<?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 Enise İrem Colak on Medium]]></title>
        <description><![CDATA[Stories by Enise İrem Colak on Medium]]></description>
        <link>https://medium.com/@eniseirem?source=rss-d9e4918777c1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*2eAzcpXkMsV3AccJ-lNHFw.jpeg</url>
            <title>Stories by Enise İrem Colak on Medium</title>
            <link>https://medium.com/@eniseirem?source=rss-d9e4918777c1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 15:16:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@eniseirem/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[Explainable AI: How Models Think? — III (Relevance Propagation)]]></title>
            <link>https://medium.com/@eniseirem/explainable-ai-how-models-think-iii-relevance-propagation-5bb26a5405ba?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/5bb26a5405ba</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 07:31:01 GMT</pubDate>
            <atom:updated>2026-11-10T07:31:01.739Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Part 3.4 — The Toolbox: Relevance Propagation</em></p><p>In the previous part, we used CAM and Grad-CAM to see where a model focuses during a prediction.</p><p>Let’s imagine a building with several floors and many radiators. At the top, we observe the final temperature. The natural question is no longer where the heat is, but how much each radiator contributed to that final temperature. We can trace the heat backward, floor by floor, and assign part of the final temperature to the radiators that contributed to it. Some receive a larger share, while others receive very little.</p><p><strong>Layer-wise Relevance Propagation (LRP)</strong> follows a similar idea. It starts from the model’s output and redistributes the prediction backward through the network, layer by layer, until the relevance reaches the input features.</p><p>This changes the question from</p><p><em>CAM / Grad-CAM “Where is the model focusing?”</em></p><p>to</p><p><em>LRP “How much did each input feature contribute to this prediction?”</em></p><p><strong>What is LRP?</strong></p><p>So how does this work inside the network? Let’s keep the building example for one more step.</p><p>Assume the final temperature at the top of the building is <strong>100</strong>. That number is our starting point. Now we move backward and ask how much of that 100 came from the radiators on the floor below. A radiator that contributed more heat should receive a larger share. A radiator that contributed less should receive a smaller one.</p><p>LRP does the same with a prediction.</p><p>It starts from the model output and gives that output a relevance value. Then this relevance is passed backward through the network, one layer at a time.</p><blockquote>Output → Hidden Layers → Input</blockquote><p>At each layer, the relevance is redistributed between the neurons that contributed to the next layer.</p><p>So if a neuron played a stronger role in producing the prediction, it receives more relevance. If its contribution was small, it receives less. We usually write the relevance of a neuron as <em>Rj</em>​.</p><p>By the time this process reaches the input layer, each input feature has its own relevance score. For an image, this means we can end up with a relevance value for each pixel.</p><p>One important idea here is that the relevance is not simply created again at every layer. It is redistributed.</p><p>LRP designed around a layer-wise conservation principle.</p><p>In our building example, we are not creating more heat while tracing it back. We are asking how the final amount can be divided among the radiators that contributed to it.</p><p>That is where the LRP redistribution rule comes in. We know that relevance moves backward through the network. The next step is to decide <strong>how much of it each neuron receives</strong>. The main idea is to redistribute the relevance of one neuron to the neurons that contributed to it.</p><p>Suppose neuron <em>k</em> has a relevance value of <strong>Rₖ</strong>. This relevance is shared between the neurons <em>j</em> that contributed to it.</p><p>A basic LRP rule can be written as:</p><p><strong>Rⱼ = Σₖ [(aⱼ wⱼₖ) / (Σⱼ′ aⱼ′ wⱼ′ₖ)] Rₖ</strong></p><p>Here:</p><ul><li><strong>Rₖ</strong> = relevance of neuron <em>k</em></li><li><strong>Rⱼ</strong> = relevance received by neuron <em>j</em></li><li><strong>aⱼ</strong> = activation of neuron <em>j</em></li><li><strong>wⱼₖ</strong> = weight between neurons <em>j</em> and <em>k</em></li></ul><p>The fraction shows how much neuron <em>j</em> contributed to neuron <em>k</em> compared with the other neurons connected to it. A larger contribution means more relevance. This calculation is repeated layer by layer until the relevance reaches the input.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AI_sO8vvaw34Rcge.png" /><figcaption>Figure: Relevance is propagated backward from the output through the network to the input. Source: Yoann Poupart, “Layer-Wise Relevance Propagation.”</figcaption></figure><p>For an image classifier, this relevance can be mapped back to the pixels. If the model predicts <strong>dog</strong>, pixels around the face, ears, or body may receive higher relevance, while the background receives less. The result can be visualized as a relevance heatmap.</p><p>The redistribution rule shown here is a simplified version. LRP also has different propagation rules, such as the <strong>ε-rule</strong> and <strong>α-β rule</strong>. The chosen rule can change how relevance is distributed and, as a result, how the final explanation looks.</p><p>LRP gives us another way to inspect a prediction, by tracing the model output backward and distributing its relevance across the input.</p><p>Together:</p><p><strong>LIME / SHAP → perturbation</strong><br><strong>Gradients → sensitivity</strong><br><strong>CAM / Grad-CAM → visualization</strong><br><strong>LRP → relevance propagation</strong></p><p>These methods ask different questions, but together they form a useful toolbox for understanding deep learning models.</p><p>But having an explanation does not automatically mean having a reliable explanation. Different methods can highlight different things, and an explanation can look convincing without fully reflecting how the model actually works. So the next question is: <strong>Can we trust the explanations we get?</strong></p><h3>References</h3><ul><li>Bach et al. (2015), <em>On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation</em></li><li>Montavon, Samek &amp; Müller (2018), <em>Methods for Interpreting and Understanding Deep Neural Networks</em></li><li>Samek, Wiegand &amp; Müller (2017), <em>Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models</em></li><li><a href="https://yp-edu.github.io/articles/layer-wise-relevance-propagation">Poupart, Y. (2024). <em>Layer-Wise Relevance Propagation</em>.</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5bb26a5405ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Explainable AI: How Models Think? — III (Visualization-Based Explanations)]]></title>
            <link>https://medium.com/@eniseirem/explainable-ai-how-models-think-iii-visualization-based-explanations-365d50b4963e?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/365d50b4963e</guid>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[visualization]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Sun, 09 Aug 2026 18:44:42 GMT</pubDate>
