<?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 Arohan Paul on Medium]]></title>
        <description><![CDATA[Stories by Arohan Paul on Medium]]></description>
        <link>https://medium.com/@arohanpaul?source=rss-986b602bb823------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*r3N5q9Xa4mS6SejclOR68Q.jpeg</url>
            <title>Stories by Arohan Paul on Medium</title>
            <link>https://medium.com/@arohanpaul?source=rss-986b602bb823------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 09:10:21 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@arohanpaul/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[Newton’s Bucket Argument — What is it and why was it needed?]]></title>
            <link>https://medium.com/@arohanpaul/newtons-bucket-argument-what-is-it-and-why-was-it-needed-4095bb2b0f6e?source=rss-986b602bb823------2</link>
            <guid isPermaLink="false">https://medium.com/p/4095bb2b0f6e</guid>
            <dc:creator><![CDATA[Arohan Paul]]></dc:creator>
            <pubDate>Sun, 03 Aug 2025 16:09:45 GMT</pubDate>
            <atom:updated>2025-08-03T16:09:45.648Z</atom:updated>
            <content:encoded><![CDATA[<h3>Newton’s Bucket Argument — What is it and why was it needed?</h3><p>Before I start with this article, I will first answer the question, <em>Why was the Bucket argument needed?</em> It was needed to construct an irrefutable argument for the existence of an <em>inertial reference frame</em>. Now, let&#39;s see why that is the case.</p><p>When I first studied Newtonian mechanics, the most common expression I would hear was that there is no meaning to the term absolute motion. <em>All motion is relative</em> was the maxim to adhere to. We were taught not to talk of absolute velocity, only of relative velocity.</p><p>And then came this magical term called <em>inertial reference frames</em>. The term had me thinking for a long time. What is an inertial reference frame? A frame with respect to which Newton’s first law is valid,i.e., an object moving with uniform velocity in a straight line is unchanged (in terms of velocity and direction) without an external force.</p><p>The natural question that arises is that where are these reference frames located? And what makes them special and inertial? We know for sure that Earth is not an inertial reference frame. The Earth’s rotation on its axis makes it non-inertial.</p><p>These questions are difficult to answer, but Newton answered them in a very interesting way in his book <em>Mathematical Principles of Natural Philosophy. </em>Well, to put it straight, he answered the question of the existence of an inertial reference frame by just asserting the existence of one. He called it “absolute space”.</p><p>Did he give any experimental proof of absolute space? No. Did he suggest any experiment <em>that can be performed in practice</em> to prove the existence of such a space? No. Did he suggest any experiment to prove the existence of such a space? Yes, a thought experiment. And the bucket argument that is the title of this article is an argument (thought experiment) to prove the existence of an absolute space.</p><p>The argument goes as follows: <em>Imagine a bucket hanging from a rope is turned in such a way that the rope is fully twisted. Assume it is filled with water and kept at rest with respect to the bucket. Then, with a gentle force, whirl the entire bucket-water system in the opposite direction by untwisting the rope. Initially, the water and the bucket will be at rest with respect to each other. But as the motion of the bucket starts, it will slowly transfer its motion to the water. As the water also starts to rotate, it will begin to rise upward and assume a concave shape. Eventually, the motion of the bucket will be fully transferred to the water, and its speed of rotation will equal that of the bucket. It will remain concave, as it is rotating. The concavity of the water’s surface is due to its rotation. An interesting question that Newton asks here is: Rotation with respect to what? His answer is “Rotation with respect to absolute space”.</em></p><p>Let’s unpack the above argument a bit. How did Newton reach the conclusion of the existence of absolute space? His reasoning is as follows -</p><ol><li>At the beginning of the experiment, the water and the bucket were at rest with respect to each other, and the water’s surface was flat.</li><li>In the middle of the experiment, there is relative motion between the water and the bucket, and the water’s surface is beginning to concave.</li><li>At the end of the experiment, the water and the bucket are at relative rest with respect to each other (same speed of rotation), but the water is still concave.