<?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 Minsung Kim on Medium]]></title>
        <description><![CDATA[Stories by Minsung Kim on Medium]]></description>
        <link>https://medium.com/@kimchiwave?source=rss-f687e4de1182------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*hN81-ng9FqNZXzpNK2kTKQ.jpeg</url>
            <title>Stories by Minsung Kim on Medium</title>
            <link>https://medium.com/@kimchiwave?source=rss-f687e4de1182------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 16:58:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kimchiwave/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[1/5 — EEG vs fNRIS]]></title>
            <link>https://medium.com/@kimchiwave/1-5-eeg-vs-fnris-44c99c84a156?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/44c99c84a156</guid>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Tue, 06 Jan 2026 03:54:59 GMT</pubDate>
            <atom:updated>2026-01-06T03:54:59.388Z</atom:updated>
            <content:encoded><![CDATA[<h3>1/5 — EEG vs fNRIS</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*90hz3-bTXSzqs5FO" /></figure><p>At an early stage of this project, EEG naturally emerged as the primary sensing modality for measuring user focus during desktop work. EEG is widely used for attention and vigilance research, offers high temporal resolution, and aligns well with real time feedback requirements. From a system perspective, it appeared to be the most straightforward and conventional choice.</p><p>However, as the design constraints became more concrete, several fundamental concerns surfaced. The device is limited to forehead-only placement, uses dry electrodes, and must support frequent donning and doffing in everyday use. Under these constraints, EEG signal quality becomes highly sensitive to facial muscle activity, eye movements, and subtle changes in electrode contact. In particular, beta-band activity, which is commonly associated with attention and alertness, overlaps significantly with electromyographic signals from the forehead and jaw muscles. This creates a structural risk in which facial tension rather than cognitive attention could dominate the measured signal.</p><p>These concerns shifted the focus from simple feasibility toward the validity of the resulting metrics. The question was no longer whether EEG could be made to work, but whether it could reliably measure the intended construct under realistic usage conditions. This prompted a broader examination of alternative neurophysiological sensing approaches.</p><p>Functional near infrared spectroscopy emerged as a compelling candidate in this context. Unlike EEG, fNIRS does not measure electrical activity directly and is therefore largely immune to electromyographic interference. In addition, fNIRS has a strong research foundation for assessing prefrontal cortical workload and sustained cognitive effort, which are highly relevant to prolonged desktop tasks. From a signal interpretation standpoint, it offered a cleaner separation between cognitive demand and facial muscle activity.</p><p>At the same time, adopting fNIRS introduces a different set of challenges. Its temporal resolution is inherently limited by hemodynamic response dynamics, resulting in slower signal changes compared to EEG. This shifts the nature of feedback from instantaneous state detection to trend-based assessment. Moreover, the engineering challenges move away from low-noise analog electronics toward optical coupling, motion robustness, and mechanical design. In other words, fNIRS does not simplify development, but rather changes where the difficulty lies.</p><p>For these reasons, fNIRS is not viewed as a direct replacement for EEG, but as a technology that answers a slightly different question. EEG is well suited for detecting rapid fluctuations in attention and vigilance, while fNIRS is more appropriate for tracking sustained cognitive workload over longer time scales. Recognizing this distinction has been an important step in clarifying both the technical direction and the conceptual framing of the product.</p><p>As a result, fNIRS is now being considered as a viable sensing option that merits deeper evaluation, not because EEG is inadequate in principle, but because the constraints of a wearable, forehead-only, everyday device place unusually strict demands on signal validity. This reflection has helped refine the problem definition itself and has provided a clearer foundation for future design decisions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=44c99c84a156" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[1/4 Confirmed Decisions]]></title>
            <link>https://medium.com/@kimchiwave/1-4-confirmed-decisions-be9cdcfdb872?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/be9cdcfdb872</guid>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Sun, 04 Jan 2026 23:53:14 GMT</pubDate>
            <atom:updated>2026-01-04T23:53:14.710Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qn_63yUZA9CY6ccJ" /></figure><p>I am not gonna waste my life anymore.</p><p>Slow and steady wins the race.</p><p><strong>1. Product Definition</strong></p><ul><li>Goal: Real-time estimation of the user’s focus state during desktop work</li><li>Output: Continuous focus score (0–100)</li><li>Use case: Personal productivity UX (not clinical or diagnostic)</li></ul><p><strong>2. Platform</strong></p><ul><li>Target OS: Windows and macOS</li><li>Application type: Desktop application</li><li>Development stack: Python</li></ul><p><strong>3. Signal Scope</strong></p><ul><li>EEG-based measurement</li><li>Core frequency bands<br>: Theta (4–7 Hz) &amp; Beta (13–30 Hz)</li></ul><p><strong>4. Wearability and Electrode Constraints</strong></p><ul><li>Placement: Forehead only</li><li>Electrode type: Dry electrodes only</li><li>Gel usage: Not allowed</li><li>Number of channels: 4 channels</li></ul><p><strong>5. Electrode Configuration</strong></p><ul><li>Active electrodes: 4</li><li>Reference electrode: 1</li><li>Bias electrode: 1</li><li>Total electrodes: 6</li></ul><p>Reference and Bias placement</p><ul><li>Forehead-centered + Flat pad-style surface contact</li></ul><p><strong>6. Data Strategy</strong></p><ul><li>Raw EEG data is the primary signal source</li></ul><p>Raw data is required for</p><ul><li>Beta and Theta computation</li><li>User-specific calibration</li><li>Algorithm tuning and iteration</li></ul><p><strong>7. Transmission Method</strong></p><ul><li>Transport protocol: Bluetooth Low Energy (BLE)</li></ul><p>Data format</p><ul><li>Raw EEG data: Binary</li><li>Summary metrics: Optional, may be sent in parallel</li></ul><p><strong>8. BLE Raw Data Specification</strong></p><ul><li>Sampling rate: 250 Hz</li></ul><p>Packetization</p><ul><li>20 samples per packet</li></ul><p>Reliability</p><ul><li>Sequence number included for packet drop detection</li></ul><p>Packet structure</p><ul><li>Fixed-length packets</li><li>No explicit timestamp</li><li>Time reconstructed from sequence index</li></ul><p><strong>9. UX Output Specification (MVP)</strong></p><ul><li>Focus score range: 0–100</li><li>Update interval: 1 second</li></ul><p><strong>10. Baseline Calibration</strong></p><ul><li>Baseline type: Session-based</li></ul><p>Procedure</p><ul><li>1 minute of normal, natural work behavior at session start</li></ul><p>Purpose</p><ul><li>Used as the reference baseline for focus scoring</li></ul><p><strong>11. Data Storage Policy</strong></p><ul><li>Raw EEG data is stored</li><li>Daily focus reports are generated and saved</li></ul><p>Raw data lifecycle</p><ul><li>Automatically deleted after 30 days</li></ul><p>Estimated storage usage</p><ul><li>Approximately 80–120 MB per day (8-hour workday)</li><li>Approximately 2–4 GB per month</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=be9cdcfdb872" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Annual Report 2024]]></title>
