<?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 Elyes Manai on Medium]]></title>
        <description><![CDATA[Stories by Elyes Manai on Medium]]></description>
        <link>https://medium.com/@manai-elyes?source=rss-2542a6ba20e3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*6j6yOAwvZh9YVfVEsTLuGA.jpeg</url>
            <title>Stories by Elyes Manai on Medium</title>
            <link>https://medium.com/@manai-elyes?source=rss-2542a6ba20e3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 20:03:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@manai-elyes/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[Making AI Safer: Reducing Critical Misclassifications with Targeted Loss Functions]]></title>
            <link>https://manai-elyes.medium.com/making-ai-safer-reducing-critical-misclassifications-with-targeted-loss-functions-afcc22667bb3?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/afcc22667bb3</guid>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Thu, 25 Sep 2025 23:47:34 GMT</pubDate>
            <atom:updated>2025-09-25T23:47:34.583Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Problem: AI Models Still Make Dangerous Mistakes</h3><p>Deep learning models power everything from facial recognition to medical diagnostics. Their accuracy is impressive, but when they fail, they can fail <strong>catastrophically</strong>. In 2015, Google Photos infamously mislabeled a Black couple as “gorillas.” More recently, Google Gemini was criticized for generating historically inaccurate “diverse Nazis.”</p><p>These errors aren’t just embarrassing — they’re <strong>dangerous</strong> in safety-critical domains like healthcare, finance, or cybersecurity. The root issue? AI models treat all misclassifications as equal. But in real life, <strong>some mistakes are far more harmful than others</strong>.</p><h3>Our Idea: Penalize the Right Mistakes</h3><p>Instead of retraining endlessly or relying only on post-hoc explainability, we propose (<a href="https://ieeexplore.ieee.org/document/10707892/">link</a>) a <strong>proactive training method</strong>:</p><ul><li>Identify label pairs where misclassification is unacceptable (e.g., “human” → “gorilla”).</li><li>Modify the <strong>loss function</strong> so that these errors are penalized more heavily.</li><li>Guide the model to avoid these mistakes — without retraining from scratch or adding costly manual fixes.</li></ul><p>This method works on <strong>already trained models</strong> too, making it cheaper and faster for organizations to audit and improve AI before deployment.</p><h3>How It Works</h3><p>We add targeted penalties into the loss function:</p><ul><li><strong>Linear penalty</strong> — for mild but important cases.</li><li><strong>Square penalty</strong> — for cases where errors are serious.</li><li><strong>Exponential penalty</strong> — for absolutely critical errors where even a single mistake is unacceptable.</li></ul><p>Example: If label A should never be classified as label B, and the model makes 10 mistakes, applying a <em>square penalty</em> treats it as if it made 100 mistakes. This forces the model to adjust its decision boundaries.</p><h3>Case Study: Synthetic Dataset, Real Results</h3><p>We tested this on a synthetic dataset (1M samples, 20 classes). Three label pairs with the highest misclassification rates were penalized.</p><ul><li><strong>Baseline (original model):</strong> 379 critical misclassifications.</li><li><strong>Continued training (1 epoch with penalties):</strong> only <strong>133 misclassifications</strong> (a 65–82% reduction across pairs).</li><li><strong>Retrained from scratch with penalties:</strong> results were unstable, showing the method works best when applied to existing trained models.</li></ul><p>Performance metrics (accuracy, precision, recall, F1) barely dropped (~0.1–1.5%), proving that large safety gains can be achieved with minimal trade-off.</p><h3>Why This Matters</h3><p>This approach provides:</p><ul><li><strong>Fairness &amp; robustness</strong> — reduces harmful biases and critical errors.</li><li><strong>Low cost &amp; speed</strong> — works within a single epoch on existing models.</li><li><strong>Practicality</strong> — easy to implement in real-world AI pipelines with Keras/TensorFlow.</li></ul><p>For companies, this means <strong>auditing and strengthening AI before deployment</strong> without expensive retraining or complex explainability pipelines.</p><h3>Looking Ahead</h3><p>Next steps include:</p><ul><li>Smarter penalty assignment strategies.</li><li>Real-time sensitivity analysis to balance performance vs. safety.</li><li>Testing on high-stakes domains (medical imaging, fraud detection, autonomous driving).</li></ul><h3>Final Takeaway</h3><p>Not all mistakes are equal. By <strong>teaching AI which errors are unacceptable</strong>, we can make models safer, fairer, and more accountable — without slowing innovation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=afcc22667bb3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Helping Vulnerability Analysts Work Faster with Explainable AI]]></title>
            <link>https://manai-elyes.medium.com/helping-vulnerability-analysts-work-faster-with-explainable-ai-74a2a1060173?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/74a2a1060173</guid>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Thu, 25 Sep 2025 23:46:50 GMT</pubDate>
            <atom:updated>2025-09-25T23:46:50.778Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Growing Problem: Too Many Vulnerabilities, Too Little Time</h3><p>Every year, the number of reported cybersecurity vulnerabilities keeps growing. In the U.S. National Vulnerability Database alone, there are now <strong>over 200,000 CVEs</strong> — and the number is accelerating</p><p>Each vulnerability must be assigned a <strong>CVSS score</strong> by a CVE Numbering Authority (CNA). This score determines how severe the vulnerability is and whether organizations treat it as a critical risk or a minor issue.</p><p>But here’s the catch: scoring is slow. A single CVE can take <strong>weeks</strong> to analyze, and severe vulnerabilities take even longer. This backlog leaves organizations exposed to risks and stretches analysts thin.</p><h3>Why Explainability Matters</h3><p>Some researchers tried automating CVSS scoring with Deep Learning. The models achieved high accuracy — but they were <strong>black boxes</strong>. Analysts can’t trust a model if they don’t understand how it made a decision, especially when their reputations and compliance obligations are on the line (think <strong>GDPR in Europe</strong> or <strong>DARPA’s Explainable AI push in the U.S.</strong>)</p><p>Our goal was different: build not only an accurate system but also an <strong>explainable one</strong>.</p><h3>What We Built</h3><p>We designed an approach (<a href="https://www.mdpi.com/2076-3417/14/20/9231">link</a>)using:</p><ul><li><strong>XGBoost</strong>, a lightweight and interpretable ML model.</li><li><strong>SHAP (Shapley values)</strong> to generate explanations at both global and local levels.