</li><li>This implies that the shape of the water’s surface has nothing to do with its state of rest or motion with respect to the bucket.</li><li>Recall that we did not imagine anything else around the bucket or water while describing the experiment. Hence, according to Newton, this experiment can be performed in an empty space. This implies that the concavity in the shape of the water’s surface is due to rotation with respect to “absolute space”.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4095bb2b0f6e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Learning a new concept directly from ChatGPT or any other LLM Model might be a bad idea]]></title>
            <link>https://medium.com/@arohanpaul/why-learning-a-new-concept-directly-from-chatgpt-or-any-other-llm-model-might-be-a-bad-idea-0fa79ca784d6?source=rss-986b602bb823------2</link>
            <guid isPermaLink="false">https://medium.com/p/0fa79ca784d6</guid>
            <category><![CDATA[dsa-training]]></category>
            <category><![CDATA[llm-usage]]></category>
            <dc:creator><![CDATA[Arohan Paul]]></dc:creator>
            <pubDate>Thu, 31 Jul 2025 06:53:16 GMT</pubDate>
            <atom:updated>2025-08-01T05:37:32.016Z</atom:updated>
            <content:encoded><![CDATA[<p>You might find it a bit strange that I am stating the conclusion in the title itself. The reason is that I am not comparing the pros and cons of learning a new concept using GPT and coming to a conclusion. Rather, I am making leaps and bounds in arguing in the negative with very little evidence.</p><p>This is because, when it comes to learning a new concept, there is no acceptable number of cons before you can say that the cons outweigh the pros. Rather, even a single con is enough to raise serious alarms. Learning a new concept is like laying the foundations of an upcoming building. We cannot tolerate a single wrong component.</p><p>Okay, enough philo talks. Let’s get straight to the recent occurrence that made me so whiny about GPT use for concept learning. I was studying Data Structures and Algorithms and wanted to check the logical correctness of a simple algorithm I had written. Let me give the credit in the beginning itself, I was going through GeeksforGeeks Python DSA course.</p><p>The problem was “<em>Given an array arr[] of size N containing positive integers and an integer X, find the element in the array which is smaller than X and closest to it.</em>”. I had a misconception while writing the solution, that in case X itself is present in the list, we will just return -1. It turns out that the problem asked to only print a number if it is <em>immediately</em> <em>less than X </em>in the list.</p><p>Anyways, I wrote the following code —</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/672/1*8-jh3JdR-JxKWeuPaMTUtg.png" /></figure><p>and asked the following question to ChatGPT -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/644/1*oPd5pbJ8_zAz5D6AAekHgg.png" /></figure><p>I asked to show problems in the above logic using examples. It generated many examples (images attached below), but the code worked in all these examples since the basic logic is actually correct.</p><p>Generated example -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nOaSCrbn8q9gikgkEiA3kA.png" /></figure><p>The code worked -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jKp97vG8kxJoNGszdDN5CA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m1zAwSoH2dY-N2Pc4JtojQ.png" /></figure><p>The code worked -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/963/1*wdHA8CGLU5YNSIy9wLkGIQ.png" /></figure><p>Generated Example —</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7991EmQOgvJAvSFziVwMNw.png" /></figure><p>The code worked again, but GPT bro changed the very laws of established mathematics to come to the following grand conclusion. It ends up saying in the last line that 13 lies between 14 and 15 just to prove me wrong —</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ducq4vBMbvv5wxNg553egw.png" /></figure><p>Apart from giving me a reason to giggle, it also sent me on this rant (the point of this whiny article) that these probabilistic models have no idea of a fundamental concept. They just keep generating tokens. If you have already learned a concept from <em>human </em>authors, it might be okay for you to revise or quickly recollect something using GPT. But if you are a student who is learning a concept for the first time, learning it using GPT might be damaging. You might get introduced to a lot of misconceptions that you aren’t even aware of.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0fa79ca784d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding Juggling Algorithm for Array Rotation By Group Theory]]></title>