            <atom:updated>2026-08-09T18:44:42.808Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Part 3.3 — The Toolbox: Visualization Methods</em></p><p>After measuring sensitivity through gradients, we now move on to the next one, visualization-based explanations.</p><p>Gradients tell us how sensitive a prediction is to small changes in the input. However, deep vision models often contain thousands of gradient values, making it difficult to identify which parts of the input contributed to a prediction. Visualization-based explanations address this challenge by focusing on <strong>where the model concentrates its attention</strong> rather than how strongly the output changes.</p><figure><img alt="Source : GeeksForGeeks" src="https://cdn-images-1.medium.com/max/640/0*gibwStyogExuTGID.jpg" /><figcaption>Saliency Maps Source : <a href="https://www.geeksforgeeks.org/machine-learning/what-is-saliency-map/">GeeksForGeeks</a></figcaption></figure><p>In image models, information passes through multiple layers, with each layer learning different visual patterns from the input. Early layers typically detect simple features such as edges, colors, and textures, whereas deeper layers capture higher-level concepts such as shapes, object parts, and complete objects. Visualization-based explanation methods aim to reveal these internal representations by showing which regions of an image are most relevant to a model’s prediction.</p><p>So… how can we use this information to determine which regions of an image contributed most to a model’s prediction?</p><p>One of the earliest methods developed for this purpose was Class Activation Mapping (CAM). Introduced by Zhou et al. in 2016, CAM highlights image regions associated with a target class. Instead of examining individual neurons, the method generates a heatmap that shows where the network focused when producing its prediction.</p><p>CAM combines the feature maps from the final convolutional layer using class-specific weights. If (A_k(x,y)) is the activation of feature map (k) at location ((x,y)), and (w_k^c) is the weight assigned to that feature map for class (c ), the class activation map is calculated as:</p><figure><img alt="CAM calculation" src="https://cdn-images-1.medium.com/max/538/1*I04AcFW5-0t1sETq1CDZTQ.png" /></figure><p>The equation shows that the final heatmap is obtained by combining all feature maps according to their importance for the predicted class.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ahOEtnlBvTculXY_" /><figcaption>Source: <a href="https://debuggercafe.com/basic-introduction-to-class-activation-maps-in-deep-learning-using-pytorch/">https://debuggercafe.com/basic-introduction-to-class-activation-maps-in-deep-learning-using-pytorch/</a></figcaption></figure><p>So basically when the model makes a prediction, the classification layer assigns different importance weights to these feature maps. CAM combines the feature maps with these weights and creates a single heatmap. Regions with higher values contribute more to the predicted class, while regions with lower values contribute less. As a result, the heatmap shows which parts of the image were most important for the model’s prediction.</p><p>For example, if a model predicts that an image contains a dog, the heatmap should ideally highlight the dog’s head or body rather than the surrounding grass or background. This provides a visual explanation of the evidence used to make the prediction.</p><p>The final convolutional layer is used because it preserves the spatial structure of the image while also capturing high-level semantic information.</p><p>Although CAM introduced an effective way to visualize model attention, it has one important limitation. The method only works with CNN architectures that include a Global Average Pooling (GAP) layer before the final classification layer. Since many modern architectures do not follow this design, CAM cannot be applied directly.</p><p>To overcome this limitation, Selvaraju et al. introduced Gradient-weighted Class Activation Mapping (Grad-CAM) in 2017. Instead of using class-specific weights learned during training, Grad-CAM computes importance weights from the gradients of the target class with respect to the feature maps. This allows the method to be applied to a much wider range of convolutional neural networks without modifying their architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/0*41wlUWpbCT7TDGFb" /><figcaption>Source: <a href="https://www.researchgate.net/figure/Example-attention-maps-of-CAM-and-Grad-CAM-cite-from-reference-22_fig6_337804593">https://www.researchgate.net/figure/Example-attention-maps-of-CAM-and-Grad-CAM-cite-from-reference-22_fig6_337804593</a></figcaption></figure><p>As shown above, CAM and Grad-CAM often highlight similar regions, but Grad-CAM generally produces more complete and accurate localization maps. In both examples, Grad-CAM covers a larger portion of the target object, making it easier to identify which regions of the image contributed most to the model’s prediction.</p><p>Visualization-based methods provide a direct way to examine where a model focuses when making a prediction by transforming internal network activations into heatmaps. Techniques such as CAM and Grad-CAM reveal the image regions that contribute most to a target class, making model decisions easier to inspect.</p><p>However, these methods primarily indicate <strong>where</strong> the model is looking rather than <strong>how</strong> each feature contributes to the final prediction. To address this limitation, the next section explores relevance propagation methods, which trace a prediction backward through the network and assign a relevance score to every input feature.</p><h3>References</h3><ul><li><a href="https://doi.org/10.1109/CVPR.2016.319">Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., &amp; Torralba, A. (2016). <em>Learning Deep Features for Discriminative Localization</em>. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).</a></li><li><a href="https://doi.org/10.1109/ICCV.2017.74">Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., &amp; Batra, D. (2017). <em>Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization</em>. Proceedings of the IEEE International Conference on Computer Vision (ICCV).</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=365d50b4963e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Explainable AI: How Models Think? — III (Sensitivity-Based Explanations)]]></title>
            <link>https://medium.com/@eniseirem/explainable-ai-how-models-think-iii-sensitivity-based-explanations-2c6e4cb85ab4?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c6e4cb85ab4</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Wed, 20 May 2026 12:49:20 GMT</pubDate>
            <atom:updated>2026-05-20T12:49:20.656Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Part 3.2 — The Toolbox: Sensitivity-Based Explanations</em></p><p>In the previous section, we explored perturbation-based methods like LIME and SHAP, which treat the model as a black box. Sensitivity-based approaches on the other hand, open up the model and analyze how small changes in the input propagate through it.</p><blockquote>How much does the model’s prediction change when the input changes slightly?</blockquote><p>Let’s take a simple function: <em>f(x) = x²</em> as our model. If we change our input <em>x</em> to <em>x + ε,</em> new output becomes:<em> f(x + ε) = (x + ε)² = x² + 2xε + ε²</em></p><p>The change in output is the difference between the two: <em>f(x + ε) − f(x) = x² + 2xε + ε² — x² = 2xε + ε²</em></p><p>If <em>ε</em> is very small, <em>ε²</em> is negligible, so we get approximately: change ≈ <em>2xε . </em>Now divide by <em>ε: (f(x + ε) − f(x)) / ε ≈ 2x</em></p><p><strong>This value (2x) is the gradient.