</li><li>A <strong>multi-class explanation method</strong> tailored for CVSS metrics.</li><li><strong>IF rules</strong> derived directly from Shapley values, giving analysts actionable insights.</li></ul><p>We also released:</p><ul><li>The <strong>largest open CVSS dataset</strong> at the time of publishing (197,755 CVEs, version 3.1).</li><li>Full <strong>open-source code</strong> for data collection, preprocessing, modeling, and explainability.</li></ul><h3>Key Contributions</h3><ol><li><strong>Global Explanations:</strong> Show how close a vulnerability was to being classified differently — helping analysts catch anomalies.</li><li><strong>Local Explanations:</strong> New SHAP-based methods that highlight the <em>exact words</em> driving each classification.</li><li><strong>Overcoming SHAP’s text limitations:</strong> We enhanced reliability by mixing Shapley values with frequency and confidence statistics.</li><li><strong>IF Rules:</strong> Generated directly from explanations, giving analysts shortcut checks for subsets of vulnerabilities.</li><li><strong>Performance:</strong> Achieved <strong>state-of-the-art accuracy and F1</strong> scores with XGBoost — without the heavy cost and opacity of deep learning.</li><li><strong>Efficiency:</strong> Training models took <strong>seconds to minutes</strong> on a laptop, making continuous updates feasible for real-world CNAs</li></ol><h3>Why This Matters for CNAs</h3><ul><li><strong>Speed:</strong> Analysts can generate CVSS scores much faster.</li><li><strong>Confidence:</strong> Transparent explanations let them defend decisions if challenged.</li><li><strong>Compliance:</strong> Meets legal and operational requirements for explainability.</li><li><strong>Scalability:</strong> Lightweight models can be retrained frequently without costly infrastructure.</li></ul><p>In short: <strong>AI doesn’t replace analysts — it makes them faster and more confident.</strong></p><h3>What’s Next</h3><p>Future work will focus on:</p><ul><li>Extending explanations to CVSS v4.</li><li>Building analyst-friendly dashboards.</li><li>Combining XAI with automated <strong>alert correlation</strong> for SOCs.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74a2a1060173" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Faster, Smarter, and Explainable: Sampling Strategies for Intrusion Detection]]></title>
            <link>https://manai-elyes.medium.com/faster-smarter-and-explainable-sampling-strategies-for-intrusion-detection-a9aa7bf8b5e7?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/a9aa7bf8b5e7</guid>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Thu, 25 Sep 2025 23:45:24 GMT</pubDate>
            <atom:updated>2025-09-25T23:45:24.319Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Challenge: Speed Meets Explainability</h3><p>Intrusion Detection Systems (IDS) sit at the frontline of cybersecurity. They must spot attacks fast, explain <em>why</em> they flagged something, and keep pace with ever-changing traffic patterns.</p><p>Modern IDS powered by AI can analyze more data than ever, but their <strong>explanations come at a cost</strong>. Techniques like SHAP provide powerful feature attributions (which feature influenced a prediction and by how much), but generating these explanations can take minutes per sample — far too slow for real-time defense.</p><p>This creates a pressing question: <em>Can we make explainability faster without losing accuracy?</em></p><h3>Our Approach: Smarter Sampling</h3><p>Instead of running feature attribution on an entire dataset, what if we could select a <strong>small but representative subset</strong> of data?</p><p>We studied (<a href="https://ieeexplore.ieee.org/document/11141916/">link</a>) different <strong>sampling strategies</strong> to see which delivers explanations that closely mirror full-dataset results but at a fraction of the time:</p><ul><li><strong>Random Sampling</strong> — simple, but inconsistent.</li><li><strong>Monte Carlo Sampling</strong> — strong results, but high variance.</li><li><strong>Weighted Adaptive Sampling</strong> — tuned for diversity, sometimes unstable.</li><li><strong>Latin Hypercube Sampling (LHS) &amp; Improved LHS</strong> — structured, efficient, and surprisingly effective.</li></ul><p>We tested these on the <strong>UNSW-NB15 intrusion dataset</strong>, using <strong>XGBoost + SHAP</strong> to generate explanations.</p><h3>Key Findings</h3><ol><li><strong>Tiny samples can be enough.</strong><br> LHS and Improved LHS delivered <strong>near-identical feature importance explanations using only 0.1% of the training set</strong>.</li><li><strong>Stability matters.</strong><br> While Monte Carlo and Weighted Sampling sometimes outperformed others, they also showed high variance and produced unreliable outliers.</li><li><strong>Speed is no longer the bottleneck.</strong><br> Even with tens of thousands of samples, explanation times stayed near one second. For cybersecurity, that’s a game-changer.</li><li><strong>Better generalization.</strong><br> In some cases, models trained on LHS-selected subsets <strong>outperformed models trained on the full dataset</strong>, proving that smarter sampling can improve learning by removing noise.</li></ol><h3>Why This Matters for Cybersecurity</h3><p>For SOC analysts, explanations are not optional — they are the key to <strong>trusting AI alerts</strong>. But explanations must also be <strong>fast enough</strong> to be actionable in live scenarios.</p><p>This research shows that with the right sampling strategy:</p><ul><li><strong>Explainability can keep up with real-time detection.</strong></li><li><strong>Smaller data = faster retraining.</strong> Models can adapt more quickly to new attack patterns.</li><li><strong>Lower resource costs.</strong> Less data means lighter computation — ideal for environments with constrained resources or federated learning setups.</li></ul><h3>Looking Ahead</h3><p>The results point to three future directions:</p><ol><li><strong>Benchmarks:</strong> Establish standardized tests for sampling strategies across XAI methods and datasets.</li><li><strong>Hybrid sampling:</strong> Combine model-driven insights with structured methods like LHS.</li><li><strong>Deployment in SOCs:</strong> Integrate fast-sampling explainability into operational IDS pipelines.</li></ol><h3>Final Takeaway</h3><p>Intrusion detection doesn’t have to choose between speed and explainability. With optimized sampling strategies like <strong>Latin Hypercube Sampling</strong>, we can have both — <strong>fast, precise, and trustworthy AI explanations</strong> that keep defenders one step ahead.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a9aa7bf8b5e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Smarter Intrusion Detection: Explainability Through Ensemble Learning]]></title>
            <link>https://manai-elyes.medium.com/smarter-intrusion-detection-explainability-through-ensemble-learning-64ab94de8826?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/64ab94de8826</guid>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Mon, 22 Sep 2025 04:22:37 GMT</pubDate>