            <link>https://medium.com/@arohanpaul/understanding-juggling-algorithm-for-array-rotation-by-group-theory-e2f8bf1a8467?source=rss-986b602bb823------2</link>
            <guid isPermaLink="false">https://medium.com/p/e2f8bf1a8467</guid>
            <dc:creator><![CDATA[Arohan Paul]]></dc:creator>
            <pubDate>Sat, 19 Jul 2025 13:53:00 GMT</pubDate>
            <atom:updated>2025-07-19T13:53:00.138Z</atom:updated>
            <content:encoded><![CDATA[<p>I was studying this algorithm called “juggling algorithm” for array rotation. It&#39;s a straightforward algorithm once you understand the idea behind it.</p><p>The code is straightforward, and it is available <a href="https://www.geeksforgeeks.org/dsa/juggling-algorithm-for-array-rotation/">here</a>, and it&#39;s very nicely explained.</p><p>What I have tried to do here is come up with a mathematical explanation of the code given above. Please read through the argument below after you have read the above code. If you understand it directly there itself, great! In case you don’t, you can go through my explanation. If you still have doubts or think I have made a mistake, please feel free to comment.</p><p>The rotation of an array of size n can be understood in terms of cyclic groups. Rotation by one place can be taken as the generator element, and the number of elements in this cyclic group is n, with the identity element being ‘no rotation’. If rotation by one place is denoted by ‘a’ then the cyclic group G can be written as G = ⟨a⟩ = {e, a, a^2, . . . , a^(n−1)}.</p><p>Rotation of the array by k places to the left can also be understood as the path of the generator element a while jumping ‘k’ steps at a time,i.e., the subgroup SG = ⟨a⟩ = {e, a, a^k, . . . }. The Number of elements in this subgroup is equal to the number of times we have to move by ‘k’ steps to reach the lowest common multiple of ’n’. Let’s say the number of elements in this subgroup is ‘m’, then ‘m’ will be given by the following equation — m⋅k=LCM(n,k). This implies m=LCM(n,k)/k.</p><p>Keep the above equation in mind, now we have to ask ourselves, how many such <em>distinct </em>cycles exist? The answer is given by the relation: number of cycles (nc) * m = n. Why I wrote the previously used symbol ‘m’ here? Because the length of each cycle is exactly the number of elements in the subgroup SG as discussed above. This gives us — nc*(LCM(n,k)/k)=n.</p><p>This implies nc*LCM(n,k)=n*k. We know that GCD(n,k)*LCM(n,k)=n*k.</p><p>This gives us the number of cycles ‘nc’ as GCD(n,k). Once we establish the number of <em>distinct </em>cycles present, the code given above is easy to follow.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e2f8bf1a8467" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Logistic Regression Cost Function]]></title>
            <link>https://medium.com/@arohanpaul/logistic-regression-cost-function-9d1e2f091b7d?source=rss-986b602bb823------2</link>
            <guid isPermaLink="false">https://medium.com/p/9d1e2f091b7d</guid>
            <category><![CDATA[supervised-learning]]></category>
            <category><![CDATA[binary-classification]]></category>
            <category><![CDATA[logistic-regression]]></category>
            <category><![CDATA[cost-function]]></category>
            <dc:creator><![CDATA[Arohan Paul]]></dc:creator>
            <pubDate>Mon, 14 Oct 2024 05:18:06 GMT</pubDate>
            <atom:updated>2024-10-20T05:02:32.608Z</atom:updated>
            <content:encoded><![CDATA[<h4>A quick explanation</h4><p>Let’s start by writing out the monster :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/615/1*B_QEM8otFi35CMnWNHwHyA.png" /></figure><p>It looks like an unexplainable monster, but trust me when you break it down, it appears cute and simple. To understand this, let’s split it into two parts -</p><p>Before we delve into each part, let’s first make a very simple observation. Part A(see the two parts given below) is useful only if the actual label(y(i)) is <strong>1</strong>. Part B is useful when the actual label (y(i)) is <strong>0. </strong>You can check this by making a very simple calculation of substituting <strong>0 </strong>and <strong>1</strong> in Part A and Part B respectively.</p><p>PART A</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/286/1*aZ3E2TqYdhVAr5BRfxcZPg.png" /></figure><p>Let’s delve into part A first. As mentioned above, this part is useful only when the actual label is 1. Let’s see how it behaves for the two cases, i.e., when the prediction matches the actual label and the one where it doesn’t.