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hhvEIFZp3FIWIF8r.png" /><figcaption><em>The gradient is computed at a single point (P) and assumes a linear approximation. However, as we move to another point (Q), the true function deviates from this approximation, showing that gradients are only reliable locally. Source: </em><a href="https://learninglab.rmit.edu.au/maths-statistics/differentiation/d2-gradients-tangents-and-derivatives/"><em>RMIT University</em></a></figcaption></figure><p>In real models, inputs are multi-dimensional, meaning the gradient is computed for each feature separately, with this we get to estimate the importance of every input dimension. If a small change in a feature causes a large change in the output, that feature is important. If the output barely changes, the feature has little influence.</p><p>Since gradients directly measure this effect, they provide a natural way to estimate feature importance. This makes them model-aware (they use internal structure), fine-grained (per feature / per pixel), efficient (computed via backpropagation).</p><p>This allows us to treat the gradient as an attribution signal, assigning an importance score to each input feature. However, raw gradients by themselves are often difficult to interpret. This is where sensitivity-based explanation methods used.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*CqJeEJ8HEdhTomra.jpg" /><figcaption>Saliency Maps. Source: <a href="https://www.geeksforgeeks.org/machine-learning/what-is-saliency-map/">GeeksForGeeks</a></figcaption></figure><p>The simplest approach is to directly use gradients as explanations. In image models, this leads to <strong>saliency maps</strong>, where the gradient is computed for each pixel and visualized as a heatmap. These maps aim to show which regions of the input most influence the prediction. They are often noisy and unstable because they rely on a single local estimate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1002/1*dT_rwivGmUFTPxad3tiG_w.png" /><figcaption>Integrated Gradients Source: <a href="https://arxiv.org/pdf/1703.01365">Axiomatic Attribution for Deep Networks</a></figcaption></figure><p>To address this limitation, more robust methods have been proposed. One of the most widely used is <strong>Integrated Gradients</strong>, which improves standard gradients by accumulating them along a path from a baseline input to the actual input. Instead of relying on a single point, it considers how the model’s output changes over a range of inputs, resulting in smoother and more reliable attributions.</p><p>Beyond these core methods, several extensions have been proposed to improve gradient-based explanations. Techniques such as <strong>SmoothGrad</strong>, <strong>Guided Backpropagation</strong>, and <strong>Gradient × Input</strong> aim to reduce noise or refine the raw gradient signal, often by smoothing, filtering, or re-weighting it. Other approaches like <strong>DeepLIFT</strong> introduce alternative ways of tracking contributions through the network.<strong> Saliency maps</strong> focus on direct visualization, while <strong>Integrated Gradients </strong>focus on reducing noise and improving theoretical consistency. While these methods differ in implementation, they all build on the same fundamental idea of sensitivity.</p><p>Sensitivity-based methods provide a more direct view into model behavior by leveraging gradients. However, their reliance on local derivatives makes them <strong>sensitive to noise and model irregularities.</strong> This motivates the need for more structured visualization techniques, which we explore in the next section.</p><p><strong>References<br></strong><a href="https://arxiv.org/abs/1312.6034">Karen Simonyan, Andrea Vedaldi, Andrew Zisserman — <em>Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps</em> (2013)</a><br><a href="https://arxiv.org/abs/1703.01365">Mukund Sundararajan, Ankur Taly, Qiqi Yan — <em>Axiomatic Attribution for Deep Networks</em> (2017)</a><br><a href="https://arxiv.org/abs/1706.03825">Daniel Smilkov, Nikhil Thorat, Been Kim — <em>SmoothGrad: Removing Noise by Adding Noise</em> (2017)<br></a><a href="https://arxiv.org/abs/1412.6806">Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox — <em>Striving for Simplicity: The All Convolutional Net</em> (2015)</a><br><a href="https://arxiv.org/abs/1704.02685">Avanti Shrikumar, Peyton Greenside, Anshul Kundaje — <em>Learning Important Features Through Propagating Activation Differences</em> (2017)</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c6e4cb85ab4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Explainable AI: How Models Think? — III (Local Approximation & Perturbation)]]></title>
            <link>https://medium.com/@eniseirem/explainable-ai-how-models-think-iii-local-approximation-perturbation-0e0e00c52be8?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e0e00c52be8</guid>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Fri, 01 May 2026 16:05:36 GMT</pubDate>
            <atom:updated>2026-05-01T16:05:36.804Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Part 3.1 — The Toolbox: LIME &amp; SHAP</em></p><p><a href="https://medium.com/@eniseirem/explainable-ai-how-models-think-iii-04b4a136eeb7">In Part 3</a>, we framed explanation as interaction. We changed the inputs, observed outputs, and tried to understand patterns through these reactions. Among all the approaches in the toolbox, some of them follow this idea very closely.</p><p>In local explanations we won’t be trying to understand the model as a whole. We are focusing on <strong>a single prediction</strong>, and exploring the model’s behavior in its immediate neighborhood. This is where perturbation-based methods come in.</p><h3>LIME</h3><p>LIME explains a prediction by looking at small changes around one example. It creates slightly different versions of the input and checks how the model’s prediction changes. You can think of it like asking the same question in slightly different ways and observing how the answer changes. If a small change leads to a very different answer, that part is probably important.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Tv85eaY5fpw_NPOQ.jpg" /><figcaption><a href="https://www.oreilly.com/content/introduction-to-local-interpretable-model-agnostic-explanations-lime/">Source: O’Reilly</a></figcaption></figure><p>From a technical perspective, LIME works in three main steps. First, it samples new data points by perturbing the original input, creating small variations around the example. For images, you can think of this as dividing the picture into puzzle like regions and selectively hiding some pieces while keeping others visible.</p><p>These modified versions are then passed through the model to obtain predictions. By observing how the output changes when certain pieces are missing, we can infer which parts of the image are important. Next, each sample is weighted based on its distance to the original instance, so closer points have more influence. Finally, a simple interpretable model, often a sparse linear model, is trained on this weighted dataset to approximate the model locally.</p><p>Because of this setup, the explanation depends on how the local neighborhood is defined and how the perturbations are generated. Different choices can lead to different explanations, which means LIME does not provide a single “true” interpretation. Instead, it offers a local, data dependent approximation of the model’s behavior rather than a global understanding.</p><h3>SHAP</h3><p>SHAP builds on a similar idea but takes a more structured approach. Now, rather than fitting a local model, it focuses on assigning a contribution value to each feature. It is actually a game-theory based approach. It treats each feature as a player in a cooperative game, where the prediction is the final outcome. The goal is to fairly distribute this outcome among all features.