            <atom:updated>2025-09-22T04:22:37.310Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why Explainability Matters in Cybersecurity</h3><p>Cybersecurity today depends heavily on <strong>AI-driven intrusion detection systems (IDS)</strong>. These models sift through massive volumes of network traffic, spotting malicious patterns in real time. But the catch? The more accurate the models, the harder they are to understand.</p><p>For security analysts, this “black box” problem is unacceptable. It’s not enough to know <em>what</em> the AI predicts — we need to know <em>why</em>. Without explainability, trust erodes, false alarms increase, and effective response slows down.</p><p>This is where our research (<a href="https://ieeexplore.ieee.org/document/10708602/">link</a>)steps in: combining <strong>ensemble learning</strong> with <strong>explainability techniques</strong> to get the best of both worlds.</p><h3>The Core Idea: Confusion Matrix + Ensemble Learning</h3><p>Our method introduces a <strong>new ensemble approach</strong> that leverages:</p><ol><li><strong>Confusion Matrix</strong> — to capture how each model performs (true positives, false negatives, etc.).</li><li><strong>Statistical Means</strong> — to measure how close a new input is to past patterns.</li><li><strong>Dynamic Model Selection</strong> — instead of blindly voting or stacking, we pick the <em>most suitable model</em> per input based on these insights.</li></ol><p>This makes the ensemble both <strong>smarter and more transparent</strong>: we can explain <em>why</em> a particular model was chosen for a given prediction.</p><h3>Case Study: UNSW-NB15 Intrusion Detection</h3><p>We tested our approach on the <strong>UNSW-NB15 dataset</strong>, a benchmark in intrusion detection research.</p><h4>Models considered:</h4><ul><li>KNN (nearest neighbors)</li><li>SVM (support vector machine)</li><li>Logistic Regression</li><li>Naive Bayes</li><li>Decision Tree</li><li>XGBoost</li></ul><p>After baseline testing, <strong>SVM and XGBoost</strong> emerged as the strongest candidates. Instead of merging them in the usual ways (hard voting, soft voting, or stacking), we applied our <strong>confusion-matrix-driven ensemble</strong>.</p><h4>Results:</h4><ul><li>SVM: Accuracy 83.1%, Recall 82.0%, F1 83.5%</li><li>XGBoost: Accuracy 82.1%, Recall 76.8%, F1 82.7%</li><li>Stacking: Accuracy 82.3%, Recall 77.1%, F1 82.9%</li><li>Hard Voting: Accuracy 75.9%, Recall 67.3%, F1 76.7%</li><li><strong>Our method:</strong> Accuracy <strong>85.2%</strong>, Recall <strong>85.2%</strong>, F1 <strong>85.5%</strong></li></ul><p>The gains weren’t just numbers — our method made the decision process more interpretable, showing which model was chosen and why.</p><h3>Trade-offs and Insights</h3><p>One observation: while accuracy, recall, and F1 improved, <strong>precision dropped</strong>. This suggests the system caught more true positives but also flagged more false positives. In cybersecurity, this trade-off is often acceptable: missing attacks (false negatives) is usually riskier than raising extra alerts.</p><p>The key point: our method explains <em>why</em> certain alerts are flagged, giving analysts confidence in filtering noise.</p><h3>Why This Matters for Industry</h3><p>Unlike many XAI methods that require model internals, our approach is:</p><ul><li><strong>Model-agnostic:</strong> works with any algorithms.</li><li><strong>Scalable:</strong> can handle multiple models and datasets.</li><li><strong>Transparent:</strong> confusion matrix and means are easy to understand, even outside research labs.</li></ul><p>For Security Operations Centers (SOCs), this means explainable AI that can be plugged into existing workflows without exposing proprietary models or requiring extra compute resources.</p><h3>What’s Next</h3><p>We see three exciting directions:</p><ol><li><strong>Robust statistics:</strong> Testing median and standard deviation instead of just mean, to handle noisy data better.</li><li><strong>Genetic algorithms:</strong> Optimizing ensemble composition dynamically for evolving threats.</li><li><strong>False positive control:</strong> Adding constraints to balance recall gains with precision stability.</li></ol><h3>Final Takeaway</h3><p>This research shows that <strong>explainability and performance are not at odds</strong>. By rethinking ensemble learning with confusion matrix insights, we created an IDS approach that’s not only more accurate but also more trustworthy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=64ab94de8826" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Confusion Matrix as a Tool for Smarter Intrusion Detection]]></title>
            <link>https://manai-elyes.medium.com/confusion-matrix-as-a-tool-for-smarter-intrusion-detection-eedc901524d1?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/eedc901524d1</guid>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Mon, 22 Sep 2025 04:19:33 GMT</pubDate>
            <atom:updated>2025-09-22T04:19:33.453Z</atom:updated>
            <content:encoded><![CDATA[<p>AI has become central to modern cybersecurity, especially in <strong>network intrusion detection systems (NIDS)</strong>. These systems must detect attacks in real time, reduce false alarms, and scale with ever-growing network traffic.</p><p>But there’s a problem: the best-performing AI models are often <strong>black boxes</strong>. Analysts see predictions, but not the reasoning behind them. This lack of transparency undermines trust and slows down decision-making.</p><p>That’s where <strong>Explainable AI (XAI)</strong> comes in. And in our recent research (<a href="https://ieeexplore.ieee.org/document/10708595/">link</a>), we showed that one of the simplest tools in machine learning — the <strong>confusion matrix</strong> — can do far more than just report metrics.</p><h3>From Metric to Methodology</h3><p>Traditionally, the confusion matrix is used to summarize performance with four numbers:</p><ul><li><strong>TP (True Positives)</strong> — Attacks correctly detected.</li><li><strong>TN (True Negatives)</strong> — Normal traffic correctly identified.</li><li><strong>FP (False Positives)</strong> — False alarms.</li><li><strong>FN (False Negatives)</strong> — Missed attacks.</li></ul><p>Instead of treating it as a <strong>post-hoc scorecard</strong>, we used the confusion matrix as a <strong>diagnostic tool</strong> to analyze <em>why</em> the model makes errors and <em>how</em> to fix them. Our method:</p><ol><li><strong>Group predictions</strong> into TP, TN, FP, FN.</li><li><strong>Analyze feature distributions</strong> for each group.</li><li><strong>Spot anomalies</strong> (features where FP/FN differ sharply from TP/TN).</li><li><strong>Propose preprocessing tweaks</strong> (e.g., new encodings, outlier removal, selective filtering).</li><li><strong>Retrain and measure improvements.