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/558/1*ybwr44ahBUFb0Q3ybDj6dA.png" /></figure><p>As we can see from the above table, the cost function is as we would expect it to be.</p><p>PART B</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/442/1*HmmkGiR3dfS7aGrFMqcaKA.png" /></figure><p>Let’s delve into part B now. As mentioned above, this part is useful only when the actual label is 0. Let’s see how it behaves for the two cases, i.e., when the prediction matches the actual label and the one where it doesn’t.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/543/1*CFzSi8qW8OecL3lAunfmlQ.png" /></figure><p>As we can see from the above table, the cost function is as we would expect it to be.</p><p>After we are clear with these two major (note monstrous looking :) ) parts of the equation, the rest of it is quite clear. The summation from i =1 to j and the division by m just gives the averaging over m training examples.</p><p>And the negative sign at the beginning exists to convert the negative values received from the log functions in parts A or B, since the log is negative for values between 0 and 1.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9d1e2f091b7d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Kalman Filters]]></title>
            <link>https://medium.com/@arohanpaul/kalman-filters-de2097a6affa?source=rss-986b602bb823------2</link>
            <guid isPermaLink="false">https://medium.com/p/de2097a6affa</guid>
            <dc:creator><![CDATA[Arohan Paul]]></dc:creator>
            <pubDate>Wed, 18 Sep 2024 10:15:58 GMT</pubDate>
            <atom:updated>2024-09-18T10:15:58.740Z</atom:updated>
            <content:encoded><![CDATA[<h4>What are they and how to use them</h4><p>I came across Kalman Filters while working on a time-series problem. Although there is plenty of theoretical material available online, I found it difficult to pinpoint the exact algorithm to solve my problem.</p><p><em>Hence, my blog is very</em> <em>application-focused. My aim is to quickly take you to the algorithm and show it’s working.</em></p><p>Before delving into Kalman filters, let’s first create a hypothetical scenario that resembles the kind of problems where Kalman filters would be applicable.</p><p>Imagine a train leaving a station with a fixed acceleration. Our task is to continuously access the current position of the train after pre-defined fixed intervals of time (e.g. 1 hour). Also, let’s consider we have some kind of an enormous scale along the tracks to measure the train’s position at each instant of time. By train’s position, we mean the position of its engine.</p><p>To estimate the train’s position, we have two different tools at our disposal.</p><p>1. Prediction of its position through equations of motion.</p><p>2. Measurement of its position using our hypothetical long scale along the tracks.</p><p>Both of the ways of getting the current position will introduce some errors.</p><p>In the case of prediction through equations of motion, the error would be introduced by the bias or simplicity in our assumptions that lead us to those equations. For example, in this case, we are not considering the resistance from the tracks in our equation and many other factors that might influence the position of the train.</p><p>In the same way, measurement has its own associated errors, such as those introduced by the limitations of the measurement scale.</p><p>We obtain two sets of distance values: one from the equation and one from measurement, each with its own errors. Our task is to estimate the true distance using these two sets. The Kalman Filter helps by averaging these errors to provide a more accurate estimate of the quantity we’re interested in — in this case, the distance.</p><p>The example that I took to give a feel for Kalman Filters was chosen for a specific reason namely that the equation of motion that comes from the above situation falls into a class of equations we call <em>Linear Discrete Dynamical Systems </em>which is a pre-requisite for the application of Kalman Filters.</p><p>To see how we get a linear discrete dynamical system from the above example, let’s start by working out some equations -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/724/1*_6cPO8tvqlqMmkUS_uZvpA.png" /></figure><p>If we take the time steps to be a standard value let’s say 1, we get</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/398/1*gs8JY5ysTStZhhIxszi-9Q.png" /></figure><p>which means the distance and velocity in the next time step are dependent on the distance and velocity in the current time step. The matrix form of the equations is given below -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/585/1*04g8xUxFG00mAGMcGaP34Q.png" /><figcaption>Equation 1</figcaption></figure><p>Let’s keep this equation aside for a while and take a straight dive into the theoretical heart of Kalman Filters. We will come back to this equation after taking a detour from Kalman theory and then use that theory to solve the above problem.