<?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 Mircea Florescu on Medium]]></title>
        <description><![CDATA[Stories by Mircea Florescu on Medium]]></description>
        <link>https://medium.com/@mrc.florescu?source=rss-1013853f1f59------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*6MBrcaUqYLfIMO4E</url>
            <title>Stories by Mircea Florescu on Medium</title>
            <link>https://medium.com/@mrc.florescu?source=rss-1013853f1f59------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 13:55:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mrc.florescu/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[Humanity’s Last Revolution]]></title>
            <link>https://medium.com/@mrc.florescu/humanitys-last-revolution-d9372d737eab?source=rss-1013853f1f59------2</link>
            <guid isPermaLink="false">https://medium.com/p/d9372d737eab</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[humanitys-future]]></category>
            <category><![CDATA[ai-revolution]]></category>
            <dc:creator><![CDATA[Mircea Florescu]]></dc:creator>
            <pubDate>Sun, 02 Feb 2025 20:23:43 GMT</pubDate>
            <atom:updated>2025-02-10T09:19:03.400Z</atom:updated>
            <content:encoded><![CDATA[<p>Throughout history, humanity has been defined by its revolutions -transformative periods when the trajectory of our species fundamentally changed. These pivotal turning points have shaped not just how we live, but who we are as a species. From the earliest Oldowan stone tools (more than 2.5 million years ago) to the digital age, each revolution has been driven by human ingenuity, creativity, and determination.</p><blockquote><em>But now, we stand at the threshold of what may be humanity’s last self-driven revolution: the rise of Artificial Intelligence (capital letters used intentionally).</em></blockquote><h4>The anatomy of revolution</h4><p>A revolution, in its essence, is a fundamental shift that transforms society’s foundations altering the very fabric of human existence by rapid, profound changes. They rewrite the rules of what’s possible and reshape our relationship with the world around us.</p><h4>The great human revolutions</h4><p>Our journey as a species has been marked by several transformative revolutions, as documented by leading thinkers in human development (Diamond, 1997; Harari, 2014; Pomeranz, 2000).</p><p>The <strong>cognitive revolution</strong> (70,000 BCE) transformed us from mere primates into storytellers and dreamers, then the<strong> agricultural revolution</strong> (10,000 BCE) transformed nomadic hunter-gatherers into settlers, leading to the birth of civilization itself. For the first time, humans could exert control over their food supply, leading to population growth and the emergence of complex societies. The<strong> scientific revolution</strong> (16<em>ᵗʰ-</em>17<em>ᵗʰ</em> centuries) challenged ancient authorities and established the foundations of modern knowledge, changing how we understand the universe and our place within it. Then the <strong>industrial revolution</strong> (18<em>ᵗʰ</em>-19<em>ᵗʰ</em> centuries) came with its mechanized production, urbanized society, creating unprecedented material wealth. It transformed not just how we work, but how we live, leading to modern cities and global commerce.</p><p>Not too long ago, the <strong>digital revolution</strong> (20<em>ᵗʰ</em> century) connected humanity through global networks of information, democratizing knowledge and enabling instant global communication. It laid the groundwork for what was to come.</p><h4>The AI revolution and the end of human-driven progress</h4><blockquote><strong>November 30, 2022</strong>, marks a different kind of revolution. With the public release of ChatGPT and the dawn of generative AI, humanity crossed a threshold unlike any before. This revolution is unique because it represents the first time we’ve created something that can potentially surpass our own cognitive capabilities.</blockquote><p>The <strong>AI revolution</strong> is fundamentally different through its self-improving nature: unlike previous technologies, AI systems can learn and improve themselves, potentially at rates far exceeding human capability. Previous revolutions gave us tools that remained tools — a steam engine never became an autonomous entity that could design better steam engines.</p><p>Before reaching the era of Artificial General Intelligence (AGI), where AI systems will autonomously drive discoveries and technological breakthroughs without human intervention, we live in a brief window in human history when progress emerges from an unprecedented collaboration between human and artificial intelligence. This temporary partnership phase represents a crucial bridge between human-led progress and the future era of pure AI-driven advancement. As AI systems become increasingly sophisticated, they will gradually assume the primary role in scientific discovery and technological innovation. Eventually, human participation in breakthrough research and development will become unnecessary, as AI systems will possess capabilities far surpassing human cognitive limits and research abilities.</p><blockquote>The AI revolution represents humanity’s last solo act on the stage of progress. From this point forward, technological and scientific advancement will be a collaborative endeavor between human and artificial intelligence, with AI increasingly taking the lead in many areas. This isn’t a dystopian prediction, but rather a recognition of a fundamental shift in how progress will occur.