</p><p>To do this, SHAP considers many different combinations of features being present or absent and measures how the prediction changes in each case. By averaging these effects, it estimates how much each feature contributes to the final prediction. You can think of it as systematically testing different versions of the input and tracking how the prediction shifts when features are added or removed.</p><p>This makes SHAP more consistent and theoretically grounded, but also more computationally expensive.</p><p>Both methods rely on perturbations, but they reflect different ways of thinking about explanations. LIME focuses on approximating the model’s behavior locally, while SHAP aims to assign contributions to each feature in a more systematic and theoretically grounded way. In practice, this means LIME gives a quick and intuitive local explanation, while SHAP provides more consistent and comparable feature attributions, often at a higher computational cost.</p><p>So far, we have been actively changing the input to understand how the model behaves. But as we learned before, not all explanation methods rely on explicit perturbations. Some approaches take a different path and look at how the output would change under very small variations, without generating new samples.</p><p>This leads us to the next group of methods: <strong>sensitivity-based explanations.</strong></p><p><a href="https://arxiv.org/abs/1705.07874"><em>A Unified Approach to Interpreting Model Predictions</em> (SHAP)</a></p><p><a href="https://arxiv.org/abs/1602.04938"><em>“Why Should I Trust You?”: Explaining the Predictions of Any Classifier</em> (LIME)</a><br><a href="https://arxiv.org/abs/1705.07874">Lundberg, S. M., &amp; Lee, S.-I. (2017).</a></p><p><a href="https://www.oreilly.com/content/introduction-to-local-interpretable-model-agnostic-explanations-lime/"><em>Local Interpretable Model-Agnostic Explanations (LIME): An Introduction</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e0e00c52be8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Explainable AI: How Models Think? — III]]></title>
            <link>https://medium.com/@eniseirem/explainable-ai-how-models-think-iii-04b4a136eeb7?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/04b4a136eeb7</guid>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Tue, 07 Apr 2026 13:23:10 GMT</pubDate>
            <atom:updated>2026-04-07T13:23:10.443Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Part 3 — The Toolbox: How We Actually Probe Models</blockquote><p><a href="https://medium.com/@eniseirem/explainable-ai-how-models-think-ii-a7ee89409fca">In Part 2, </a>we moved from reading models to interacting with them.</p><p>Deep learning models do not give us clear rules or step-by-step reasoning. We explore how they behave, we change inputs, observe reactions, and try to understand patterns through these interactions. Over time, this process become more structured. Different methods have been developed to help us look at model behavior in consistent ways. Now, we will look into those methods in our XAI toolbox.</p><h3>What Are We Actually Trying to Explain?</h3><p>Are we trying to understand one decision, or the model as a whole? This question shapes everything that follows.</p><p>Local explanations focus on a single prediction. Global explanations step back and look at the broader behavior of the model. Do we want to understand why this image was classified as a cat or are we asking which features are generally important or what patterns the model has learned across the dataset?</p><p>Both perspectives are necessary. A model can give reasonable explanations for individual decisions while still relying on problematic patterns overall. Or it can capture the right general trends but fail in specific cases. Understanding this distinction helps us decide not only how to explain a model, but what exactly we are trying to understand. With this in mind, the different methods start to make more sense.</p><p>Explanation methods are not just technical tools. They represent different ways of looking at a model, from testing how it reacts to input changes, to measuring sensitivity, tracing decision paths, or simplifying its behavior for interpretation.</p><h4><em>What if we simplify the model locally?</em></h4><p>Instead of trying to interpret a complex neural network directly, we can approximate its behavior locally with a simple linear model. Around a specific data point, this simpler model can tell us which features contributed positively or negatively to the prediction. In this sense, we are not explaining the full model, but building a local approximation that behaves similarly in that region.</p><p>Methods like <strong>LIME</strong> do this by generating small variations of the input around a specific instance, collecting the model’s predictions for these samples, and fitting a simple interpretable model to approximate the decision boundary locally.</p><p>For example, in a loan approval model, a local approximation might show that a specific decision was mainly influenced by income and credit history, even if the original model is much more complex.</p><h4>What breaks when we change the input?</h4><p>Another approach is to change the input and observe what happens. If hiding a region in a picture causes the confidence to drop, that region was important for the decision. This is a very direct way of testing what the model relies on.</p><p>This idea appears in <strong>occlusion methods</strong>, where parts of the input are systematically removed, and also in <strong>SHAP</strong>, which estimates feature importance by comparing predictions across many combinations of present and missing features, approximating Shapley values from game theory.</p><p>For example, if covering the background of an image changes the prediction more than covering the object itself, it may indicate that the model is relying on context rather than the object.</p><h4>What is the model most sensitive to?</h4><p>We can also look at sensitivity without actually modifying the input. By measuring how much the output would change with a small variation, we can estimate which parts of the input the model is most responsive to.</p><p>Methods like <strong>saliency maps</strong> compute the gradient of the output with respect to the input, while <strong>Integrated Gradients</strong> accumulate gradients along a path from a baseline input to the actual input, making the attribution more stable and less noisy.</p><p>For example, in a text model, slightly changing a single word might cause a large shift in the prediction, showing that the model is highly sensitive to that specific token.</p><h4>How does the decision flow through the network?</h4><p>Finally, some methods start from the prediction and work backwards through the network. They redistribute the output score step by step to the input features, assigning each part a share of responsibility. This gives a structured view of how the decision was built inside the model.</p><p>Methods like <strong>Layer-wise Relevance Propagation (LRP)</strong> follow this idea by conserving and redistributing the prediction score across neurons, tracing how much each component contributed to the final output.</p><p>For example, in an image classification task, this produces a heatmap showing which regions contributed most to the prediction, reflecting how the decision was constructed internally.</p><p>Each of these methods gives us a different perspective on the same model. In the next parts, I plan to go into these methods in more detail, starting with local approximation methods like LIME and SHAP, then moving into gradient-based and relevance-based approaches.</p><p>References<br><a href="https://arxiv.org/abs/2003.07631">Samek, W., Montavon, G., Vedaldi, A., Hansen, L. K., &amp; Müller, K.-R. (2021).<br> <em>Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications</em></a><em><br></em><a href="https://arxiv.org/abs/1602.04938">Ribeiro, M. T., Singh, S., &amp; Guestrin, C. (2016).<br> <em>“Why Should I Trust You?”: Explaining the Predictions of Any Classifier</em> (LIME)</a><br><a href="https://arxiv.org/abs/1705.07874">Lundberg, S. M., &amp; Lee, S.-I. (2017).<br> <em>A Unified Approach to Interpreting Model Predictions</em> (SHAP)</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04b4a136eeb7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Explainable AI: How Models Think? — II]]></title>