</p><p>Kalman Filters are used to estimate the states of a linear discrete dynamical system where the state to be estimated follows an equation of type -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/478/1*OWKBgd5c6WY36IWNz13GfA.png" /></figure><p>with a measurement</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/275/1*n9-dymGd43LIokAM1_K5_Q.png" /></figure><p>The matrices A and B are assumed to be constant over varying k. The variable u over all k is known as the control variable (which we won’t delve into here). z is the term we are measuring which is linearly related to our desired variable x by a coefficient matrix H. One way to understand z with our train example is if instead of measuring the distance (our desired variable), we would rather measure the time taken by light to reach us from the train. I know it’s almost impossible to measure that kind of small time interval. I am saying this just to give a sense of the term z.</p><p>w and v are prediction and measurement errors respectively which we will assume to be gaussian in most cases.</p><h3><strong>Algorithm</strong></h3><p>The algorithm works similarly to numerical methods for differential equations, like <em>predictor-corrector</em> methods. Starting with an initial estimate x0, we iterate step by step to estimate xk​ from the previous value xk−1​. I’ll first directly introduce the algorithm used in real-world problems and then explain how it works, so stay with me.</p><p><strong>Time Update</strong></p><ol><li>Project the state ahead -</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/326/1*YEONZiUMS9GKDLdc3rhXag.png" /></figure><p>2. Project the error covariance ahead -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/336/1*JYQahAMGK1djO9JgSd8MOQ.png" /></figure><p><strong>Measurement Update</strong></p><ol><li>Compute the Kalman Gain-</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/479/1*nrcxyflWB3bE-7N-GADgcQ.png" /></figure><p>2. Update the estimate -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/421/1*Cd0zl6K0-wiDYppoXpEtOQ.png" /></figure><p>3. Update the error covariance -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/345/1*I52a7qEDUBViKWYCknLFsw.png" /></figure><p>Among all the terms used in the 5 equations above, three of them have not been described before -</p><p>Q -&gt; Variance of the process error or error arising from equation predictions, i.e.,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/268/1*tQTkq8Jt_4qlwI6VjFQkEg.png" /></figure><p>R -&gt; Variance of the measurement error, i.e., errors arising from measurements</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/252/1*sOtgssPr9zVoZ276sxIfBw.png" /></figure><p>$\bar{P}$ and P -&gt; Covariance matrices of the equation and measurement errors respectively.</p><h4>Application of the Algorithm</h4><p>We can apply the above algorithm to multiple real-life problems. In the train example described above (<strong>equation 1</strong>), we have two unknowns, position and velocity. And since we are measuring the distance directly (using our imaginary scale), the measurement equation becomes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/365/1*rlwjEjbqsqjtvnw4bqrW7Q.png" /><figcaption>Equation 2</figcaption></figure><p>From equation 1 above, we have</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*YVZYWINlRtt2HtPk8bSnwA.png" /></figure><p>From equation 2, we have</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/194/1*9SFqJ7BZywUdwQ7RB1xypQ.png" /></figure><p>For simplicity, let’s ignore the process term uk, and its associated noise variance term Q, which implies B and Q are zero. Assume R which is the variance of the measurement error as 0.5 consistently, i.e. R=0.5.</p><p>After the coefficients are set, let’s decide on our initial conditions, i.e.,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/191/1*dmy3Nslc3plBlsqhztQGUg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/229/1*It9MByM4dV4-RH5_mXMrhw.png" /></figure><p>Let’s take a dummy set of observations of distance given in the below table -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mfDbTTIui0U_ayxbnh4LRQ.png" /></figure><p>Let me confess that the values given in the above assumptions and table are completely arbitrary and don’t reflect any real scenario (may not even make any numerical sense). Given the algorithm, initial conditions, coefficients, and observations, we can successively calculate the distance estimate by finding the kalman gain (K) through successive implementation of the algorithm.</p><h4>References</h4><ol><li><a href="https://www.cs.cmu.edu/~motionplanning/papers/sbp_papers/kalman/kleeman_understanding_kalman.pdf">Understanding and Applying Kalman FIltering</a></li><li><a href="https://www.cs.unc.edu/~welch/media/pdf/kalman_intro.pdf">An introduction to the Kalman Filter</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de2097a6affa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Very Short History of Artificial Intelligence]]></title>