            <link>https://medium.com/@kimchiwave/annual-report-2024-cfc3518ca17a?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/cfc3518ca17a</guid>
            <category><![CDATA[dreams]]></category>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Tue, 31 Dec 2024 13:53:29 GMT</pubDate>
            <atom:updated>2024-12-31T13:53:29.566Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8cEjx_t85e1TPXCG" /><figcaption>Photo by <a href="https://unsplash.com/@kryszpin?utm_source=medium&amp;utm_medium=referral">Andrzej Kryszpiniuk</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>As the year that was filled by being emptied comes to an end, 2024 was a special year for me. In the past, I used to feel anxious when minor cracks appeared in human relationships or when the future was uncertain. This year, I had a truly unique experience. Despite being in the midst of more uncertainty than ever, my mind was strangely calm. Everything I did was because I wanted to, so even though I was incredibly busy, working 100+ hours a week, I felt like I wasn’t stressed. In terms of relationships, I think I was able to express myself more freely while maintaining my true self. I’m glad I’ve matured a bit more than last year.</p><p>Here are some memorable things I did this year:</p><ul><li>My company received $100K in funding from the government. In November, I had a fun experience traveling to Vietnam as part of the program.</li><li>Although it wasn’t very successful, I went through the medical device certification process.</li><li>I interned at the Neuroimaging lab at Yonsei University’s Severance Hospital for a semester, gaining experience in academia.</li><li>I graduated! I lost 7kg in a week for the graduation photos because I had gained too much weight, lol.</li><li>I worked really hard to earn money.</li><li>I’m applying for a master’s program in the U.S. I’ve completed one application and plan to apply to several more. I’ve decided on Engineering Management as my field.</li><li>I took the TOEFL and GRE.</li><li>I started an online bachelor’s program in computer science and completed 24 credits.</li><li>I played ping-pong a lot with my friend.</li></ul><p>Next year, I aim to achieve the following goals perfectly:</p><ol><li><strong>I will establish a financial foundation to complete my master’s in the U.S. without any issues.<br></strong>Although I can’t disclose it here, I’m working very hard to earn money, and if things go as planned, I should be able to cover the costs of settling in the U.S. The master’s program in the U.S. is very expensive due to the business model of American universities, but I believe I can manage it all on my own. If I accomplish this, I will feel very proud of myself.</li><li><strong>I plan to study more engineering to follow the Engineering Management courses well.</strong> <br>I’m continuing to take computer science courses, but I think I’ll need to work on a few projects. Someday, I also want to study physics.</li><li><strong>I want to have the stamina to run 10km without stopping.</strong></li><li><strong>I want to create a printmaking artwork.</strong></li></ol><p>I firmly believe that striving to live freely while making constant efforts leads to happiness and, ultimately, success. That’s why I liked 2024. I hope my 2025 will be busy and, at the same time, free.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cfc3518ca17a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding Oligo synthesis, Sanger sequencing, Nanopore sequencing in 5 minutes]]></title>
            <link>https://medium.com/@kimchiwave/understanding-oligo-synthesis-sanger-sequencing-nanopore-sequencing-in-5-minutes-db29394b800a?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/db29394b800a</guid>
            <category><![CDATA[sanger-sequencing]]></category>
            <category><![CDATA[sequencing]]></category>
            <category><![CDATA[biotech]]></category>
            <category><![CDATA[oligo-synthesis]]></category>
            <category><![CDATA[nanopore-sequencing]]></category>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Thu, 26 Sep 2024 05:18:13 GMT</pubDate>
            <atom:updated>2024-09-26T05:20:53.381Z</atom:updated>
            <content:encoded><![CDATA[<h3>Understanding Oligo Synthesis, Sanger Sequencing and Nanopore Sequencing in 5 minutes</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CNL1U6vTKGtlFDt_" /><figcaption>Photo by <a href="https://unsplash.com/@nci?utm_source=medium&amp;utm_medium=referral">National Cancer Institute</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>DNA is composed of three main components, and its sequence is determined by four nitrogenous bases: adenine (A), guanine (G), cytosine (C), and thymine (T). These four bases serve as the “alphabet” of DNA, and the process of reading this sequence is called sequencing. Rather than observing the molecular structure of these bases directly under a microscope, which is inefficient, sequencing is performed indirectly through chemical reactions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/444/0*VMnNaD4jSdjlnd58" /><figcaption><a href="https://atdbio.com/nucleic-acids-book/Nucleic-acid-structure">ATDBio — Nucleic Acids Book — Chapter 1: Nucleic acid structure</a></figcaption></figure><p>In DNA, phosphate groups link sugar molecules to form the backbone, and the nitrogenous bases pair together via hydrogen bonds, forming a double helix structure. Adenine pairs with thymine (A-T), and guanine pairs with cytosine (G-C), ensuring that information is not lost. The directionality of the DNA strands is crucial — strand runs in a 5&#39; to 3&#39; direction.</p><h3>Oligo Synthesis</h3><p>Oligonucleotide(Oligo) synthesis is the process of chemically synthesizing short strands of DNA or RNA. The word “oligo” means “short” and refers to a short piece of nucleic acid, typically 20 to 200 bases in length. Oligos are used for a variety of purposes, including genetic analysis, PCR, gene editing, sequencing, and as probes or primers for research.</p><p><strong>How is it possible to create a DNA or RNA strand in any desired order, even though bases only pair as A-T and G-C?</strong></p><p>In natural DNA replication, bases pair complementarily (A with T, and G with C), but in <strong>oligonucleotide synthesis</strong>, chemical methods are used to assemble the strand in any desired sequence, regardless of base-pairing rules. During this process, bases (A, T, G, C, or U) are added one by one in a specific order as dictated by the user.</p><p>Each base is chemically activated in its phosphoramidite form and sequentially attached to the growing strand. Protective groups ensure that only the correct base is added at each step, preventing unwanted reactions. Since the synthesis follows a programmed sequence, the rules of complementary base pairing do not apply. Instead, the desired sequence is built by chemical bonding, allowing the creation of custom DNA or RNA strands.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/673/1*m1NWq_wSDbURCNZWoo4low.png" /><figcaption><a href="https://www.tcichemicals.com/KR/ko/support-download/tcimail/application/187-19">DNA Phosphoramidites for Oligonucleotide Synthesis | Tokyo Chemical Industry Co., Ltd. (tcichemicals.com)</a></figcaption></figure><p>In this process, short, single-stranded DNA sequences, typically between 20 and 200 bases long, are chemically synthesized in the lab. This is usually done using specialized machines or external services because manual synthesis is costly. Researchers can order a specific DNA sequence, and it will be synthesized and delivered in a tube. Oligo synthesis is particularly useful when there is no complementary strand available to serve as a template.</p><h3>PCR</h3><p><strong>PCR (Polymerase Chain Reaction)</strong> is fundamental to <strong>Sanger sequencing</strong> and is used to amplify DNA. In PCR, a single strand of DNA can be used to generate large amounts of DNA by making complementary strands. When a high concentration of DNA is needed, PCR amplifies it exponentially through a process that involves adding nucleotides (dNTPs) and other components.