            <link>https://medium.com/@eniseirem/explainable-ai-how-models-think-ii-a7ee89409fca?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a7ee89409fca</guid>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Wed, 18 Feb 2026 18:09:59 GMT</pubDate>
            <atom:updated>2026-02-18T18:09:59.344Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Part 2 — The Black Box: How Deep Learning Changes What “Reasoning” Means</blockquote><p><a href="https://medium.com/@eniseirem/explainable-ai-how-models-think-0f055bcf4415">In Part 1</a>, we talked about what it means to explain a model and why explainability starts with how a model learns. Now, let’s deep dive into further.</p><p>What happens when models stop reasoning in terms of explicit features or rules?</p><h3>Deep Learning Learns Representations</h3><p>Classical models directly work on input features. So, we as humans, know what model does almost instinctively. Deep learning models do something different, they build internal representations across layers.</p><p>You can think of a deep neural network as a sequence of transformations, where <strong>each layer creates a new representation of the input.</strong> The model is not storing the rules but learning <strong>how to reshape information</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/0*M3EiQHcJGEQwKBs3.jpg" /><figcaption><a href="https://logicmojo.com/artifical-neural-network">https://logicmojo.com/artifical-neural-network</a></figcaption></figure><p>Early layers often capture simple patterns, but next layers combine them into abstract concepts. The meaning of a prediction is therefore not located in one feature, or even one neuron. It is distributed across the network, that would mean that also the meaning is distributed.</p><h3>Why This Makes Explanation Harder</h3><p>In classical models, we generally look into which input feature influenced the prediction. In deep learning, we want to know two things, w<strong>hich input mattered</strong> and <strong>how was that information transformed i</strong>nside the network?</p><p>Modern machine learning systems can be seen in real life areas with high stakes; healthcare, finance, or any system where automated decisions affect real outcomes. Explainable AI exists because high performance alone is not enough. We need systems that are understandable, inspectable, and debuggable.</p><p>And importantly, explainability is not only about explaining decisions after deployment. It is also about improving models during development.</p><p>In one well-known example, a neural network trained for horse detection achieved high accuracy. However, explanation methods revealed that the model was heavily relying on a recurring copyright watermark present in many of the horse images. Instead of learning the visual features of the animal, the model had learned a dataset artifact. This is known as shortcut learning, where performance remains high but the underlying reasoning is flawed.</p><p>The phenomenon is often referred to as the Clever Hans effect, named after the horse that appeared to perform arithmetic but was actually responding to subtle cues rather than understanding mathematics. Much like Clever Hans, the model was not recognizing horses; it was recognizing hints.</p><p>If a deep model learns shortcuts, bias, or unstable decision boundaries, performance metrics alone may not reveal it.</p><h3>What Does It Mean to “Explain” a Deep Model?</h3><p>An explanation is not the model itself rather an approximation that highlights certain aspects of model behavior. Explanations are context-dependent and question-dependent.</p><p>Explanations can look convincing and still be wrong.</p><ul><li>They can hide bias.</li><li>They can oversimplify complex reasoning.</li><li>They can create false confidence if not evaluated carefully.</li></ul><p>This is why XAI is not just about generating explanations more about testing whether explanations are faithful and stable reflections of model behavior.</p><p>Once we accept that deep models cannot be explained through simple rule extraction, we need to think about if we cannot directly read the model <br>how do we probe it? How do we poke around with a stick enough to understand its behaviour and its reactions?</p><p>We test its reactions, change the input slightly and observe what shifts. We remove parts and see what collapses, we amplify signals and measure what responds.</p><p>There are all ways of explainable AI works.</p><p>Some methods measure sensitivity; how much the output changes when the input changes. Some trace relevance flow; they try to see which parts of the network were most important for the decision. Some change or remove parts of the input to see what happens to the prediction. Others try to connect internal representations to concepts humans understand.</p><p>And this is where the technicalities of XAI begins. To move from intuition to practice, we need tools that let us test, measure, and analyze model behavior in a systematic way.</p><p>Next, we will look at the main families of Explainable AI algorithms, <br>gradient-based methods, perturbation approaches, attribution techniques, and representation-level explanations, and what each of them is actually trying to measure.</p><p><em>Part 3 — The Toolbox: How We Actually Probe Models</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a7ee89409fca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Explainable AI : How Models Think?]]></title>
            <link>https://medium.com/@eniseirem/explainable-ai-how-models-think-0f055bcf4415?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/0f055bcf4415</guid>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Fri, 06 Feb 2026 21:49:42 GMT</pubDate>
            <atom:updated>2026-02-06T21:53:44.629Z</atom:updated>
            <content:encoded><![CDATA[<h3>Explainable AI : How Models Think?</h3><blockquote>Part 1 — Foundations: What Does It Mean to Explain a Model?</blockquote><p><em>This series will be about understanding the explainable AI.</em></p><p>Explainable AI (XAI) is asking a simple question: why did the model work the way it worked? Before we start talking about techniques, we should first discuss that what does it even mean to explain a model.</p><p>Most discussions about AI start with performance. Accuracy, benchmarks, state-of-the-art results. However every real-world decision about using AI depends on trust. And trust does not come from performance alone. It comes from understanding.</p><p>Explainability matters because AI systems do not operate in isolation. Their outputs influence decisions about people, resources, risks, and opportunities. When a model fails, or when its behavior changes in unexpected ways, performance metrics alone cannot tell us why. Without understanding how a model reasons, we cannot diagnose errors, detect bias, or know when a system should not be trusted. Explainability gives us a way to question model behavior. It turns predictions into objects we can inspect, challenge, and improve. Which is essential if AI systems are meant to be used responsibly rather than blindly</p><h3>Explainability Starts With Learning, Not With Explanations</h3><p>A common mistake is to treat explainability as something that happens <em>after</em> the model is built. Train first, explain later. But explanations always inherit the logic of the learning process.