcomm dominating the mobile semiconductor industry? — VYRIAN. (2022, March 16). <em>Vyrian — Your Primary Source of Semiconductors, Connectors and LEDs</em>. <a href="https://www.vyrian.com/qualcomm-dominating-mobile-semiconductor-industry/#:~:text=Qualcomm%20is%20leading%20the%20mobile">https://www.vyrian.com/qualcomm-dominating-mobile-semiconductor-industry/#:~:text=Qualcomm%20is%20leading%20the%20mobile</a></p><p>Infographic: Q2–2022 | Global Smartphone AP Market Share. (2022, September 21). <em>Counterpoint Research</em>. <a href="https://www.counterpointresearch.com/infographic-global-smartphone-ap-market-share-q2-2022/">https://www.counterpointresearch.com/infographic-global-smartphone-ap-market-share-q2-2022/</a></p><p>Parashar, S., &amp; Sharma, P. (2022, December 12). Global Smartphone AP-SOC Shipment &amp; Forecast tracker by Model — Q3 2022. Counter Point. <a href="https://www.counterpointresearch.com/global-smartphone-ap-market-share/">https://www.counterpointresearch.com/global-smartphone-ap-market-share/</a></p><p>Sims, G. (2022, June 5). Why are Apple’s chips faster than Qualcomm’s? Android Authority. <a href="https://www.androidauthority.com/why-are-apples-chips-faster-than-qualcomms-gary-explains-802738/">https://www.androidauthority.com/why-are-apples-chips-faster-than-qualcomms-gary-explains-802738/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b5ae7a9b54e6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“InceptionNet”]]></title>
            <link>https://medium.com/@ujjawalpoudel/inceptionnet-c6b5aa40eccf?source=rss-6f9583b7544b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6b5aa40eccf</guid>
            <category><![CDATA[convolution-neural-net]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[inception]]></category>
            <dc:creator><![CDATA[Ujjawal Poudel]]></dc:creator>
            <pubDate>Tue, 17 Dec 2019 09:24:19 GMT</pubDate>
            <atom:updated>2019-12-17T09:24:19.040Z</atom:updated>
            <content:encoded><![CDATA[<p>The evolution in <strong>Convolutional Neural Network</strong> had started in 1989 in a modest network. The first paperwork on modern Convolutional Neural Networks (CNNs) was inspired by Yann LeCun et al. in the 1990s in their paper “<strong>Gradient-Based Learning Applied to Document Recognition</strong>”. It consisted of a layered network composed of convolutional and subsampling (i.e. Pooling layers) operations followed by fully connected layers and ultimately classifier layers for classifying handwritten digits.</p><p>After a gradual increase in research of Convolutional Neural Network, many competitions were organized for the development of CNNs, <strong>ImageNet</strong> competition is also one of them. ImageNet has been running an annual competition in visual recognition (<strong>ILSVRC</strong> — <strong>ImageNet Large Scale Visual Recognition Challenge</strong>) from 2010 onwards, where participants are provided with 1.4 million images. Below are some popular CNNs architectures that won in ILSVRC competition :</p><ul><li><strong>LeNet-5</strong></li><li><strong>AlexNet</strong></li><li><strong>VGGNet</strong></li><li><strong>GoogLeNet</strong></li><li><strong>ResNet</strong></li></ul><p>First of all, let’s be clear about the Convolution Parameter. Convolution layer accepts a volume of size W*H*D and requires four hyperparameters as follows:</p><p>Number of filters:- K</p><p>Spatial Extent:- Fw, Fh (Filter width, Filter height)</p><p>Stride:- Sw, Sh (Stride width, Stride height)</p><p>Padding:- P</p><p>To calculate output of Convolution, the formula is as follows</p><p>Output width :- ((W-Fw+2P)/Sw)+1</p><p>Output height :- ((H-Fh+2P)/Sh)+1</p><p>To calculate pooling layers, the formula is as follows</p><p>Om :- ((Im+2P-F)/S)+1</p><p>Om:- Output Matrix</p><p>I’m:- Input Matrix</p><p>F:- Filter</p><p>S:- Stride</p><p>Let us review the architecture of conventional neural networks up to <strong>GoogLeNet</strong>. A typical design pattern for Convolutional Neural Networks is as follows:</p><ul><li>Stack of Convolution Layers with kernel sizes as high of 7*7 and 11*11 and strides such as 1 or 2.</li><li>After that max-pooling layer is implemented.</li><li>The last layer is fully connected layers with an output layer, which include <strong>Cost Function</strong> calculation, <strong>Accuracy Matrix</strong> calculation and prediction output using <strong>Softmax</strong>, <strong>Binary</strong>, etc function.</li><li>In fully connected layers, the dropout function is implemented to reduce the number of connections and overcome problems of overfitting.</li></ul><p>Now let’s try to understand GoogLeNet architecture, GoogLeNet is the winner of the “<strong>ImageNet Large Scale Visual Recognition Challenge</strong>” in the year 2014, used 22 layers with parameters which are the micro-network embedded within a macro network. GoogLeNet architecture is highly based on the approach of <strong>Network-in-Network</strong> proposed by Lin et al. to increase the computational power of neural networks. In GoogLeNet architecture, having parallel convolutions with different filters, followed by concatenation, captures different features at 1×1, 3×3 and 5×5, after that ‘clustering’ them. Convolutional Computation is done by addition of 1 × 1 convolutional layers followed by the rectified linear activation, which is used mainly as dimension reduction function to remove computational bottlenecks and it finally solved the problem (That is, in a deep neural network, if two convolutional layers are chained, any uniform increase in the number of their filters results in a quadratic increase of computation).