</strong></li></ol><h3>Case Study: UNSW-NB15 Dataset</h3><p>We tested this method on the <strong>UNSW-NB15 dataset</strong>, a benchmark for intrusion detection. Using XGBoost as the base model:</p><ul><li><strong>Baseline accuracy:</strong> 81.9%</li><li><strong>Baseline recall:</strong> 76.4%</li><li><strong>Baseline F1-score:</strong> 82.5%</li></ul><p>By analyzing the confusion matrix:</p><ul><li>We found that <strong>“ackdat” and “synack”</strong> values above 10,000 caused frequent false negatives.</li><li>We discovered that if <strong>“is_im_ips_ports” ≠ 0</strong>, the traffic was almost always benign — yet the model didn’t learn this clearly.</li></ul><p>From these insights, we tried two fixes:</p><ul><li><strong>Binary encoding</strong> of “ackdat” and “synack” (thresholding).</li><li><strong>Removing rows</strong> where “is_im_ips_ports” &gt; 0 (only 1.1% of data, but noise-heavy).</li></ul><h3>The Results</h3><ul><li>Suggestion A (binary encoding): negligible effect.</li><li>Suggestion B (row removal): <strong>major gains</strong>.</li><li>Accuracy jumped to <strong>89.6%</strong></li><li>Recall to <strong>89.2%</strong></li><li>F1-score to <strong>89.8%</strong></li></ul><p>That’s an <strong>8-point leap in recall</strong> — critical for intrusion detection where missed attacks are the most dangerous.</p><h3>Why This Matters</h3><p>This approach turns the confusion matrix into more than just a report card — it becomes an <strong>explainability engine</strong>:</p><ul><li><strong>For practitioners:</strong> It shows exactly where the model struggles and how to fix it without complex tools.</li><li><strong>For industry:</strong> It’s lightweight, requires no access to the model internals, and can be deployed in SIEM or SOC workflows.</li><li><strong>For research:</strong> It bridges explainability with performance, showing they can reinforce each other instead of being a trade-off.</li></ul><h3>Looking Ahead</h3><p>Our next step is automating this process — building systems that continuously analyze confusion matrices and generate actionable improvement suggestions. Beyond network intrusion, this method could apply to fraud detection, malware classification, and medical diagnostics.</p><p>In all these domains, <strong>explainability + performance = trust</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eedc901524d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Feature Encoding Matters for Explainable Malware Detection]]></title>
            <link>https://manai-elyes.medium.com/why-feature-encoding-matters-for-explainable-malware-detection-7dd85714c4dc?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/7dd85714c4dc</guid>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Mon, 22 Sep 2025 04:13:58 GMT</pubDate>
            <atom:updated>2025-09-22T04:13:58.017Z</atom:updated>
            <content:encoded><![CDATA[<p>Explainable AI (XAI) is about making black-box models transparent so that humans can understand how decisions are made. Most research focuses on algorithms and explanation methods, but one critical factor is often overlooked: <strong>data preprocessing</strong>. Specifically, the way we <strong>encode features</strong> before training can change not only the performance of a model but also how interpretable its explanations are.</p><p>Our recent study (<a href="https://ieeexplore.ieee.org/document/10193964">link</a>) explored this question in the context of <strong>malware classification</strong> — a domain where explainability is crucial for analysts tasked with stopping evolving cyber threats.</p><h3>Label Encoding vs One-Hot Encoding</h3><p>We compared two popular encoding methods:</p><ul><li><strong>Label Encoding (LE):</strong> assigns each category an integer. Simple and compact.</li><li><strong>One-Hot Encoding (OHE):</strong> creates a binary column for each possible value, making features explicit.</li></ul><p>Using the Microsoft Malware Classification dataset (19,611 files, 78 features), we trained two <strong>XGBoost</strong> models, one for each method, and analyzed the results with <strong>SHAP</strong> explanations.</p><h3>Key Findings</h3><p><strong>1) Performance trade-off was minimal.</strong></p><ul><li>Label Encoding F1-score: <strong>0.991</strong></li><li>One-Hot Encoding F1-score: <strong>0.988</strong><br> A small drop in accuracy with OHE, but not game-breaking.</li></ul><p><strong>2) Explanations get sharper with OHE.</strong></p><ul><li>With LE, we know which <em>feature</em> matters.</li><li>With OHE, we know which <em>feature value</em> matters.<br> For instance, instead of just “Operating System Version” being important, OHE pinpoints that <em>Version 3</em> drives decisions.</li></ul><p><strong>3) Smaller and faster explanation files.</strong><br> Despite OHE’s higher dimensionality, the extracted IF-rules and explanation files were shorter and lighter, reducing analyst workload.</p><p><strong>4) Better local and global insights.</strong><br> OHE explanations revealed specific rules and anomalies faster, making it easier to spot potential vulnerabilities and biases.</p><h3>Why This Matters for Cybersecurity</h3><p>In cybersecurity, explainability isn’t academic — it’s operational. Analysts must trust AI tools, and small improvements in clarity can mean faster response times during an attack. Our results suggest that even a <strong>tiny performance trade-off is worth the gain in interpretability</strong>.</p><p>By switching feature encoding, we unlocked:</p><ul><li>Clearer rules for analysts.</li><li>Faster file analysis.</li><li>Stronger foundations for building trust in AI-driven malware detection.</li></ul><h3>Looking Forward</h3><p>This research is just a first step. Future work should:</p><ul><li>Explore <strong>more encoding methods</strong> (binary, embeddings, hybrids).</li><li>Build <strong>visualizations</strong> tailored for cybersecurity analysts.</li><li>Test how encoding impacts other XAI tasks beyond malware detection.</li></ul><p>The takeaway? <strong>Preprocessing choices shape the story that AI tells us.</strong> If we want AI to be both powerful and explainable, we need to treat feature encoding as more than a technical detail — it’s a design decision that impacts trust, usability, and ultimately, security.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7dd85714c4dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The ability to change people’s lives and leave one’s mark]]></title>
            <link>https://medium.com/google-developer-experts/the-ability-to-change-peoples-lives-and-leave-one-s-mark-a1d2533afb66?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1d2533afb66</guid>
            <category><![CDATA[growth]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[change]]></category>
            <category><![CDATA[impact]]></category>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Wed, 08 Dec 2021 21:22:53 GMT</pubDate>