</p><p>Key components of PCR include:</p><ul><li><strong>Taq polymerase</strong>: A heat-stable enzyme that synthesizes new DNA strands by adding nucleotides to the growing strand.</li><li><strong>Primers</strong>: Short, synthetic DNA sequences that bind to the template strand and provide a starting point for Taq polymerase to extend the DNA. They attach at the 3&#39; end of the target sequence, allowing complementary base pairs to attach.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/0*yqNjsi8HE2g1Uemh.png" /><figcaption><a href="https://byjus.com/biology/pcr/">PCR -Principle, Steps, Types, Components And Applications Of PCR (byjus.com)</a></figcaption></figure><p>PCR involves a cycle of heating and cooling. The DNA is first <strong>denatured</strong> by heating to separate the strands. Then, the temperature is lowered for <strong>annealing</strong>, allowing primers to bind to the template DNA. Finally, during the <strong>extension</strong> phase, Taq polymerase synthesizes the new DNA strand. With each cycle, the amount of DNA doubles exponentially (2^n), significantly increasing the concentration of the target DNA.</p><h3><strong>Sanger Sequencing</strong>:</h3><p>Sanger sequencing is the most traditional and widely used method for determining the sequence of DNA bases (A, T, G, C). Essentially, the goal is to read the order of these nucleotides. During this process, the same principles of PCR (Polymerase Chain Reaction) apply, where complementary strands are synthesized. The difference here is that fluorescent labels are attached to the nucleotides. Each time a nucleotide is added during the synthesis, it emits a distinct color corresponding to its base (A, T, G, or C). By observing the color that appears at each step, scientists can determine which nucleotide was added and thus infer the DNA sequence in real-time and in a time series manner. This method allows for a direct, visual readout of the nucleotide sequence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r1CSqL7ZEwYWvKxA.jpeg" /><figcaption><a href="https://microbenotes.com/sanger-sequencing/">Sanger Sequencing: Principle, Steps, Applications, Diagram (microbenotes.com)</a></figcaption></figure><p>However, Sanger sequencing has its limitations. The fluorescent dyes used can overlap and become indistinguishable when dealing with very long sequences, making it difficult to accurately read beyond a few hundred base pairs. As a result, the length of DNA that can be sequenced is usually limited to a few hundred base pairs. To overcome this, carefully designed primers are used to segment the DNA for effective PCR amplification, ensuring better sequence accuracy within these shorter ranges.</p><h3><strong>Nanopore Sequencing</strong>:</h3><p>Nanopore sequencing is a cutting-edge technology designed to handle much larger datasets and sequence longer stretches of DNA. Unlike Sanger sequencing, it does not rely on light or fluorescent markers. Instead, the DNA is split into single strands and passed through a nanoscale pore (hence the name “nanopore”). As each nucleotide (A, T, G, or C) passes through the pore, it generates a distinct change in the electrical current. These changes are detected and used to identify the specific nucleotide passing through the pore based on its unique electrical signature.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*s1NCeIsM7nNR7la8.png" /><figcaption><a href="https://www.yourgenome.org/theme/what-is-oxford-nanopore-technology-ont-sequencing/">What is Oxford Nanopore Technology (ONT) sequencing? (yourgenome.org)</a></figcaption></figure><p>This technology offers significant advantages, especially when dealing with large genomes or long DNA sequences. Since it doesn’t rely on fluorescence and can sequence long stretches of DNA, it avoids the limitations of Sanger sequencing. Nanopore sequencing is often preferred for projects requiring large-scale sequencing due to its ability to process vast amounts of data.</p><p>In the sequencing industry, the key to success lies in building strong <strong>connections</strong>, offering competitive <strong>pricing</strong>, and ensuring <strong>speed</strong> in delivering results. These factors are crucial in attracting and retaining clients, especially as sequencing demands grow and evolve.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db29394b800a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Operating System 101 — Bird eye’s view]]></title>
            <link>https://medium.com/@kimchiwave/operating-system-101-bird-eyes-view-c717b7a56d29?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/c717b7a56d29</guid>
            <category><![CDATA[operating-systems]]></category>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Thu, 19 Sep 2024 03:15:48 GMT</pubDate>
            <atom:updated>2024-09-19T03:15:48.235Z</atom:updated>
            <content:encoded><![CDATA[<h3>Operating System 101 — Bird eye’s view</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*btjMsiOIWlzMI0rH" /><figcaption>Photo by <a href="https://unsplash.com/@purzlbaum?utm_source=medium&amp;utm_medium=referral">Claudio Schwarz</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Operating System</h3><p>A system software that serves as an intermediary between computer hardware and computer users.</p><p><strong>‘User’:</strong> Refers to a person, device, or another computer using the computer.<br><strong>‘Software’:</strong> All programs necessary to perform functions on a computer.<br><strong>‘Hardware’:</strong> The resources provided by the computer, such as the processor (CPU), memory, and peripheral devices.</p><h3>Roles of the Operating System</h3><ol><li><strong>Coordinator:</strong><br>Controls the use of operating elements and allows communication between users and applications, only providing an environment for tasks to be performed.</li><li><strong>Resource Allocator or Manager:</strong><br>Allocates resources needed for each application and determines how resources are allocated.</li><li><strong>Controller of Applications and Input/Output Devices:</strong><br>Controls various input/output devices and applications, ensuring that user programs run properly to prevent inappropriate use or errors within the computer system.</li></ol><ul><li>Provides an interface between hardware, users, applications, and system programs.</li><li>Coordinates and manages computer resources such as processors, memory, input/output devices, and communication devices for effective use.</li><li>Controls applications that handle high-level services such as mail transmission, file system checks, and server tasks.</li><li>Controls input/output to protect the computer system from multiple users and manages data.</li></ul><h3>Objectives of the Operating System</h3><ul><li><strong>Convenience:</strong> Provides a convenient environment and interface (GUI).</li><li><strong>Efficiency:</strong> Improves system performance, throughput, reliability, availability, and reduces response time.</li><li><strong>Control Services:</strong> Enhances management of input/output device operations and prevents system errors.</li></ul><h3>Types of Operating Systems</h3><ol><li><strong>By Number of Simultaneous Users:</strong></li></ol><ul><li><em>Single-user system</em> (e.g., Windows): The OS supports only one user at a time, with all resources dedicated to a single user. Primarily used in microcomputers and personal computers.</li><li><em>Multi-user system</em> (e.g., Unix): Supports multiple users at the same time. It requires management of ownership rights to various system resources (e.g., files). Multi-tasking is typically necessary, making the OS’s functions and structure more complex.</li></ul><p><strong>2. By Task Processing Method:</strong></p><ul><li><em>Single-tasking system</em>: Only one task (program or process) exists and runs at a time. After finishing one program, the next program starts.