</blockquote><p>The history is now clearly divided into two eras: before and after November 30, 2022. In the pre-AI era, human progress was limited by our biological cognitive capabilities. In the post-AI era, the limits of progress are being redefined by the combination of human creativity and AI capabilities.</p><h4>Looking forward: the human role in an AI-driven future</h4><p>Does this transition diminish humanity’s importance? I like to believe that rather it transforms our role - we become the architects of our own succession, the creators of our intellectual heirs. Our challenge now is not just to push the boundaries of knowledge ourselves, but to guide and direct the artificial intelligences we’ve created to ensure they serve humanity’s best interests.</p><p>We will have to learn these skills so the next AI-driven revolutions won’t be the end of human progress (and the start of machines’ progress) but a partnership between human and machine intelligence in which human wisdom, ethics, and guidance still plays an important role.</p><p>As we stand at this threshold, we must recognize that we are not just witnesses to history - we are its architects. If we really understand what this means, the decisions we make now, as we navigate this transition, will shape not just our future, but potentially the future of all intelligence and life, either human or artificial, in our corner of the universe.</p><p>Our last revolution may be our greatest achievement: creating the tools that will help us transcend our own limitations and open up possibilities we can barely imagine. The question is no longer what humans alone can achieve, but what humans and AI together can accomplish.</p><p>In this new era, our greatest challenge - and opportunity - is to ensure that this partnership between human and artificial intelligence serves to enhance rather than diminish our humanity, creating a future that honors both our past achievements and our future potential, and is not just a future of the machines.</p><p><em>References:<br>Diamond, J. M. (1997). Guns, germs, and steel: the fates of human societies. New York, W.W. Norton &amp; Co.<br>Harari, Y. N. (2014). Sapiens: A brief history of humankind. Random House.<br>Pomeranz, K. (2000). The great divergence: China, Europe, and the making of the modern world economy.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d9372d737eab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI training without labels: self-supervised vs. supervised learning]]></title>
            <link>https://medium.com/@mrc.florescu/ai-training-without-labels-self-supervised-vs-supervised-learning-8042409e829d?source=rss-1013853f1f59------2</link>
            <guid isPermaLink="false">https://medium.com/p/8042409e829d</guid>
            <category><![CDATA[supervised-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[self-supervised-learning]]></category>
            <dc:creator><![CDATA[Mircea Florescu]]></dc:creator>
            <pubDate>Tue, 08 Oct 2024 06:46:10 GMT</pubDate>
            <atom:updated>2024-10-08T06:46:10.012Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Supervised </strong>and <strong>self-supervised</strong> learning both play vital roles in machine learning, each suited to different data constraints and learning goals.</p><blockquote>Supervised learning is the go-to for tasks where labeled data is plentiful and specific predictions are needed. Self-supervised learning excels in scenarios where labeled data is scarce, allowing models to extract meaningful representations from data itself.</blockquote><p>Though self-supervised learning shares structural similarities with supervised learning, its unique ability to generate labels from data sets it apart. This flexibility is essential for models to learn autonomously, especially in fields requiring large-scale understanding of data without extensive human intervention.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/697/1*DzfGVKJxXNlIjOq5SpTs9w.jpeg" /><figcaption>Self-supervised: robot extracting meaningful representations from data itself. (DALL-E generated)</figcaption></figure><h4>Supervised learning, a traditional approach</h4><p>Supervised learning is perhaps the most well-known and widely used form of machine learning. In supervised learning, an AI model is trained on a dataset that contains <strong>labeled data</strong>, where each input has an associated output label provided by humans. The model’s goal is to learn a mapping between inputs and their corresponding labels so it can accurately predict the labels for new, unseen data.</p><p>For example, consider training an AI model, such as a neural network, to classify images of animals. In this scenario, the model is given thousands of labeled images, such as “cat,” “dog,” and “bird.” At the beginning of the training process, the model’s algorithms are far from perfect, often producing inaccurate predictions. These initial predictions are then compared to the actual labels in the dataset to calculate an error. Using this error, the model adjusts its parameters through a process called backpropagation, iteratively refining its algorithms to improve accuracy over time. This cycle continues until the model can reliably classify new images.</p><h4>Self-supervised learning: labels from the data itself</h4><p>Self-supervised learning is a subfield of unsupervised learning, but with an interesting twist: it creates labels from the data itself. Rather than relying on external labels provided by humans, self-supervised learning generates <strong>pseudo-labels</strong> based on the structure or content of the data. This approach allows the model to learn from the data without needing manually labeled datasets.