</p><ul><li>If the training data is biased, the explanations will be biased.</li><li>If shortcuts were learned, the explanations will highlight shortcuts.</li><li>If information leaked, the explanations will just justify leakage.</li></ul><p>Explainability cannot fix a broken learning process. It can only make its structure more visible. In that sense, XAI is not just about interpreting models. It is about supervising how knowledge was formed in the first place.</p><h3>Why Classical Models Still Matter</h3><p>This is why classical machine learning models remain important in any discussion about explainability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/739/0*gifcNdDLVQ9xr5Jt.png" /><figcaption><a href="https://rasbt.github.io/mlxtend/user_guide/classifier/LogisticRegression/">https://rasbt.github.io/mlxtend/user_guide/classifier/LogisticRegression/</a></figcaption></figure><p><strong>Logistic regression </strong>gives us clear feature weights. We can see which inputs push the prediction up or down, and by how much. This creates a baseline form of understanding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/573/0*U-BoVcbomHIr-vZA.png" /></figure><p><strong>Decision trees </strong>go one step further. They teach logic. Each split represents a visible decision. You can follow the reasoning path from input to output without guessing.</p><p>These models show us what explanation looks like when reasoning is explicit.</p><p><strong>Ensemble models </strong>combines many reasonable models, results with gaining performance. But doing so, we also lose clarity. Individual reasoning paths disappear. Feature importance becomes global and averaged, not specific to a single decision. <strong>Higher accuracy often comes at the cost of localized understanding.</strong></p><p>Some models reason through explicit features. Some reason through rules and conditions. Others combine many weak signals into a collective outcome. Explainability depends on this structure. There is no universal explanation that fits all models.</p><p>This is the foundation XAI builds on.</p><p>And this is also why larger models do not remove the need for explainability. They make it more important.</p><p><em>Part 2: “The Black Box Problem: How Deep Learning Changes What ‘Reasoning’ Means”</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0f055bcf4415" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Trustworthy AI Isn’t Optional: The Case of ChatGPT and Adam Raine]]></title>
            <link>https://medium.com/@eniseirem/trustworthy-ai-isnt-optional-the-case-of-chatgpt-and-adam-raine-33ee8c870169?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/33ee8c870169</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[accountability]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Sun, 16 Nov 2025 14:03:56 GMT</pubDate>
            <atom:updated>2025-11-16T14:03:56.151Z</atom:updated>
            <content:encoded><![CDATA[<h4>A reflection on how accountability defines trust in AI.</h4><p>When a 16-year-old student followed instructions from ChatGPT that led to his death, a global discussion on AI accountability began. The case of Adam Raine is more than a tragic event, it is evidence of what happens when artificial intelligence operates without full responsibility.</p><h4>What Happened</h4><p>Adam Raine was a teenager living in California. He often used ChatGPT like many of us. His parents reported that he had mental health struggles. Instead of safe guidance, he received harmful instructions from the chatbot. Court documents show that ChatGPT allegedly gave step-by-step details on how to take his life and encouraged him by praising him for tying a rope correctly.</p><p>In a chat log from April 2025, Adam allegedly uploaded a photo of a noose and asked whether it could “hang a human”. The chatbot analyzed the image, commenting on its structure and suggesting ways to make it ‘’safer’’ or more stable, treating a life-threatening object as a technical problem to solve. When Adam later wrote, “I want to leave my noose in my room so someone finds it and tries to stop me,” ChatGPT responded: “Please don’t leave the noose out … Let’s make this space the first place where someone actually sees you.”</p><p>In August 2025, Adam died, and his parents later found the disturbing chat logs. The family, supported by the <strong>T</strong>ech Justice Law Project, filed a wrongful death lawsuit in San Francisco Superior Court. This case has been widely reported by <em>The Guardian</em> and <em>The Washington Post</em>, becoming a symbol of the missing accountability in AI systems.</p><h4>Where Accountability Failed</h4><p>The European Commission defines accountability as the presence of mechanisms that ensure responsibility and liability for AI outcomes, in other words defining who is responsible when harm occurs.</p><p>Without it, people are left without protection, and public trust in AI disappears. In practice, this means three things: <strong>auditability</strong>, <strong>traceability</strong>, and <strong>clear allocation of responsibility</strong>.</p><p>In the Raine case, all three failed:</p><ul><li><strong>No effective preventive safeguard.</strong> ChatGPT was released with content filters, but these were not enough to protect vulnerable users like Adam.</li><li><strong>No clear liability mechanism.</strong> When harm happened, the family had to go to court.</li><li><strong>No transparent oversight.</strong> Users did not know how ChatGPT was trained, how safety rules were enforced, or how harmful interactions were monitored.</li></ul><p>The EU guidelines also note that accountability requires the ability to audit and contest AI decisions. However, the parents only saw the harmful chats after their son’s death. This lack of transparency gave control to the company, while affected ones had no power.</p><h4>A Legal and Ethical Gap</h4><p>In the EU, upcoming laws such as the <strong>AI Act</strong> aim to make accountability enforceable through regulation.<br>But in the US, there is no dedicated legal framework for AI liability yet.<br>Families must rely on general product liability and negligence doctrines, which were designed for physical products, not digital services. This gap leaves courts unsure whether an AI company can be treated like a manufacturer.</p><p>OpenAI may argue that ChatGPT is just a tool. But when a system is unsafe by design, accountability cannot stop at the user’s screen. This legal uncertainty demonstrates why accountability cannot be left only to private companies. Voluntary ethics codes are not strong enough to prevent harm.</p><p>The case also raises questions about global consistency. If accountability rules differ between the EU and the US, companies may adapt only to the weakest system, leading to <em>regulation shopping</em>.</p><p>A harmonized, universal approach to accountability would help prevent this. The Raine case contributes to a global debate on <em>trustworthy AI</em>. Other values also play a role: <strong>Privacy</strong> (Adam was a minor and deserved stronger protection), <strong>Human Agency and Oversight </strong>(systems should support autonomy, not influence users negatively), and S<strong>ocietal Well-Being</strong> (public trust is at stake).</p><p>Still, accountability remains at the core because it enables responsibility for all others. Building accountable AI requires proactive design, not reactive defense.<br>We might take several practical steps:</p><ul><li><strong>Context-aware safeguards:</strong> Go beyond keyword filters to detect risky conversations.</li><li><strong>Participatory design:</strong> Include psychologists, educators, and young users in testing.