            <link>https://medium.com/@arohanpaul/development-of-artificial-intelligence-e544477c5f37?source=rss-986b602bb823------2</link>
            <guid isPermaLink="false">https://medium.com/p/e544477c5f37</guid>
            <dc:creator><![CDATA[Arohan Paul]]></dc:creator>
            <pubDate>Thu, 12 Oct 2023 06:03:36 GMT</pubDate>
            <atom:updated>2024-09-01T18:29:57.831Z</atom:updated>
            <content:encoded><![CDATA[<p>The idea of intelligent machines is not new. The idea itself is neither an invention of the Industrial Revolution nor an emergent phenomenon from rapid digitization. People in ancient Greece who predate us by thousands of years have thought and attempted to build machines that are ‘intelligent’ in some sense but obey human orders.</p><p>But the specific origins of the innovative ideas that led to the current boom in Machine Learning can be traced back to the seventeenth century to people like Leibniz, Thomas Hobbes, Rene Descartes, and, many more. Their main focus was to theorize about the formation and mechanism of rational thought processes. They came up with the idea that rational thought generation in the mind is a computational process in the sense that it is the emergent property of formal systems. Leibniz went as far as envisioning a universal language system that reduces logical arguments to formulas.</p><p>A formal system is a collection of symbols with rules to manipulate them. Some basic associations of symbols are taken as ‘true’ statements and are called axioms. And the rest of the ‘true’ statements are derived from the rules. The systematic study of formal systems is called Mathematical Logic, a field that grew rapidly in the nineteenth and twentieth centuries.</p><p>Drawing on the thoughts of Leibniz, Descartes, and others, and from the advancement in mathematical logic, Allen Newell and Herbert Simon formulated the Physical Symbol System Hypothesis (PSSH) which states that a formal system has the necessary and sufficient conditions to generate intelligent action. PSSH had both admirers and bitter critics.</p><p>A famous critic of the PSSH came from John Searle in what is known as the Chinese Room Thought Experiment. The crux of the argument raised in the thought experiment was that a computer program executing instructions or a formal system performing symbol manipulations cannot have an ‘understanding’ of the processes in a way that can be compared to the ‘understanding’ possessed by human beings.</p><p>A parallel strand of thought that paved the way for modern AI is called connectionism. Neural Networks and their popularity come from this school of thought. Unlike PSSH, connectionism posits that thinking is a result of connected structures and signal processing among them. The connection strength can be measured by weights and updated using learning algorithms.</p><p>The debate between symbolic and connectionist AI is still raging. Both the methods have their own pros and cons. While symbolic AI provides great explainability, they have limited learning power. On the other hand, connectionist approaches are built on learning a model from data, but they lack explainability. Regardless of the debate, it is true that in applications both approaches are used extensively being the dominant powers in their own arenas.</p><p>To conclude, in this write-up, I have attempted to give a brief pathway of the development of Artificial Intelligence. I spent more time exploring the development of the symbolic approach as opposed to the connectionist approach as it has a longer history of development. It would be interesting to see how this age-old attempt to mimic the human mind and develop intelligent machinery develops over time. Will one of the two approaches overtake the other, or will they work in cooperation, or is there a third yet undiscovered way? Only time will tell.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e544477c5f37" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>