            <atom:updated>2021-12-08T21:22:53.226Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How Google helped me accelerate my impact</em></p><p>“Will this ever be useful for your career?” is what my mom would tell me each time she saw me coming home tired from organizing an event or staying up all night to prepare a workshop.</p><blockquote>Well, guess what mom? It was! and I’ve now become the youngest Google Developer Expert in Africa and MENA! (<a href="https://developers.google.com/community/experts/directory/profile/profile-elyes-manai">Link to my profile</a>)</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/539/1*twDXQe1oATRMRKNzHn5OsA.png" /></figure><p>Like most youngsters out there, I grew up in an environment where school grades were the most important things to have and everything else a waste of time. Be it my family, my friends or my colleagues, all of them saw that those that dedicate time to do community activities were wasting their time. And why wouldn’t they? When you grow up in such an ecosystem, success stories are brushed off as luck, and failure stories are showcased to prove their point.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t_QEvG1yNCGgO8oH_MCUmQ.png" /><figcaption>Screenshot from the film “3 idiots” I highly recommend watching it</figcaption></figure><p>This results in generations of people that are scared of exploring what is possible and relying on the existing system. That system however can’t give scholarships to everyone, and only rewards the top achievers out of tens of thousands. In that case, what can you do if you’re neither a top student nor have connections in industry? You’ll have to accept whatever options are left, and as sad as it sounds, it’s the reality for millions of people and thousands of Tunisians.</p><p>I was lucky to have been involved in events and cultural activities from a young age, where I managed to befriend older people who would share their experiences. I remember the first time hearing that I could actually find any course material online and not need to follow the teacher as the sole source of information, I was shocked! When I actually started exploring this option, it was like a whole new world of opportunities opening up. Being able to learn what I want when I want however I want, just by typing some keywords online and being patient. By the time I reached University, I was joining classes to ask detailed questions and start discussions with the professor instead of discovering the subject.</p><p>This allowed me to actually like what I study and get good grades while having the flexibility of doing extra activities, which later turned out to be a luxury when compared to my peers. Adults in my parent’s generation weren’t brought up with the access to information that we now have with Google, YouTube, Medium, etc.. and so my parents found my lifestyle weird, but they would allow me to live it as long as I kept getting good grades. Thank you mom and pops!</p><p>Anyways, with that superpower in hand and my network of older friends, I managed to know, at the very start of my University life, what to focus on for Industry, which allowed me to be efficient in balancing my studies, my social life, and my career. The most important thing I learned very early on was that however technologically advanced we may be and become, life is still about communicating with humans, and that HR managers, developers, and CEOs are all in the end humans, with different biases. I, therefore, focused a lot on improving my communication skills which led to experimenting with Education, hosting small-scale workshops, and doing an internship teaching Web Development that then turned into a part-time management job.</p><p>I completely fell in love with Education and knew that is what I wanted to do. This early revelation helped me choose my path in life. My love language being achievements, I tried my best to level up as fast as I could, from small workshops to big conferences, to YouTube videos and Medium articles, to organizing big AI events.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/699/0*0FAYjhWzu56VXSHC" /><figcaption>Some of my talks and events</figcaption></figure><p>As I kept doing this, I started getting recognition for my work and was contacted for opportunities. This reinforced my belief that I was doing the right thing. After 2 years of climbing up the ladder and putting myself out there, I woke up to find an email that I was recommended by a Google employee for the Google Developer Expert program, a prestigious community of technical experts with only 3 in Tunisia at the time. After 4 months, I became a GDE, the youngest in Africa and MENA.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/735/1*kCwAICvDFWA2Aca8oe-jdg.png" /></figure><p>That was the turning point for me. A sign that my work was not in vain. While my peers were sacrificing their fun and time to get recognition, it got to me by itself. This was proof that what I was believed in was the right thing. A light of hope for every free spirit out there that wanted to follow my steps but were too afraid of the consequences. I was the living proof that giving back to the community and devoting time and soul to the creation of added value for others was recognized, rewarded and respected, and by one of the biggest names in Industry.</p><p>Ever since, I continued meeting more GDEs and I started seeing how like-minded we are, and how devoted to helping we were even when most were very busy. I started being invited to exclusive events, getting tons of perks, and having access to international experts has become easy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/874/1*2padL-chYVHjIyTJpyAQjg.png" /></figure><p>I was invited to become a mentor at the Google Startup Accelerator, a task I loved.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/497/1*rk4EftCh64mcTfa2Einw9w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/434/1*Svu6onE0C0AXVKGvEt9WHg.png" /></figure><p>I was given vouchers and assistance to take certifications. I was able to sponsor events by handing out some credits I got and was able to help projects progress.</p><p>I was given the ability to invite big names to local events. All of this was very difficult for us Tunisians as we did not have powerhouses locally that allowed easy access to the international market.</p><p>Adding GDE to my introductions has made it easier for me to find better job opportunities and invitations to events. It allowed me to more easily contact people I would have normally a hard time reaching, and it gave me a sense of credibility every time I was presented and invited. All in all, Becoming a Google Developer Expert has made my life better and allowed me to help more efficiently the local Tech Community, especially in Data Science.</p><p>This is a call to all the free spirits roaming out there, facing difficulties growing in constrained environments, and having to sacrifice in order to push their careers. The world needs you, and organizations like Google are here to help you reach your potential. Don’t let the negative echo chambers affect you, because everything you’re doing will come back to you in the future and you’ll be glad you persisted.</p><p>Google Developers Experts is but one of the many programs out there, and the people you’ll find in these programs are the ones that will listen to you and suggest solutions. We’ve been where you are, and we won’t let your potential go to waste. If you’re still reading this, we’re here for you. Contact us and keep doing the great job you’re currently doing! Godspeed :)</p><p>-Elyes Manai</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1d2533afb66" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-developer-experts/the-ability-to-change-peoples-lives-and-leave-one-s-mark-a1d2533afb66">The ability to change people’s lives and leave one’s mark</a> was originally published in <a href="https://medium.com/google-developer-experts">Google Developer Experts</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Get recognized as an ML Expert with the Google Professional ML Engineer Certificate!]]></title>