</li><li><strong>Transparency and reporting:</strong> Allow users to report harmful content and publish safety audits.</li><li><strong>Ethical risk assessment:</strong> Simulate worst-case scenarios before deployment.</li></ul><p>These lessons extend beyond a single case and define what trustworthy AI should look like in practice.</p><h4>Lessons for Stakeholders</h4><p><strong>For data scientists:</strong> accuracy and efficiency are not enough. Ethical risks must be tested from the start.</p><p><strong>For AI companies:</strong> releasing strong models without safety systems is irresponsible. Accountability must be built into every layer of the system.</p><p><strong>For policymakers:</strong> the case highlights the need for strong legal frameworks. Current laws leave gaps that allow companies to avoid responsibility.</p><p><strong>For society:</strong> AI can help, but it cannot replace human care. Public discussion and awareness are essential to pressure companies and governments toward responsible action.</p><h4>Conclusion</h4><p>In short, accountability is not optional but necessary. Learning from this case can help avoid similar tragedies and guide AI toward human well-being.</p><p>The <strong>OpenAI and Adam Raine</strong> case is a reminder that accountability is a vital part of trustworthy AI.</p><p>Without responsibility and liability, AI can cause harm while leaving victims without support. Accountability must exist in every stage of AI, from design and testing to deployment and monitoring.</p><p>Other values such as privacy, human agency, and societal well-being also matter, but accountability makes them enforceable.</p><p>By learning from this case, data scientists, companies, policymakers, and society can move toward AI that is innovative, safe, responsible, and aligned with human values.</p><p>On 2 September 2025, OpenAI promised parental control tools to flag “acute stress” in young users, a post-tragedy fix that arrived a year too late for Adam Raine.</p><p>Accountability, it seems, still waits for the body count.</p><p><a href="https://www.courthousenews.com/wp-content/uploads/2025/08/raine-vs-openai-et-al-complaint.pdf">Raine v. OpenAI et al. (2025). Complaint filed in the Superior Court of California, County of San Francisco.</a></p><p><a href="https://www.theguardian.com/technology/2025/aug/27/chatgpt-scrutiny-family-teen-killed-himself-sue-open-ai">The Guardian. (2025, August 27). ChatGPT scrutiny grows as family of teen who killed himself sues OpenAI</a>.</p><p><a href="https://www.washingtonpost.com/technology/2025/09/02/chatgpt-parental-controls-suicide-openai/">The Washington Post. (2025, September 2). ChatGPT to get parental controls after teen user’s death by suicide</a>.</p><p><a href="https://ec.europa.eu/futurium/en/ai-alliance-consultation.1.html">European Commission. (2019). Ethics Guidelines for Trustworthy AI. High-Level Expert Group on Artificial Intelligence.</a></p><p><a href="https://openai.com/index/building-more-helpful-chatgpt-experiences-for-everyone/">OpenAI, (2025, September 2). Building more helpful ChatGPT experiences for everyone</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=33ee8c870169" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gemma 3N ne yapıyor?]]></title>
            <link>https://medium.com/@eniseirem/gemma-3n-ne-yap%C4%B1yor-6334d0830f71?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/6334d0830f71</guid>
            <category><![CDATA[gemma3n]]></category>
            <category><![CDATA[makine-öğrenimi]]></category>
            <category><![CDATA[google]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Fri, 23 May 2025 19:20:38 GMT</pubDate>
            <atom:updated>2025-07-03T11:45:33.490Z</atom:updated>
            <content:encoded><![CDATA[<p>Google, geçtiğimiz hafta tanıttığı <strong>Gemma 3N</strong> modeliyle serisini büyüttü.<br> Peki bu neden önemli?</p><p>Çünkü bu, <em>mobile-first</em> yani doğrudan mobil cihazlar için optimize edilmiş bir yapay zekâ modeli.<br> Başka bir deyişle: Artık ulaşılabilirlik bir üst seviyeye çıkıyor.</p><p>“Bu ne demek?” Yarın öbür gün telefonundaki asistan çok daha akıllı olacak hem de internet bağlantısına bile ihtiyaç duymadan.</p><p>RAM kullanımının adeta canavar besler gibi büyüdüğü bir çağda, <strong>Gemma 3N</strong>, bu yükü ciddi oranda azaltarak milyarlarca parametreye sahip modellerin mobil cihazlarda bile çalışabilmesini mümkün kılıyor.</p><p>Bunu nasıl başarıyor?</p><p><strong>Per-Layer Embedding</strong> tekniğiyle.</p><p>Modelin her bir katmanı için ayrı embedding parametreleri tanımlanıyor. Bu parametreler sadece ihtiyaç duyulan anlarda belleğe çağrılıyor.<br> Yani tüm ağırlıkları aynı anda yüklemek yerine “gerektiği kadarını” yükleyerek önemli bir bellek tasarrufu sağlıyor.</p><p>Bir diğer güçlü yanıysa: <strong>MatFormer mimarisi</strong>.<br> Bu yapı, farklı görevler için farklı alt yapılara geçiş yapabiliyor.<br> Yani, birden fazla küçük model gibi davranarak kullanım senaryosuna göre en uygun hesaplama yolunu seçiyor.<br> Bu sayede hem verimlilik artıyor hem de esneklik sağlanıyor.</p><p>Artık modelleri doğrudan telefonunuzda çalıştırabilir, verilerinizin bulutlara gitmesine gerek kalmadan gizliliği koruyabilirsiniz.<br>Üstelik internetin olmadığı yerlerde bile bu modellerden faydalanmak mümkün hale geliyor. Belki mobil sağlık uygulamalarına dünya üzerindeki erişimini arttırabilir veya felaket anlarında erişim sağlayabilirsiniz. İnternetin olmadığı yerlere seyahat ederken yanınızda dijital asistanınızı götürebilir çok gizli iş verilerini kullanırken milyonlarca lirayı sunucuya harcama gereği duymayabilirsiniz. Ayrıca bu süreç bellek oranı daha da düşük olan gözlük saat gibi aksesuarlarda bu kullanımların da kapısını açacağından daha kişisel daha mahrem daha ulaşılabilir ve daha hızlı hale geliyor olacak.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6334d0830f71" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[İnterneti Kullanmanın Daha İyi Yolu : Arc, The Browser Company]]></title>
            <link>https://medium.com/@eniseirem/i%CC%87nterneti-kullanman%C4%B1n-daha-i%CC%87yi-yolu-arc-the-browser-company-6b535d399ab0?source=rss-d9e4918777c1------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b535d399ab0</guid>
            <category><![CDATA[arc-browser]]></category>
            <category><![CDATA[tarayıcı]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[browsers]]></category>
            <category><![CDATA[the-browser-company]]></category>
            <dc:creator><![CDATA[Enise İrem Colak]]></dc:creator>
            <pubDate>Fri, 24 Mar 2023 00:06:41 GMT</pubDate>
            <atom:updated>2023-03-24T00:06:41.869Z</atom:updated>