</li><li><em>Multi-tasking system</em> (Multiprogramming system): Supports the execution of multiple tasks simultaneously. It requires managing issues that arise from parallel execution, such as multiple requests for the same resources.</li></ul><p><strong>3. By Usage Environment:</strong></p><ul><li><em>Batch processing system</em>: Gathers user requests for tasks over a certain period and processes them all at once. Once a task starts, nothing else can be done until it completes. Used in payroll processing. The system efficiently uses CPU by processing large batches of tasks using auxiliary storage devices like magnetic tape and disk, but all tasks must be of the same type. Input of all types may not be possible, and since input/output devices are slower than the processor, there may be CPU idle time.</li><li><em>Time-sharing system</em>: Alternates the processor among multiple users in short intervals, making each user feel like they are monopolizing the processor. This allows multiple users to use a single computer system simultaneously. It supports interactive systems and has the advantage of shared CPU usage, though it requires control and management of inter-process communication and I/O device use.</li><li><em>Distributed processing system</em>: Distributes the functions of a large computer to multiple mini-computers located in different regions, communicating and processing via a network. It has the advantages of low cost and high performance, but building the network can be difficult.</li><li><em>Real-time system</em>: Processes data immediately upon receiving requests from terminals or control objects, outputting results or responding to demands. It incorporates the concept of online systems, reducing user effort and processing time. Since data arrives randomly, temporary storage and queuing of input/output data are necessary.</li></ul><h3>Evolution of Operating Systems</h3><ol><li><strong>1940s: No Operating System (Sequential Processing per Task)</strong></li></ol><ul><li>Programs were written and executed directly in machine language using a sequential processing system for each task.</li><li>Instructions explicitly specified everything, including how to store commands in the processor, the targets for calculations, how and where to store results, and when and where to output them.</li></ul><p><strong>2. 1950s: Batch Processing System</strong></p><ul><li>IBM 701 (GM), IBM 704 (GM, North American Aviation)…</li><li>The early operating system, Batch Processing System, focused on reducing the time required to load and release tasks (methods such as batch processing, buffering, etc.). Data was collected and processed all at once.</li></ul><p><strong>3. 1960s: Development of Multi-programming, Time-sharing, Multiprocessing, and Real-time Systems</strong></p><ul><li>Introduction of hardware management using device independence (running programs with different input/output devices) and multi-programming.</li><li>Multiple programs were loaded into memory, and the processor was alternately allocated to them, maximizing processor usage and allowing multiple programs to run simultaneously.</li><li>Improved processing capabilities through time-sharing, multiprocessing, and real-time systems.</li><li>Example: SABRE reservation system used by American Airlines.</li></ul><p><strong>4. 1970s: Distributed Processing System</strong></p><ul><li>Wide use of computer networks and online processing methods.</li><li>Enabled remote usage of computers over a network.</li><li>Microprocessors were introduced, allowing personal computers.</li><li>Users could communicate with remote systems, making information security a key concern.</li></ul><p><strong>5. 1970s-1990s</strong></p><ul><li><strong>1970s:</strong> Command-line based systems.</li><li><strong>1980s:</strong> Menu-based systems, more user-friendly.</li><li><strong>1990s:</strong> GUI-based systems.</li></ul><p><strong>6. 2000s Onwards</strong></p><ul><li>Popularization of mobile devices (Android, iOS…).</li><li>Emergence of IoT technology, where OS performance in low-cost, low-power, and low-memory environments became critical.</li><li>Cloud computing enabled users to transmit personal files and data via operating systems from anywhere.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6jDodanRD2X176sC" /><figcaption>Photo by <a href="https://unsplash.com/@warrenumoh?utm_source=medium&amp;utm_medium=referral">Warren Umoh</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Functions of an Operating System</h3><ul><li><strong>Memory Management</strong><br><em>Main memory:</em> Refers to the primary storage that the processor can directly address. The OS checks which parts of memory are in use and by whom, determines which processes to load into memory, and decides how to allocate and reclaim memory.<br><em>Secondary storage:</em> Since main memory is limited, the system relies on secondary storage for continuous data and program storage. The OS manages free space, allocates storage space for new files, schedules memory access requests, and creates and deletes files.</li><li><strong>Process Management</strong><br><em>Process:</em> Consists of resources like the processor, memory, files, and input/output devices. These resources are provided when the process is created or allocated during execution. <br>Scheduling of processes and threads / Creation and removal of user and system processes / Stopping and restarting processes / Synchronization and communication between processes / Prevention of deadlocks.</li><li><strong>Device Management</strong><br><em>Input/output control:</em> The OS uses device drivers, which are special programs, to interact with I/O devices. Device drivers provide an interface for communication with specific hardware devices. The OS offers features like buffer-caching, generic device drivers, and device-specific drivers.</li><li><strong>File Management<br></strong>Manages the storage, retrieval, and location of input/output files. Computers can store files in various physical formats, and the OS provides a unified storage format for efficient data use.<br>Usually, files are organized in directories, and the OS controls access when multiple users attempt to access files.<br>OS creates and deletes files, manages directory creation/deletion, maps files to secondary storage, and stores files securely on non-volatile storage.</li><li><strong>System Protection<br></strong>Controls access to computer resources, such as programs, processes, and users. Multiple processes can run simultaneously, so mutual protection is required.<br>Example: Windows 7’s system protection feature helps recover individual folders affected by ransomware or viruses.</li><li><strong>Network (Communication Functionality)<br></strong>Processors use various network configurations to communicate. The OS must consider aspects like path settings, connection policies, collision management, and security when processors are connected and communicating.</li></ul><h3>Operating System Services</h3><ol><li><strong>Booting/Bootstrapping</strong></li></ol><ul><li>The process of loading the operating system into main memory.</li><li>A bootloader, a small program stored in ROM, loads the OS from secondary storage (like a hard drive) into main memory.<br>Process:</li><li>The BIOS program in ROM runs &gt; The boot program is loaded into main memory &gt; The operating system is loaded into main memory &gt; Programs are loaded by the OS &gt; The CPU executes the program based on the OS commands.</li></ul><p><strong>2. User Interface</strong></p><ul><li><em>Menu Interface:</em> Interacts with the system through menus.</li><li><em>GUI:</em> Uses visual elements like icons, visual indicators, widgets, etc.</li><li>Executes programs by loading them into memory and allocating processor time.</li><li>Provides methods for performing I/O operations.</li><li>Manages file system operations like opening, saving, and deleting files.</li><li>Network management, error detection (e.g., memory overflows, bad sectors on a hard drive, improper memory access).</li><li>Allocates resources when multiple users or tasks are run simultaneously.</li></ul><p><strong>3. System Services</strong></p><ul><li>Account management, protection, and security.</li></ul><p><strong>4. System Call Service</strong></p><ul><li>Interface (API) between running programs and the OS.</li><li>User programs make system calls to request OS functions.</li><li>These system calls allow communication with core kernel services, process creation, hardware services, etc.</li><li>System calls enable processes to interact with the OS in ways they can’t do at the user level, such as calling commands, subroutines, or using a command interpreter.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c717b7a56d29" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Database 101-DBS & DB users]]></title>