            <link>https://medium.com/google-developer-experts/get-recognized-as-an-ml-expert-with-the-google-professional-ml-engineer-certificate-c85a67e9270d?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/c85a67e9270d</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[certification]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Tue, 23 Nov 2021 01:28:05 GMT</pubDate>
            <atom:updated>2021-12-09T20:42:50.735Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JsyoB9X03fd6XdilWJmaGQ.png" /></figure><p>As the current market demand for Machine Learning engineers keeps rocketing, and the immense amount of learning content keeps getting created, the Machine Learning Sphere is becoming more and more competitive. It is even harder for recruiters to evaluate the soundness and authenticity of a candidate’s application, especially for non-domain experts. Even Tech Recruiters can have a difficult time evaluating candidates as thresholds, references, and third-party courses aren’t always reliable and audited, especially when the need is either niche or very very specific.</p><p>That’s why there is an effort by known and trusted Tech Entities to create Methods to objectively certify a person&#39;s technical skills. These methods are also known as Certificates and are very helpful to source and select the right candidates in record time. Be it tech giants like Google, Amazon, Microsoft that have a broad list of skills or specialized companies like Oracle, that focus on a few things, everyone is trying their best to keep up with the demand and help us tech enthusiasts get recognized for our skills and get our dream jobs.</p><p>In Machine Learning specifically, we have the Tensorflow Developer Certificate, which we detailed in <a href="https://medium.com/google-developer-experts/become-a-tensorflow-certified-developer-its-easier-than-you-think-5a5b1e9e2105">another post</a>, and then company-specific certifications. For Google, there’s the <a href="https://cloud.google.com/certification/data-engineer">Google Cloud Professional Data Engineer </a>and the <a href="https://cloud.google.com/certification/machine-learning-engineer">Professional Machine Learning Engineer </a>that we will detail in this post. These certifications focus on Data &amp; ML on the Google Cloud Platform or GCP.</p><h3>GCP?</h3><p>Yes, Google Cloud Platform (GCP). It is one of the most popular and used Cloud Platforms that allow you to rent infrastructure &amp; services on the Cloud. You can use GCP to rent storage, servers, cloud functions, containers and even notebooks and models. This is not counting the already existing AI services that you can roll out to your apps in some clicks. Since it is made by Google and thus natively supports Tensorflow as their ML language of choice.</p><p>GCP thus has a ton of features that can work together and is made so that both beginners and experts have an easy time exploiting it. It is thus one of the go-to solutions for many startups and tech companies. However, the sheer number of features and customizations is also a problem, as choosing the best options and combinations for each problem requires in-depth knowledge of GCP. Companies wanting to use the platform will therefore need someone that knows it well and architects solutions quickly and in a cost-effective way, especially at scale. Google Cloud Certificates allow these companies to identify and contact mentioned experts and hire them for their services fast, bypassing a whole period of trouble and stress. The <a href="https://googlecloudcertified.credential.net/">Cloud Certified Directory</a> lists all the currently certified people, and for machine learning, they’ll be looking at the Google Machine Learning Professional</p><h3>Google Cloud Professional Machine Learning Engineer</h3><p>Quite a handful as a name. Not only does this certificate prove that you have a deep understanding of Machine Learning and the Google Cloud Platform, but also that you have what it takes in terms of architecture, infrastructure, costs, deployment and that you can take strategic decisions regarding company Growth. All of the previously mentioned points will be in the test, and oh boy will it be difficult. That is one of the reasons why this certificate is highly valued, especially among Google Employees (For instance, my technical interview to become a Google Developer Expert became easier when I mentioned I passed this certificate).</p><h3>Sounds cool — More details, please!</h3><p>The certificate in itself is 2 hours long and will have about 120 questions, meaning you have about 1 minute per question. You can flag questions and come back to them later to double-check before submissions.</p><p>Questions are also randomly selected so don’t worry if you get very difficult questions at the beginning and don’t become too confident if you get easy questions. It is also easy to miss a single detail that can sabotage your answer, so stay on your toes and flag any question that you’re not 90% sure of.</p><p>You can NOT use any external resource during the test. You also can NOT have anyone with you in the room. You can NOT exit the room you are in either (I&#39;ll explain later). This means that for the duration of the exam, you will have to be 1000% focused on the exam. You can bring food and drinks with you BEFORE the exam so be sure to have anything you need ready beforehand.</p><p>Finally, results will not be shown to you after submission. You will have to wait for a couple of weeks before receiving an email, so if you need it for something (application, audit, training…) make sure to take it at least 3 weeks in advance.</p><h3>Alright, no problem yet. What’s on the exam?</h3><p>Nice try. I’d love to tell you exactly what’s on the exam, but like all certified individuals, I signed an NDA prohibiting me from sharing details. What I CAN share are somewhat general pieces of information.