</p><p>For instance, in natural language processing (NLP), a self-supervised model might be trained on a book’s text, where the model learns to predict the next word in a sentence based on the preceding words. Here, each word effectively serves as the label for the word before it. In this way, the model can learn language patterns and context without needing any external labeling.</p><h4>Can self-supervised learning be considered supervised?</h4><p>An interesting perspective arises when we consider the structure of self-supervised learning. Since it involves input-output pairs, where the model minimizes an error between its predictions and an automatically generated label, self-supervised learning can resemble supervised learning in several ways:</p><ul><li><strong>Input-Output structure</strong>: both supervised and self-supervised learning involve making predictions based on inputs and comparing those predictions to a target. The difference lies in how the target is obtained. In supervised learning, targets are given by humans, while in self-supervised learning, they are derived directly from the data.</li><li><strong>Minimizing error:</strong> in supervised learning, error minimization is based on discrepancies between the model’s predictions and human-provided labels. In self-supervised learning, the error is computed based on how well the model predicts one part of the data (like the next word) given another part. This allows the model to learn useful representations from the data itself.</li><li><strong>Iterative learning</strong>: both methods rely on iterative optimization (often through backpropagation) to minimize error over time. Self-supervised learning, like supervised learning, benefits from this process, gradually improving its predictions by learning from its own internally generated targets.</li></ul><blockquote>In essence, many in the field of AI would agree that <strong>self-supervised learning can be thought of as a form of supervised learning</strong> — albeit with a twist. While self-supervised learning might not have the manually labeled target typically associated with supervised learning, it uses the data to create an <strong>internally generated target</strong>, making it resemble supervised learning in structure and process.</blockquote><p>Thus self-supervised learning can be seen as a bridge between supervised and unsupervised learning, combining the structure of supervised methods with the flexibility and scalability of unsupervised approaches.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8042409e829d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artificial Neural Networks: The Training Process Simplified]]></title>
            <link>https://medium.com/@mrc.florescu/artificial-neural-networks-the-training-process-simplified-4ed1a153e869?source=rss-1013853f1f59------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ed1a153e869</guid>
            <category><![CDATA[artificial-neural-network]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Mircea Florescu]]></dc:creator>
            <pubDate>Thu, 03 Oct 2024 04:28:55 GMT</pubDate>
            <atom:updated>2024-10-03T04:28:55.139Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/889/1*owlgrBIMkHdBerqyjCf68Q.jpeg" /><figcaption>Lost in Fog on a Hilly Terrain. (Dall-E generated)</figcaption></figure><p><strong>Artificial Neural Networks</strong> (ANNs) are powerful algorithms used to make predictions in various applications, from predicting the next word in a sentence to identifying objects in an image. While the underlying mechanics can seem complex, in this article I will explain, in simple terms, how an algorithm like an ANN automatically builds (learns) by adjusting its parameters until its predictions are good enough.</p><blockquote>The training process of an ANN involves the realization of a complex algorithm through repeated trials. These trials aim to discover the best elements of the algorithm that will make good predictions.</blockquote><p><strong>Training Data:</strong><br><strong> </strong>The process is based on a dataset that contains both the input data and the correct outputs, embodying the logic of good predictions.</p><p>This dataset can be likened to a book, where the words represent individual data points, and the order of words represents the underlying logic or patterns. For instance, if the task is to predict the next word in a sentence, the input might be a word from the book, and the output would be the word that typically follows it. During training, the algorithm repeatedly adjusts its internal parameters until it becomes a complex model capable of making accurate predictions when given new inputs.</p><p><strong>Starting with Random Parameters:</strong> <br>Initially, the algorithm doesn’t “know” the correct relationships between inputs and outputs. It starts with <strong>randomly assigned parameters</strong> (also called weights). These parameters determine how the algorithm processes the input data and generates a prediction. Since the parameters are random at first, the predictions will usually be inaccurate in the beginning.</p><p><strong>Making a Prediction:</strong> <br>The algorithm processes the input data through its series of mathematical operations using the current set of parameters. This step is often called the <strong>forward pass</strong>, where the algorithm takes the input, applies its current parameters, and outputs a prediction. In our example, this would be predicting the next word in the sequence. However, because the parameters are initially random, the algorithm’s prediction is likely wrong at the start.