            <link>https://medium.com/@kimchiwave/database-101-dbs-db-users-3dd9fed4b6f5?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/3dd9fed4b6f5</guid>
            <category><![CDATA[database]]></category>
            <category><![CDATA[computer-science]]></category>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Mon, 16 Sep 2024 05:22:25 GMT</pubDate>
            <atom:updated>2024-09-16T05:22:25.553Z</atom:updated>
            <content:encoded><![CDATA[<h3>Database</h3><p>= A meaningful collection of data representing a part of the real world.</p><ul><li>Real-time accessibility, continuous changes, concurrent sharing, reference by content.</li></ul><h3>DBMS (Database Management System)</h3><p>= A software package responsible for creating and managing the database.</p><ol><li><strong>Database Definition</strong> = The process of specifying data types, structures, constraints, etc., for the data stored in the database (metadata).</li><li><strong>Database Construction</strong></li><li><strong>Database Manipulation</strong></li><li><strong>Database Sharing</strong></li><li><strong>Database Protection</strong></li><li><strong>Database Maintenance</strong></li></ol><h3>Database System</h3><p>A Database System refers to the database, the database management system that manages it, and the application programs, collectively.</p><p>A <strong>Database</strong> is a collection of related data.</p><p>A <strong>Database Management System (DBMS)</strong> is a software package responsible for creating and managing the database.</p><h3>Components</h3><ul><li>Database, schema, database management system, database language (SQL), database computer (server, client), database users.</li></ul><p>The database organizes data records, data items, entities, and relationships. e.g., University Information</p><p><strong>Entity</strong>:</p><ul><li>STUDENT</li><li>COURSE</li><li>SECTION</li><li>GRADE</li></ul><p><strong>Relationship</strong>:</p><ul><li>SECTION belongs to a specific COURSE.</li><li>STUDENT participates in a SECTION.</li><li>COURSE has prerequisite COURSES.</li><li>INSTRUCTOR teaches a SECTION.</li><li>COURSE is offered by a DEPARTMENT.</li><li>STUDENT majors in a DEPARTMENT.</li></ul><p><strong>Record</strong>:</p><ul><li>{Course Name: Intro to CS, Course Number: CS1310, Year: 92, Instructor: Knuth…} -&gt; A set of data.</li></ul><p><strong>Table</strong>:</p><ul><li>Columns (Course Name, Number, Year…) where data is indexed.</li></ul><p>= How tables and records are structured is the key idea in building a database.</p><p>In a DBMS, database manipulation functions include querying and updating.</p><ul><li><strong>Query</strong>: Search for Smith’s grades, search for the prerequisite courses for the database course.</li><li><strong>Update</strong>: Change Smith to a sophomore, update Smith’s grade in the database course last semester to an A. =&gt; Such queries and updates are specified and processed using the query language (SQL) of the DBMS.</li></ul><h3>Characteristics of Database Systems</h3><p><strong>Data Management System</strong></p><ul><li>Largely divided into file systems and database systems.</li><li>When building a system that stores and manages actual data, you must choose between the two.</li><li><strong>File system</strong>: When changing the file structure, the application that reads that structure also needs to be updated. Each application independently manages files. In the file system processing method, files necessary for a specific software application are defined and implemented separately as part of programming the software.</li><li>Example: The academic department and the accounting department may manage the student’s grades and tuition separately. Two different files are created for a single student, and separate applications must be written for each department.</li><li>If the same data is redundantly stored in more than one file, it wastes storage space and makes maintenance difficult. If the redundant data is not updated properly, <strong>data integrity</strong> is compromised (the information in the real world does not match the information in the computer system).</li><li><strong>Program and data dependency</strong>: Applications using files are written according to the data format within the file. If the format changes, the related application must be modified. Security measures to prevent unauthorized access to specific data are also difficult to implement.</li></ul><p><strong>= This leads to the emergence of database systems.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*saO4UJEZApEZEqW5" /><figcaption>Photo by <a href="https://unsplash.com/@jankolar?utm_source=medium&amp;utm_medium=referral">Jan Antonin Kolar</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Database System</h3><ul><li><strong>Self-describing feature</strong>: The database catalog stores <strong>metadata</strong>, allowing a single DBMS to manage multiple databases.</li><li><strong>Program-data independence and data abstraction</strong>: Even if the data storage structure in the database changes, applications are not affected and do not need to be updated.</li><li><strong>Provision of multiple views of data</strong>: Users can define views of only the data they are interested in, rather than the entire database.</li><li><strong>Data sharing and multi-user transaction processing</strong>: Data used in multiple applications is integrated and stored in a single database.</li><li>This ensures <strong>logical and physical independence of data</strong>, saving memory space, maintaining data consistency and integrity, enhancing security, enabling real-time processing, and facilitating integration, management, and updates.</li><li><strong>Disadvantages</strong>: Increased computing costs, risk of overload due to concentrated access to large disks, system complexity, etc.</li></ul><h3>Representative Database Vendors and Products</h3><p><strong>Vendors</strong>: Oracle, MySQL, MariaDB, INFORMIX, DB2, MS SQL Server, PostgreSQL, ALTIBASE HDB, etc.</p><p><strong>Key Features</strong>:</p><ul><li><strong>Redundancy control</strong> (prevents duplicate storage of the same data),</li><li><strong>Access control</strong> (DBMS grants different privileges to users, controlling data access based on permissions),</li><li><strong>Interface provision</strong> (CLI, GUI, etc.),</li><li><strong>Relationship representation</strong>,</li><li><strong>Sharding/partitioning</strong>,</li><li><strong>Integrity constraints</strong>,</li><li><strong>Backup/recovery</strong>.</li></ul><h3>Classification of Database Users</h3><p>Many users are involved in defining, building, managing, and manipulating a database.</p><ul><li><strong>DBA (Database Administrator)</strong>: Responsible for managing the database (contrasts with the System Administrator, SA).</li><li><strong>DB Designer</strong>: Responsible for designing the database.</li><li><strong>End User</strong>: Responsible for querying, updating, and generating reports from the database.