</p><p>The good thing is that Google regularly updates their content. When I took the exam, the exam guide was vague and I thought it would be easy. What a shock I got when I actually took the test. The current<a href="https://cloud.google.com/certification/guides/machine-learning-engineer"> Exam guide</a> however is very detailed and similar to what’s on the test really so I really recommend reading it in detail. To not repeat what’s already there however, here is my personal recollection of the type of questions:</p><p><strong>1- Machine Learning Development with Tensorflow: </strong>These types of questions test your knowledge of actual coding with TF and Keras. They can tackle modeling, preprocessing, optimization, deployment and can either be to choose the right snippets or tell what’s wrong with the code. These are some of the easier ones and you’d have no problems if you have enough TF/Keras experience, are<a href="https://medium.com/google-developer-experts/become-a-tensorflow-certified-developer-its-easier-than-you-think-5a5b1e9e2105"> Tensorflow Certified,</a> or watched and practiced along the <a href="https://www.coursera.org/specializations/tensorflow-in-practice">Tensorflow Courses made by Google’s Laurence Moroney.</a></p><p><strong>2- GCP services: </strong>These types of questions are very tricky if you’re not used to GCP’s services like AI Platform, Notebooks, ML APIs, Kubeflow, Composer… They’re tricky because the test supposes you already worked with them and will ask you what service is the best option given a situation, what pipeline of services is most optimal given a constraint, which service to choose given some features… For these questions, it is recommended you get hands-on experience with GCP. Luckily, there’s a <a href="https://cloud.google.com/free">Free Tier</a> that gives you 300$ worth of credits, more than enough to test everything out.</p><p><strong>3- Infrastructure, Deployment, and Automation: </strong>These types of questions require in-depth hands-on experience with projects as they test your ML Architecture skills. They pin you with constraints and features and ask you to choose the optimal infrastructures, deployment methods, and automation pipelines for different use cases. You will need to take into consideration costs, bottlenecks, Cloud space, Inference Speed, communication protocols, and even application architectures. Recommended preparation is simply to get real-life experience on challenging projects or get somehow involved in advanced hackathons and projects.</p><p><strong>4- Strategic Decision Making: </strong>These types of question test your ability to take long-term, delicate company decisions. They are not about code but real-life situations where your decision will affect a company strategy for years to come. You will need to take into consideration scalability, distribution, potential technical debt, team size, architecture and technology, costs, performance, public opinion… This is for seasoned CTOs and Tech Leads who saw the impact of their work in real life and saw how each variable affects the whole system. The only way to get ready for this is to be exposed to these projects.</p><p>There’s a reason why the exam recommends 3+ years of industry experience including 1 or more years designing and managing solutions using Google Cloud. Luckily, Google Cloud prepared <a href="https://docs.google.com/forms/d/e/1FAIpQLSeYmkCANE81qSBqLW0g2X7RoskBX9yGYQu-m1TtsjMvHabGqg/viewform">sample questions</a> you can prepare yourself with. It is also recommended that you discuss the technologies in the exam guide with actual experts so they help you focus on the juicy parts and not waste time.</p><h3>No wonder it’s so valuable. how does the test happen exactly?</h3><p>There are two ways to take the test: in a certified center or online. I took the online test. If you choose that option too, you will have to install an app that, upon opening, will take you to a full-screen exam software that doesn’t allow you to switch screens, toggle apps, or use shortcuts. You can only exit through specific buttons. If you started the exam, you can only quit by stopping it. Before the exam, you will be required to show identification. A person, called a proctor, will be supervising you through your webcam. Their role is to make sure you don’t cheat, don’t quit the room, and assist you in case of a problem. The proctor will ask you to show around your surroundings through the webcam to make sure you don’t have any means to cheat. Additionally, they can pause the exam at any moment if they feel you trying to take a peak or look somewhere and ask you to show your surroundings.</p><p>Once everything has been cleared, you’re taken to the actual exam page within the app and you can click to finally start.</p><h3>I took the exam. It was gruesome. Now what?</h3><p>First of all, good job persevering until the end! You can now either take some rest or start preparing for another one because it takes a couple of weeks for the Google Cloud team to review your exam. Once done, you’ll receive an email with your results. If you did well, and I sure hope you did, you will get a certificate, a badge, and your beautiful face on the <a href="https://googlecloudcertified.credential.net/"><strong>Certified Directory.</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*ZtMabcdGJVnjNOCuf7uq4A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JsyoB9X03fd6XdilWJmaGQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5OIaWDZKrxh34OTXU7Sstw.png" /><figcaption>Examples of what you get after passing the Google Cloud Exam. It does feel cool!</figcaption></figure><p>Congrats! You’re now officially a Google Cloud Expert in Machine Learning :) You can now share your achievement with the market, apply for stronger positions, or raise your rates. Don’t forget to share this article with the people that now want to become like YOU!</p><p>Link to the certification page: <a href="https://cloud.google.com/certification/machine-learning-engineer">Click here!</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c85a67e9270d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-developer-experts/get-recognized-as-an-ml-expert-with-the-google-professional-ml-engineer-certificate-c85a67e9270d">Get recognized as an ML Expert with the Google Professional ML Engineer Certificate!</a> was originally published in <a href="https://medium.com/google-developer-experts">Google Developer Experts</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Become a Tensorflow Certified Developer — It’s easier than you think!]]></title>
            <link>https://medium.com/google-developer-experts/become-a-tensorflow-certified-developer-its-easier-than-you-think-5a5b1e9e2105?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a5b1e9e2105</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[certification]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Tue, 11 May 2021 00:43:11 GMT</pubDate>