</p><p><strong>Calculating the error:</strong> <br>After the algorithm makes a prediction, it compares the result with the actual, correct output (for example, the correct next word from the book). The difference between the predicted output and the actual output is called the <strong>error</strong> or <strong>loss</strong>. This error represents how far off the algorithm was in making its prediction.</p><p><strong>Adjusting the parameters: <br></strong>Once the error is known, the algorithm uses a process called <strong>backpropagation </strong>to adjust its parameters in a way that reduces the error.</p><p>Backpropagation doesn’t provide a one-time fix to eliminate all error instantly. Instead, it calculates how much each parameter contributed to the overall error. The algorithm uses this information to make small adjustments to the parameters in a way that should reduce the error for the next prediction. These adjustments are based on the <strong>gradients</strong>, which tell the algorithm in which direction (increase or decrease) to tweak each parameter.</p><p><strong>Taking small steps:</strong> <br>The algorithm doesn’t overhaul its parameters all at once. It makes <strong>small, incremental changes</strong> to ensure that it doesn’t overshoot and increase the error instead of reducing it. These small adjustments are guided by a process called <strong>gradient descent</strong>, which ensures that the parameters are changed in a direction that leads to a lower error in future predictions.</p><blockquote>Think of this process as walking down a hill toward the lowest point, where the error is minimal. The algorithm uses the gradient (slope) to decide which direction to step, but it takes small steps to avoid jumping too far and potentially missing the optimal point.</blockquote><p><strong>Repeating the process (iterative adjustments): <br></strong>The key to the learning process is repetition. After each prediction, the algorithm calculates the error, adjusts the parameters using backpropagation, and then repeats the process with the next batch of data. Each cycle of this process is called an <strong>epoch</strong>.</p><p>The reason the algorithm doesn’t find the best parameters after just one cycle is that the <strong>error function</strong> — the mathematical relationship between the parameters and the error — is usually very complex. It may have many peaks and valleys, and the algorithm needs to iteratively refine its parameters to approach the lowest possible error over time.</p><p><strong>Guided Adjustments Instead of Random Changes: <br></strong>What makes this process efficient is that the algorithm is not adjusting its parameters randomly each time. Instead, the parameter updates are <strong>guided by the gradients</strong>, which are calculated based on the error. These gradients provide a logical and efficient way to modify the parameters, moving the algorithm closer to a solution that minimizes the error.</p><p>Rather than trial and error, the algorithm uses the <strong>mathematical information</strong> provided by backpropagation and gradient descent to make well-informed adjustments at each step.</p><p><strong>Convergence:</strong><br>Over many iterations, the algorithm gradually reduces the error by continually adjusting its parameters. Eventually, it reaches a point where the error cannot be reduced much further, or the improvements become very small. This is when the algorithm has <strong>converged</strong> on an optimal or near-optimal set of parameters for making accurate predictions.</p><blockquote>Training an artificial neural network involves discovering the optimal algorithm starting with random parameters, making predictions, calculating the error between the predicted and actual output, and then gradually adjusting the parameters in small steps to reduce the error. The process of backpropagation ensures that these adjustments are not random but are guided by the mathematical relationship between the error and the parameters. Through this iterative process, the algorithm continuously improves its ability to make accurate predictions, ultimately converging on a set of parameters that minimize the error effectively.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ed1a153e869" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning vs Deep Learning. An Introduction.]]></title>
            <link>https://medium.com/@mrc.florescu/machine-learning-vs-deep-learning-an-introduction-44df92c1195d?source=rss-1013853f1f59------2</link>
            <guid isPermaLink="false">https://medium.com/p/44df92c1195d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Mircea Florescu]]></dc:creator>
            <pubDate>Tue, 24 Sep 2024 11:02:45 GMT</pubDate>
            <atom:updated>2024-09-24T12:43:17.650Z</atom:updated>
            <content:encoded><![CDATA[<p>Thanks to ChatGPT, artificial intelligence (AI) has become a widely recognized technology, and its popularity has also introduced new concepts that many struggle to understand. This article explains in simple terms a couple of important concepts that may cause confusion for anyone who’s not an AI expert.</p><p><strong>MACHINE LEARNING</strong> (ML) is <strong>an automated way by which computers learn from data</strong>. It is not fully autonomous, but rather <strong>semi-autonomous</strong>. Only after humans explain the computer what (data) and how (algorithms and parameters) to learn, will it learn by itself. Humans need to prepare the data, feed the computer with data, and supply the algorithms that the computer will use to learn. After these are in place, the machine learns from the data by adjusting elements of the algorithms, which constitutes the core automated aspect of ML. These adjustments that lead to learning and predictive capabilities are made without further human intervention — hence, it is “semi-autonomous.”