</li><li><strong>DBMS Designers and Implementers</strong>: Users who design and implement the DBMS software itself.</li><li><strong>Tool Developers</strong>: Design and develop tools needed to use the database (e.g., database design and construction tools, performance tools).</li><li><strong>Operations and Maintenance Personnel</strong>: Handle the operation and maintenance of the system.</li></ul><h3>Advantages of DBMS</h3><ol><li><strong>Control of redundancy</strong>: By specifying rules, the DBMS automatically controls data duplication.</li><li><strong>Prevention of unauthorized access</strong>: DBAs grant users accounts and specify access controls through the security and privilege subsystems (e.g., ensuring that only authorized users can access classified information).</li><li><strong>Provision of persistent storage for program objects</strong>: A primary goal of object-oriented database systems. For example, C++ and Java have different data types, so transforming data into a suitable file format is essential. Object-oriented DBMSs offer compatibility with programming languages and automatically convert data into the required format.</li><li><strong>Efficient storage structure for query processing</strong>: To reduce the time spent retrieving the desired records from the disk, DBMSs provide special data structures, often based on tree structures or hash structures with indexes.</li><li><strong>Backup and recovery</strong>: In case of failure during a transaction, the system rolls back to the pre-transaction state.</li><li><strong>Provision of various user interfaces</strong>.</li><li><strong>Representation of complex relationships between data</strong> (schema).</li><li><strong>Enforcement of Integrity Constraints (IC)</strong>: To accurately reflect the real world, the DBMS must define and verify constraints.</li><li><strong>Rule-based inference</strong>: By defining deductive rules, new information can be inferred from stored facts, and active rules can automatically execute when certain events or conditions occur.</li><li><strong>Reduced development time</strong>.</li><li><strong>Flexibility</strong>.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3dd9fed4b6f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2024 Q2 Earnings of My life]]></title>
            <link>https://medium.com/@kimchiwave/2024-q2-earnings-of-my-life-5fa7482b51e3?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/5fa7482b51e3</guid>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Sat, 24 Aug 2024 01:47:00 GMT</pubDate>
            <atom:updated>2024-08-24T01:47:00.752Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0BWFCYahnHpG100a" /><figcaption>Photo by <a href="https://unsplash.com/@sam?utm_source=medium&amp;utm_medium=referral">Sam Schooler</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><a href="https://medium.com/@kimchiwave/2024-q1-earnings-of-my-life-ed2f6a8a2f0d">2024 Q1 Earnings of My Life. It’s already March. | by Minsung Kim | Medium</a></p><p>I should have written my Q2 earnings post earlier, but before it gets too late, I’m going to reflect on the time from April until now.</p><p>I’m currently in Jeju, a small island in Korea and my hometown. I was born and raised here, spending my entire childhood on the island before leaving for college. Now, I’m back in Jeju, preparing for a master’s program in computer science in the U.S.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5tOpmX_orwufpIOB" /><figcaption>Photo by <a href="https://unsplash.com/@fwphotos?utm_source=medium&amp;utm_medium=referral">Felix Wiedemann</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I have a TOEFL test on September 7th and the GRE on October 11th. Every day, I go to the library with my closest friend, spending at least 10 hours there. Preparing for a master’s program has been challenging in many ways. My undergraduate degree is in public policy, but many of the programs I’m applying to highly recommend that candidates have a background in CS. Additionally, the tuition fees for international students are significantly higher — it’s a major business model for universities in the U.S. If I were from a wealthy background, this wouldn’t be a problem, but my budget is extremely tight.</p><p>However, I believe in myself now more than ever, and I’m confident that my efforts will lead to positive results. If anyone would like to chat over coffee, please reach out — I’d really appreciate it.</p><p>In March, I joined a lab at Yonsei University that primarily researches the brain. I wanted to explore my hypothesis: “Is it possible to manipulate a machine using EEG in a non-invasive way?” I assisted with several experiments involving EEG, EMG, and fMRI, and reviewed numerous articles. I tried to be in the lab as much as possible, learning something new every day. It was a fascinating experience that gave me insight into the life of scientific research. I concluded that my hypothesis might not be feasible, at least not within my lifetime. However, computational neuroscience captivated me, and I believe it could be the next big trend in AI research. I hope to study more about it in the future.</p><p>In just six days, I will graduate from college. It’s been six years since I started (due to mandatory military service ;-;). Thanks to Yonsei, I’ve had so many incredible opportunities and met many admirable people. My parents will be visiting Yonsei for the first time for my graduation ceremony, and I’m really looking forward to it. In September, I’ll start an online bachelor’s program in computer science through the Korea Academic Credit Bank System. I need to complete 48 credits to earn the degree, and I’ll be taking 24 units this upcoming semester. Since it’s online, I plan to do most of the coursework on the weekends.</p><p>The medical device business I’m working on with a friend is progressing positively, albeit slowly. I’m mainly involved in the business-related aspects, and thankfully, we received funding from the Korean government. The next step is to obtain medical device approval from the Korean Ministry of Food and Drug Safety. We’re aiming to complete the process by December. As I become more involved in this field, my desire to study STEM grows stronger. I want to be an entrepreneur with deep insight and technical literacy. For that, I believe superficial study isn’t enough — I want to delve deeply into the fundamentals. While it’s true that computer science is now an oversaturated, “red ocean” field, it’s still attractive due to its interdisciplinary nature. At the very least, I want to be someone who can offer broad insights and advice to my future employees and colleagues. So, let’s see where this journey takes me.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IejwkBdxfz8Sap1_" /><figcaption>Photo by <a href="https://unsplash.com/@rcrazy?utm_source=medium&amp;utm_medium=referral">Ricardo Rocha</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5fa7482b51e3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Some IR insights]]></title>
            <link>https://medium.com/@kimchiwave/some-ir-insights-83e2d0f6f297?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/83e2d0f6f297</guid>
            <category><![CDATA[irs]]></category>
            <category><![CDATA[investor-relations]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[investment]]></category>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Fri, 17 May 2024 14:28:57 GMT</pubDate>
            <atom:updated>2024-05-17T14:28:57.938Z</atom:updated>