</p><p>Let’s clarify <strong>why was a 1*1*1 convolution used?</strong></p><ul><li><strong>Without the use of 1×1 convolution</strong>:-</li></ul><p>Let’s consider that we have a 14*14*480 input matrix which is convoluted with a filter size of 5*5*48 and finally gives an output matrix of 14*14*48. After all, we have 112.9 Million ((14×14×48)×(5×5×480)) number of operations.</p><ul><li><strong>With the use of 1×1 convolution</strong>:-</li></ul><p>Let’s consider that we have a 14*14*480 input matrix which is convoluted with a filter size of 1*1*16 and gives an output matrix of 14*14*16, which is again convoluted with a filter size of 5*5*48 and finally gives output matrix of 14*14*48. After all, we have 1.5 Million ((14×14×16)×(1×1×480)) number of operations in the first convolution and 3.8 Million ((14×14×48)×(5×5×16)) number of operations in the second convolution. So that we have a total of 5.3 Million (1.5 Million + 3.8 Million) which is much much smaller than 112.9 Million. Using the dimension reduced inception module, a neural network architecture was built. This is popularly known as GoogLeNet (<strong>Inception v1</strong>). GoogLeNet has 9 inception modules stacked linearly. It is 22 layers deep (27 including the pooling layers). When an image’s coming in, different sizes of convolutions, as well as max-pooling, are tried. Then different kinds of features are extracted. It uses global average pooling at the end of the last inception module and in fully connected architectures, the optimal network topology can be constructed layer by layer by analyzing the correlation statistics of the activations of the last layer and clustering neurons with highly correlated outputs. After that, all feature maps at different paths are concatenated together as the input of the next module.</p><p><strong>Inception v2</strong></p><p>The authors proposed some upgrades over Inception v1 which increased the accuracy and reduced the computational complexity. Some of the important improvement in Inception v2 as compared to Inception v1 is as follows:-</p><ul><li>In deep learning, there is one universally valid concept that neural networks perform better when convolutions didn’t alter the dimensions of the input drastically but in Inception v1 there is a drastic drop down of dimensions that may cause loss of information for further prediction in that given module. That’s why Inception v2 reduce representational bottleneck.</li><li>Using smart factorization methods, that is 5x5 convolution is converted to two 3x3 convolution operations to improve computational speed. As a 5x5 convolution is 2.78 times more expensive than a 3x3 convolution. So stacking two 3x3 convolutions leads to a boost in the performance of the model.</li><li>In Inception v1, the module was made deeper instead of making wider, there would be an excessive reduction in dimensions, and hence loss of information. So that this problem is solved by using wider instead of deeper to remove the representational bottleneck.</li></ul><p><strong>Inception v3</strong></p><p>The authors noted that in the previous Inception version the auxiliary classifiers didn’t contribute much more until the end of the training process when accuracies were nearly around the saturation point. That’s why some of the important improvement is implemented in Inception v3 as compared to the Inception v2 are as follows:-</p><ul><li>In the Previous Inception version, 2 auxiliary classifiers are used for having a deeper network. But in Inception v3, only 1 auxiliary classifier is used on the top of the last 17×17 layer. So, Auxiliary Classifier acts as a regularization.</li><li>Factorized the traditional 7x7 convolution into three 3x3 convolutions.</li><li>Label Smoothing As Regularization is implemented in this version and the purpose of label smoothing is to prevent the largest logit (It is the inverse of the sigmoidal “logistic” function or logistic transform used in mathematics) from becoming much larger than all others.</li></ul><p><strong>Inception v4</strong></p><p>The authors noticed that some of the modules were more complicated than necessary. Making the modules more uniform will be able to boost performance in the Inception network. That’s why some of the important improvement in Inception v4 as compared to Inception v3 are as follows:-</p><ul><li>In this version, “<strong>Reduction Blocks</strong>” is implemented which is used to change the width and height of the grid. The earlier versions didn’t explicitly have reduction blocks, but the functionality was implemented.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6b5aa40eccf" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>