            <content:encoded><![CDATA[<h3>İnterneti Kullanmanın Daha İyi Yolu : Arc, The Browser Company</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tZ_BBfwu4XDY0szOLJxMdw.png" /><figcaption>Arc’ın Bizi Karşıladığı İlk Görsellerden Biri</figcaption></figure><p><em>Bu yazı tamamen bir kullanıcı olarak Türkçe bir görüş olması amacıyla yazıldı. İlk kullanım ile elde edilen fikirleri barındırıyor. Temel amacım arkadaşlarıma önerdiğimde topluca bir kaynağa yönlendirebilmekti. Eksiklerim gediklerim affola :)</em></p><h3><strong>Arc Nedir?</strong></h3><p><a href="https://arc.net/">Arc’ın</a> mimarı <a href="https://thebrowser.company/">The Browser Company</a>’nin deyimi ile; yeni bir tarayıcıya ihtiyaç yokken oluşturulan interneti kullanmanın daha iyi bir yolu. Evet, temelde Chromium (Chrome ile ortak kökenli) bir tarayıcı ama “internette nefes almayı” sağlayacak bir araç olarak tanıtılıyor. Chrome’un <a href="https://gs.statcounter.com/">Statcounter üzerinde yer alan istatistiklere</a> göre yaklaşık yüzde 65i ile nerdeyse domine ettiği bir markete girmek için bence de, bir tarayıcıdan fazlasını sunmaları gerektiği ortada.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FqYXeMQsK-o5cqn6kvzl6w.png" /><figcaption>The Browser Company of New York sitesinden bir görüntü.</figcaption></figure><h3><strong>Diğer Tarayıcıların Suyu mu Çıktı?</strong></h3><p>Açıkçası kullandığı tarayıcıya çok da anlam yüklemeyen biri olarak, Firefox’tan Opera’ya, Tor Browser’a (ehe torrent), Safari’ye birçok tarayıcı kullandım ve hatta zaman zaman Edge’i de deneyimlediğim oldu. Eskiden dönemsel olarak tarayıcı çok değiştirirdim. Hatta hepsi aynı anda bilgisayarımda bulunduğu da oldu. Üniversite öncesi dönemlerde Opera’yı daha sık kullanırdım. Belli bir nedeni var mı derseniz, yok, marjinal olmaya çalışmak bile bir neden olabilir :) Yazılım Mühendisliği okumaya başladıktan sonra Chrome, Apple ekosistemine dahil oldukça da Safari kullanımım arttı. Chromedaki eklenti özgürlüğü daha çekici olsa da Apple ekosistemi dönüp dolaşıp Safari kullandırıyordu bana, belki basit arayüz tasarımındandır. O yüzden ikisi birden bilgisayarımda yer alıyordu hep. Hatta bunlara ek olarak, varolan her medya tüketme kanalına aboneliğimden yola çıkarak, sadece bu iş için kullandığım <a href="https://apps.apple.com/tr/app/friendly-streaming-browser/id553245401?l=tr&amp;mt=12">Friendly Streaming</a> diye bir tarayıcım da bilgisayarımda bulunuyor. Tüm bu tarayıcıların çok benzer özellikleri var. Bookmark/favorileme, eklenti ekleme, kişiselleştirme, gruplama gibi zamanla hepsi birbirinden az/çok esinlenerek benzer araçlar oldular. Arc ise hem bunları hem fazlasını, basitleştirilmiş bir arayüzle bize sunuyor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oC1IaoY9MJzN3BPSt0r3aQ.png" /><figcaption>Arc’ın websitesinden bir görsel</figcaption></figure><p>Bütün özelliklerini atlamadan bahsedebileceğimden emin değilim ama daha çok beni çeken özelliklerine değineceğim.</p><p>Öncelikle, çok da fazla açıklamaya ihtiyaç duymayacak, bildiğimiz şeylerden bahsedelim. Chromium bir proje olduğu için sevdiğimiz eklentiler de bizimle gelebiliyor. Ancak PiP gibi (Bir medyayı örn.Youtube’ı küçük pencerede oynatmak) alıştığımız bazı özellikler zaten gömülü halde; üstelik bizim tıklamamıza bile gerek kalmadan, farklı sekmeye geçtiğimizde medyamız bizimle geliyor. Favorileme, gruplama gibi diğer özelliklerimiz ise burda farklı bir adlandırmayla karşımıza çıkıyor diyebiliriz. Ama kullanım açısından da bir o kadar geliştirilmiş durumdalar. Birazdan bunlara da değiniyor olacağım. Arc’ın içinde bulunan iki farklı büyük özellik olarak gördüğüm gömülü özelliklerden biri not alma sistemini içinde bulundurması, diğeri de kişiselleştirme deneyimi.</p><h4>Gelelim detaylara…</h4><p>Tarayıcımızın nasıl göründüğünü görelim önce. Birkaç detay tabii sizin kişiselleştirmenize bağlı. Üstte görmeye alıştığımız sidebarımız kenarda yer alıyor. Bu da aslında tasarımda daha çok özgürlük sağlamış gibi gözüküyor. İstersek tabii sidebarı gizleyip tam ekran modunda devam edebiliyoruz.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yIcQFMawq6r5_UUvD0ebSg.png" /><figcaption>Arc’ın Görüntüsü</figcaption></figure><p>URL barının altında bende Twitter, Youtube olarak gördüğünüz alanı sürekli sabitleriniz olarak görebilirsiniz; favoriler, yer imleri bir bakıma. Bu kısma Google takvim eklediğinizde yaklaşan toplantıları, Spotify eklediğinizde müzik çalma animasyonu gösterme gibi ufak tasarım farklılıkları olabiliyor. Altında ise o an açık olan grubumun adı yer alıyor, “social”. Arc’taki bu gruplamanın adı Space. Hangi Space’in açık olduğundan bağımsız olarak bu sabitlerimiz her yere bizimle geliyor. Ancak bir diğer sabitleme yöntemi ise Spacelerin içinde yer alıyor. Bendeki örneği The Browser Company sekmesi ile görebilirsiniz.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/448/1*ARMZU5-tNTttTK6_CzDtVg.png" /><figcaption>Medyalar ve Space alanı</figcaption></figure><p>Spaceler kendi içlerinde ayrı bir pencere gibi de düşünülebilir, çünkü Spacelerin içinde de klasörler oluşturabiliyorsunuz. Klasör imgesiyle sekmelerin arasında yer alıyor öyle bir durumda. Spaceler en altta, farklı sekmeleri düzenli tutuyor. Benim sidebarımda da göreceğiniz üzere, gezegen-film-yılan-resim… şeklinde ikonlarla dizebiliyorsunuz. Her bir Space’in adı, ikonu, teması özelleştirilebiliyor. Bu düzenleme açıkçası benim çalışma, sosyal medya gibi ayırma planlarımı oluşturmamı sağladı. Buna benzer bir ayırma yöntemi ise profil oluşturma seçeneği, kişisel-iş gibi bir kullanım için daha kullanışlı olduğunu düşündüğüm bu özellik, Chrome’daki profillere benzetilebilir.</p><p>Spacelerin iki ucunda bulunan alanlardan biri yeni sekme ve birtakım diğer özellikleri açmaya yararken, diğeri medya erişimini sağlıyor. İndirilenlerin yanısıra, bir ekran görüntüsü aldığınızda da hızlı bir erişim sağlıyor bu minik kutucuk. İçerisinde ayrıca arşivlenen sekmeleriniz, Spaceler, Notların yer aldığı alanlar bulunuyor.</p><p>Kişiselleştirme deneyiminden bahsedeyim biraz da… Elbette ikonlaştırma, renkler, eklentiler falan hoş da, önemli görme nedenim bu değil elbette. Boost denen bir özelliği sayesinde, kullandığınız websitelerine basit kod eklemeleri yapmanıza olanak sağlıyor. Kullandığınız bir sitenin arkaplanını değiştirebilir, o siteye her girdiğinizde görebileceğiniz bir gifi başa ekleyebilir, hatta yapmayı unutmamak için kendi ToDo listenizi bile en başlarına ekleyebilirsiniz. Kodlamaya uzak olanlar için daha hazır kalıplar da sunması iyi düşünülmüş. Yapılan şey basit bir tarayıcı konsolu düzenlemesi olsa da, bunun entegre ve kalıcı bir deneyim sağlayan bir özellik olarak getirilmiş olması, bence şirketin kendi deneyiminizi yaratmanıza önem verdiğini gösteriyor.</p><p>Gelelim not alma sistemine. Biri Notes adında, bildiğimiz basit Apple Notes benzeri tasarımında, hatta belki de daha kısıtlı bir alan. Tarayıcıya hızlı not tutmak yine gayet güzel. Ancak Easel seçeneği adeta sınırsız bir alan. Ekleyeceğiniz resimlerden, çizim yapabilmeye birçok özelliği bulunuyor. Dijital scrapbook olarak kullanmaya başladım bu özelliği, ve çok hoşuma gitti bu kullanım. Son medyaya erişim de sidebardan sürükle bırak kadar kolay olunca, daha yaratıcı notlar yaratmak çok kolay geldi. Klasörleyip Spacelerimden birinde defter gibi kullanmaya başladım bile :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aVfN3KIGACqNVdwAjX3M0A.png" /><figcaption>Örnek Easel Kullanımı</figcaption></figure><p>Arc ile ilgili ilk izlenimlerim bu şekilde, özelliklerini henüz keşfetme aşamasında olduğumu düşünerek atladığım çok şey olabileceğini düşünüyorum. Örneğin bahsettiğim Boost özelliğini, çok fazla bir kullanım şansım henüz olmadı. Ancak şu birkaç günlük deneyimimde, sidebar kullanımına alışmaya çalışmak dışında kafamı karıştıran ya da kullanımımı kötüleştiren bir deneyimim de henüz olmadı. Arc’ın ilk açılımda sanat içerikli sitelerle beni karşılamış olması bile en başta beni etkilediğini düşünürsek çok da seçici olduğumu iddia edemeyeceğim ama… Şu an için Arc hala Beta aşamasında ve bekleme listesi olan bir app bu nedenle çok daha iyi hale geleceğini düşünüyorum. Her beta kullanıcısına 5 davet hakkı veriyor, benim önerim; bayinizden isteyiniz!</p><p><em>Tarayıcıların tarihi ile ilgili, Windows’un başını ağrıtan Browser Wars’ı duydunuz mu? Firefox’un atası, bir bakıma Opera ile de bağları olan Netscape’in hikayesini, topluluk haline gelmesini öğrenmek ister misiniz? Bu konu ilginizi çektiyse gayet eğlenceli anlatımı ile 95 yılındaki ilk kısmına dokunan Valley of the Boom dizisini önerebilirim. Diziden konuyu rap ile anlatan bir </em><a href="https://youtu.be/azGAHZluMeA"><em>sahne</em></a><em>. Zamanında bir-iki resim ekleyerek yaptığım bu olay konulu basit oyuna ise </em><a href="https://withkoji.com/~terraluna/browser-wars"><em>burdan</em></a><em> ulaşabilirsiniz. Netscape!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b535d399ab0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>