            <content:encoded><![CDATA[<p>We recently had the opportunity to receive mentoring through support from Seodaemun-gu.</p><p>It was a simple project providing promotional materials for the company at first, but luckily, We are now able to continuously receive good advice from a mentor who used to be a VC associate.</p><p>I’m writing this to organize and internalize insights related to IR before I forget them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yzHodPSg5G80TwU_" /><figcaption>Photo by <a href="https://unsplash.com/@precondo?utm_source=medium&amp;utm_medium=referral">Precondo CA</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><ul><li>First of all, IR should be about better packaging of the genuinely excellent essence, and if it’s about deceiving, it’s better to stop the business.</li><li>The technology should be explained in a way that even a middle school student can understand(Not many people actually care or understand). The fastest way is to sit a non-major friend down and dictate the explanation from the basics.</li><li>Focus on instilling three things in the counterpart through the IR slides: “marketability,” “target,” and “BM (business model).”</li><li>In the case of the target, in fields with high regulatory hurdles (licenses, clinical trials, etc.) and market entry barriers (medical industry cartels, difficulty in selling to doctors) like medical devices, it’s more strategic to frame the story around “patients” rather than doctors or hospitals.</li><li>Clearly and thoroughly detail the benefits that can be obtained by each customer segmentation.</li><li>First, patents/licenses and market development should happen simultaneously. Second, the ability to keep confidentiality is also important. Therefore, instead of sending IR materials to a broad and unspecified group of doctors to verify PMF (Product-Market Fit), start narrow and deep with a trusted and reliable medical network.</li><li>The beginning of the IR materials should be “Why did we start this business?” Especially for those who are not doctors making medical devices, if this curiosity is not resolved from the very beginning, it will be hard for them to focus on the latter part of the content.</li><li>As for credence goods, the better it is to let the audience know from the start which organizations or institution we are collaborating with to develop the product.</li><li>It is also a good idea to think about the questions the counterpart might ask and prepare answers. They can be a great source for the posterior IR slides.</li><li>There are many types of IR materials. They could be for investors, for posting on the website, or for recruitment purposes. Make the one with the most impact first. Usually, it is best to assume the audience knows nothing about the business or the technology and create it accordingly. Surprisingly, writing it in a more complicated way is much easier than writing it simply.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=83e2d0f6f297" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding Neural Networks]]></title>
            <link>https://medium.com/@kimchiwave/understanding-neural-networks-1bb8662f5615?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/1bb8662f5615</guid>
            <category><![CDATA[sigmoid]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[softmax]]></category>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Sun, 12 May 2024 07:27:56 GMT</pubDate>
            <atom:updated>2024-05-12T07:27:56.925Z</atom:updated>
            <content:encoded><![CDATA[<p>Perceptron can express complex functions by stacking layers, but a disadvantage is that weights must be manually set by humans.</p><p>The most important property of neural networks is thier ability to automatically learn the appropriate values for weight parameters from data.</p><p><a href="https://medium.com/@kimchiwave/understanding-the-perceptron-3c6cde7cc95f">The concept of perceptron. A Percptron is: | by Minsung Kim | May, 2024 | Medium</a></p><h3>Neural network diagram</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UHvoa9OWlnL1N2Wc.png" /><figcaption><a href="https://otexts.com/fpp2/nnetar.html">11.3 Neural network models | Forecasting: Principles and Practice (2nd ed) (otexts.com)</a></figcaption></figure><p>The far left column (layer 0) is the input layer, the right column (layer 2) is the output layer, and the middle (layer 1) is the hidden layer.</p><p>Looking at the diagram, it doesn’t seem much different from a perceptron. So, what distinguishes a neural network from a perceptron?</p><p><strong>There is only one thing: the activation function.</strong></p><h3>Activation Function</h3><p>What is an activation function?</p><p>It is a function that transforms the total input signal into an output signal. As the name ‘activation’ suggets, the activation function determines whether the total input signal triggers activation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*I22G4Sm7YsQ68R2-.jpg" /><figcaption><a href="https://www.v7labs.com/blog/neural-networks-activation-functions">Activation Functions in Neural Networks [12 Types &amp; Use Cases] (v7labs.com)</a></figcaption></figure><blockquote>z = b + w1*x1 + w2*x2 + w3*x3 [Equation 1]</blockquote><blockquote>y = f(z) [Equation 2]</blockquote><ul><li>Eq 1 calculates the total sum of the weighted input signals and bias, referred to as z.</li><li>Eq 2 outputs f(z) by applying z to the function f().</li></ul><p>Thus, as shown in the diagram, the combined result of the weight signals becomes the node z, and passing through the activation function f() outputs the node f(z).</p><p>Now, do you see what the activation function f() does? It determines whether the total of the input signals is activated or not. Thinking in terms of composite functions might make this easier to understand. Depending on the structure of the activation function, it becomes possible to adjust the level of activation.</p><h3>Types of Activation Functions</h3><h4>1. Sigmoid Function</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Zg-zTg1L6KUEcsBw.png" /><figcaption><a href="https://medium.com/@ionutlang31/how-to-play-around-with-sigmoid-function-to-increase-its-y-max-and-shift-to-the-right-for-positive-ede40daf1fa2">How to play around with sigmoid function to increase its y max and shift to the right for positive x only | by Ionutlang | Medium</a></figcaption></figure><p>The term ‘sigmoid’ means ‘S-shaped’.</p><p>The expression exp(-x) represents e^-x, where e is the natural constant approximately equal to 2.7182…</p><p>Despite its complex appearance, the sigmoid function is essentially just a machine that takes an x and outputs a y, i.e., a function.</p><h4>2. Step Function</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/381/0*B6yJM2N4Wnzn1ZwA.png" /><figcaption><a href="https://www.codecademy.com/resources/docs/ai/neural-networks/binary-step-activation-function">AI | Neural Networks | Binary Step Activation Function | Codecademy</a></figcaption></figure><p>The step function changes its output from 0 to 1 at the threshold of 0. As dipicted, the change in values looks like steps.</p><p>Perceptrons change their output at a threshold value; hence, they utilize the step functions as their activation function.</p><blockquote>Before examining other types of activation functions, let’s note the difference between the sigmoid functiona and the step function.</blockquote><blockquote><strong>The sigmoid function smoothly transitions based on input, whereas the step function changes abruptly at 0. While the step function returns only 0 or 1, the sigmoid function can output a continuous range of values. However, both functions share the commonality that the more significant the input, the higher the output value, within the range of 0 to 1.</strong></blockquote><h4>3. ReLU Function</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4H9T5bPKAmAMjupA.png" /><figcaption><a href="https://machinelearningmastery.com/rectified-linear-activation-function-for-deep-learning-neural-networks/">A Gentle Introduction to the Rectified Linear Unit (ReLU) — MachineLearningMastery.com</a></figcaption></figure><p>The Rectified Linear Unit(ReLU) function outputs the input directly if it is greater than 0; otherwise, it outputs 0. Thus, if the input exceeds 0, it behaves as the identity function y = x. This structure is quite simple.