</p><p>Thus, machine learning refers to methods that enable computers to learn, which is <strong>identify patterns</strong> in data and <strong>make predictions or decisions</strong> <em>based on data, without being explicitly programmed for specific tasks</em>.</p><p>The stages of the ML process in which <strong>human intervention</strong> is required can be split in stages before the computer starts to learn and stages related to the quality of the learning process:</p><ol><li>Before learning:</li></ol><ul><li><strong>Data preparation</strong>: Humans must collect, clean, and preprocess the data that will be used by computers to learn.</li><li><strong>Algorithm selection</strong>: Humans choose the appropriate learning algorithm based on the problem type (e.g., classification, regression, clustering).</li><li><strong>Feature engineering</strong>: Humans often decide which features (attributes) of the data are relevant for the model.</li></ul><p>2. Quality control:</p><ul><li><strong>Model evaluation and tuning</strong>: After the initial training, humans often need to evaluate the model’s performance, fine-tune hyperparameters, and possibly retrain with different data or techniques.</li><li><strong>Bias and fairness considerations</strong>: Humans have to make sure that the model does not learn biased patterns or make unfair decisions.</li><li><strong>Deployment and Monitoring</strong>: In real-world applications, models need to be monitored by humans for performance degradation over time (due to changing data patterns).</li></ul><p>While it’s true that humans choose and implement the algorithms, it’s important to note that many machine learning models are general frameworks that can adapt to a wide variety of data and problems. The specifics of the learning process, such as which patterns are learned, are not manually programmed but instead discovered by the model through training.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tHq_yfRxSG1lVciUkXPgKQ.jpeg" /><figcaption>Guess who’s DL and who’s ML in this picture. (DALL-E generated)</figcaption></figure><p><strong>DEEP LEARNING</strong> (DL) is <strong>a subset of machine learning</strong> in which the model can <strong>automatically learn features from raw data</strong>, reducing the need for manual data engineering. This can make the process feel more autonomous than traditional ML methods where data feature selection and engineering are more manual. Deep Learning is particularly effective for problems involving unstructured data like images, audio, and text while traditional ML algorithms are more commonly applied to structured data (e.g., tables of numerical or categorical data).</p><p><strong>Deep Learning</strong> uses very complex architectures of algorithms, like neural networks with many layers. They typically <strong>require large amounts of data</strong> to perform well, because they have a vast number of parameters to learn, and more data helps improves generalization. Traditional ML algorithms like decision trees or linear models can often work with smaller datasets.</p><p>In traditional ML, selecting and transforming data is often a crucial step that requires humans to understand the data and the domain that produced the data. Deep Learning, on the other hand, automates much of this process through its layers of abstraction, which is why it needs more data to learn effectively.</p><p>Also, Deep Learning systems, such as deep neural networks and transformers, are often referred to as “<strong>black boxes</strong>” because their internal workings are complex and <strong>difficult to interpret</strong>. Designing and optimizing these models requires a deep understanding of both theory and practice, making it somewhat of an art form as well as a science.</p><p>Common or traditional Machine Learning algorithms like clustering (for grouping data points), linear regression (for predicting continuous variables), logistic regression (for binary classification), SVM (for classification and regression), decision trees, and random forests (for classification and regression) are generally more interpretable because their decision-making process can be traced back to mathematical relationships between input variables. These models are more straightforward to understand and explain compared to Deep Learning models. This is the reason why they are often used in applications where interpretability is crucial. One side note here: some complex ensemble Machine Learning methods like Random Forest or certain SVM configurations can also become difficult to interpret, though generally less so than deep neural networks.</p><p><strong>CONCLUSION</strong>. While both Machine Learning and Deep Learning are powerful tools in the field of Artificial Intelligence, they each have their strengths and optimal applications. Machine Learning offers a wide range of algorithms that can be applied to various types of data, often with good interpretability, making it suitable for many business and scientific applications where understanding the decision-making process is crucial. Deep Learning is capable to automatically extract features from complex, unstructured data, and excels in tasks involving image recognition, natural language processing, and other areas where traditional feature engineering is challenging. The choice between ML and DL depends on the specific problem at hand, the available data, and the need for model interpretability.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=44df92c1195d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>