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/660/0*AkJ035-oULfr7HbW.jpg" /><figcaption>Types Of Activation Function in ANN — GeeksforGeeks</figcaption></figure><p>Ultimately, as illustrated, the activation function takes the combined values of the input as x and re-inputs them to determine the final output value.</p><h3>Designing the Output Layer</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/829/0*Eijj1btqheEAVN4n.png" /><figcaption><a href="https://vitalflux.com/how-do-we-build-deep-neural-network-using-perceptron/">Neural Network &amp; Multi-layer Perceptron Examples — Analytics Yogi (vitalflux.com)</a></figcaption></figure><p>Now that we’ve looked at the input and hidden layers, let’s design the output layer. Typically, an identity function is used for regression, and a softmax function is used for classification.</p><h4>1. Identity Function</h4><p>This function outputs the input signal directly, that is, y = x.</p><p>It is primarily used in regression tasks.</p><h4>2. Softmax Function</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*n6YdDEaFvfGvV8sd.jpg" /><figcaption><a href="https://www.youtube.com/watch?app=desktop&amp;v=EuZZ6plg2Tk">Softmax Function in Deep Learning — YouTube</a></figcaption></figure><p>This is mainly used for classification.</p><p>’n’ is the number of neurons in the output layer, and ‘xi’ represents the ‘i’th output. The numerator is the exponential function of the input signal, and the denominator is the sum of the exponential functions of all input signals.</p><p>The output of the softmax function is a real number between 0 and 1. Additionally, the sum of the softmax function’s output is 1. By examining the relationship between the numerator and the denominator, it can be intuitively understood.</p><p>Thus, the output of the softmax function can be interpreted as ‘probabilities’.</p><p>Using this, one can derive probability-based conclusions such as “there is a 70% probability of belonging to the first class, 20% to the second class, and 10% to the third class”.</p><p>However, applying the softmax function does not change the relative orders of elements because the exponential function y = exp(x) is strictly increasing function. Hence, there is no need to apply the softmax function for classification purposes, and it is often omitted in the output layer. However, the softmax function is used in the output layer when training neural networks, but we will explore that later.</p><h4>* Key Differences Between the Sigmoid and Softmax Functions</h4><p>Although the sigmoid and softmax functions may seem similar, they differ significantly in their purpose and computational approach.</p><ol><li>Application Scope: The sigmoid function is used for binary classification, while the softmax function is used for multi-class classification.</li><li>Output Form: The sigmoid function outputs a single probability value, whereas the softmax function outputs a probability distribution for each class.</li><li>Mathematical Expression: The sigmoid function computes an output for a single input, whereas the softmax function takes multiple inputs and calculates a probability for each one.</li></ol><p>These differences mean that each function is chosen and used according to the specific type of problem being addressed.</p><h4>Reference</h4><p><strong>Saito Goki, Deep running starting from the bottom, 2017<br></strong><a href="https://www.amazon.com/Deep-running-starting-bottom-Korean/dp/8968484635">Deep running starting from the bottom (Korean Edition): Saito Goki: 9788968484636: Amazon.com: Books</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1bb8662f5615" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding the Perceptron]]></title>
            <link>https://medium.com/@kimchiwave/understanding-the-perceptron-3c6cde7cc95f?source=rss-f687e4de1182------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c6cde7cc95f</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[perceptron]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Minsung Kim]]></dc:creator>
            <pubDate>Sun, 12 May 2024 03:47:05 GMT</pubDate>
            <atom:updated>2024-05-12T03:47:05.568Z</atom:updated>
            <content:encoded><![CDATA[<h3>The concept of perceptron</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/645/0*kPlUfWch-0gLA9l3" /><figcaption><a href="https://towardsdatascience.com/what-is-a-perceptron-210a50190c3b?gi=501cfef40c3a">What is a perceptron?. A neural network is an interconnected… | by Gerry Saporito | Towards Data Science</a></figcaption></figure><h3>A Percptron is:</h3><ul><li>An algorithm that serves as the orgin of neural networks (Deep Learning)</li><li>It receives multiple signals as inputs and outputs a single signal</li><li>‘x’ denotes the input signals, ‘y’ the output signal, and ‘w’ represents weight.</li><li>The circle is called a neuron or node.</li><li>When input signals are sent to the neuron, they are multiplied by their respective weights (i.e., x1*w1, x2*w2, etc.).</li><li>The neuron only outputs 1 if the total sum of the incoming signals exceeds a certain threshold (similar to the activation spike of neuron).</li><li>Weights adjust the influence of each signal on the outcome; the larger the weight, the more important the corresponding signal is, similar to the concept of resistance in electricity.</li></ul><h4>The formula can be expressed as follows:</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/876/0*o__S6ZsoabX4uf9U.png" /><figcaption><a href="https://velog.io/@dougieduk/Basics-of-Neural-Network-Perceptron-Model">Basics of Neural Network — Perceptron Model (velog.io)</a></figcaption></figure><p>Here, ‘b’ represents the bias. Therefore, a perceptron computes the sum of the weighted inputs and bias, and outputs a specific value ‘y’ based on the sum exceeding a threshold.</p><h4>How do bias and weight differ?</h4><ul><li>w : A parameter that controls the influence of each input signals on the result.</li><li>b : A parameter that adjusts how easily a neuron activates.</li></ul><h3>Limitations of the Perceptron:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/957/0*OTiELsHUdCmzYeF6.png" /><figcaption><a href="https://medium.com/@lucaspereira0612/solving-xor-with-a-single-perceptron-34539f395182">Solving XOR with a single Perceptron | by Lucas Araújo | Medium</a></figcaption></figure><p>It cannot implement an XOR gate. Specifically, a single-layer perceptron cannot implement an XOR gate.</p><p>An XOR gate outputs 1 only when one of x1 or x2 is 1. Can a single line divide ● and ○ in the representation of XOR gate output? To solve this, layers need to be stacked, known as a <strong>multi-layer perceptron</strong>. By combining perceptrons, i.e., stacking layers, an XOR gate can be implemented.</p><h4>The following image shows a two-layer perceptron:</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WPejqqdfUzTj9cCK.jpg" /><figcaption><a href="https://www.youtube.com/watch?app=desktop&amp;v=kNPGXgzxoHw">Neural Networks 6: solving XOR with a hidden layer — YouTube</a></figcaption></figure><p>Signals are transmitted from layer 0 to layer 1, and then from layer 1 to layer 2. What couldn’t be represented with a single-layer percepron can be implemented by adding one more layer. Thus, perceptrons can express more complex functions by stacking layers.</p><p>When multiple layers of perceptrons are combined, even a computer can be built, since a computer is essentially a collection of a series of rules. By stacking layers, nonlinear expressions become possible, and in theory, any computation performed by a computer can be represented.</p><p>reference_</p><p><strong>Saito Goki, Deep running starting from the bottom, 2017<br></strong><a href="https://www.amazon.com/Deep-running-starting-bottom-Korean/dp/8968484635">Deep running starting from the bottom (Korean Edition): Saito Goki: 9788968484636: Amazon.com: Books</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c6cde7cc95f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>