            <atom:updated>2021-10-22T21:20:07.542Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Become a Tensorflow Certified Developer — It’s easier than you think!</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kLzoq63zk1pMGcCtSbTAZQ.png" /></figure><p>Starting out in Data Science and Machine Learning is still difficult because of the different approaches one can take and are suggested. Even more difficult is having to prove your proficiency to potential clients and recruiters.</p><p>While it may be easier to prove that you know what you’re doing when discussing with fellow developers, it is much harder to convince people of how good you are otherwise. Sure, participating in Kaggle/Zindi competitions and having projects in Github is a way to prove it, but someone who doesn’t know might fear that you copy-pasted existing code. After all, that’s what developers do, right? (inside joke, take it with a spoon of salt).</p><p>One objective way to showcase your skills is by becoming certified. However, most E-learning certificates like the ones on Udacity / Coursera, etc… aren’t trustworthy as one can easily repeat the exam till they get them. Those courses and certificated are meant for self-motivation and thrill of learning. They aren’t recognized per se. At the same time, certificates created by big names like Google, Amazon, IBM, and Microsoft hold much more value, at the cost of being a bit too advanced and geared towards that company’s technologies. Not really ideal for someone still fresh and starting out!</p><p>So What do I do? I’m glad you asked! Because now we have the <a href="https://www.tensorflow.org/certificate"><strong>Tensorflow Developer Certificate</strong></a>!</p><h3>Tensorflow?</h3><p>Yes, Tensorflow (TF). One of the most popular Deep Learning frameworks out there with PyTorch and Keras (although Keras &amp; TF are now shipped together).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2U5-KsSkwYwY4ZCEbXqTkQ.png" /></figure><p>One main aspect of Tensorflow is that it’s geared towards deployment to production and thus focuses more on modularity, scalability, and pipelines. It’s the perfect tool to quickly train and deploy models at scale given the existing infrastructure. I’m not dissing the other frameworks, they’re better than TF in other aspects, but when it comes to quickly training and deploying models that can scale to millions of users, Tensorflow wins hands down. That’s why it’s so popular in industry and an excellent starting point for new Deep Learning enthusiasts.</p><p>The cool thing is that once you learn to code in Tensorflow, it’s easy for you to either switch to other frameworks specialize in an aspect of ML, both allowing you to then target more difficult certifications.</p><h3>Seems cool! So how do I get certified?</h3><p>Ooooh, I see you’re all fired up! Alright, let’s get right into it then. The steps to take the $100 certificate are:</p><ul><li>Go on the official <a href="https://www.tensorflow.org/certificate"><strong>Certification Website</strong></a><strong> </strong>and read through the website</li><li>Read very carefully the <a href="https://www.tensorflow.org/extras/cert/TF_Certificate_Candidate_Handbook.pdf"><strong>Candidate Handbook</strong></a>, it contains all you need to prepare for it.</li><li>(Optional) take the <a href="https://www.coursera.org/professional-certificates/tensorflow-in-practice"><strong>Tensorflow course on Coursera.</strong></a><strong> </strong>It will greatly enhance your chance of passing the exam!</li><li>(Optional) You can apply for an education stipend which gives you free access to the above-mentioned course + brings the price down to $50</li><li><a href="https://app.trueability.com/google-certificates/tensorflow-developer"><strong>Register for the exam</strong></a><strong>. </strong>Once you do, you have 6 months to take it. After that, you’ll need to register (and pay) again.</li><li>Once you pay, you’ll get a new handbook, this time more detailed. Read it even more carefully</li><li>Install PyCharm and in it, the Tensorflow exam extension. Both of these are <strong><em>required.</em></strong></li><li>Once you’re ready, click on begin exam and you’ll have 5 hours to work on it and submit your exam.</li></ul><h3>What should I expect?</h3><p>The exam is practical, you will work on regression, classification, Computer Vision, Natural Language Processing, and Time Series. Nothing too difficult, just enough to demonstrate proficiency in the framework and mindset. You will be graded on the quality of your models, that’s all I can tell you.</p><h3>Any tips on how to get it?</h3><p>All of us who took the exam signed an NDA so I can’t tell you what’s exactly in the exam, however, I can give you the most important pieces of information to remember that are already public:</p><ul><li>You are free to use whatever resource you normally use while coding. This means you can use Google, StackOverflow, GitHub, etc. as well as Google Colab, GCP, AWS, etc. Everything goes as long as you can download the .h5 model into your local computer and put it in the right folder.</li><li>Even though you’re allowed 5 hours, you probably won’t need it. It is to take into account slow computers and bad internet connection. So if you use Colab you’ll at least win big on training time.</li><li>Get used to writing code, not copy-pasting. While it’s tempting, the benefit of writing your own code is that it builds muscle memory in your fingers and creates a mental roadmap that allows you to start new projects autonomously.</li><li>Finally, do what it takes to take the Tensorflow Developer Course. It’s very similar to the exam.</li></ul><h3>I took the exam. That wasn’t so bad. Now what?</h3><p>you’ll receive an email in the coming days (sometimes less) with your results. If you did well, and I sure hope you did, you will get a certificate, a badge, and your beautiful face on the <a href="https://developers.google.com/certification/directory/tensorflow"><strong>Certified Directory</strong></a><strong>.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*-BdaCXjlg3eLk1jmz-a2zg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kLzoq63zk1pMGcCtSbTAZQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bBDk8j4eX-IAaDHAFVgQsg.png" /><figcaption>Examples of what you get after passing the Tensorflow Exam. It does feel cool!</figcaption></figure><p>Congrats! You’re now international :) Now go share your achievement with your community and continue leveling up your conditions. Don’t forget to share this article with the people that now want to become like YOU!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a5b1e9e2105" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-developer-experts/become-a-tensorflow-certified-developer-its-easier-than-you-think-5a5b1e9e2105">Become a Tensorflow Certified Developer — It’s easier than you think!</a> was originally published in <a href="https://medium.com/google-developer-experts">Google Developer Experts</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[TPU Research Cloud]]></title>
            <link>https://medium.com/google-developer-experts/tensorflow-research-cloud-7657cf02811e?source=rss-2542a6ba20e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/7657cf02811e</guid>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Elyes Manai]]></dc:creator>
            <pubDate>Tue, 02 Mar 2021 19:16:10 GMT</pubDate>
            <atom:updated>2021-05-17T15:18:14.767Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Get free TPU Hardware for your Deep Learning Projects with TPU Research Cloud (TFRC)</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5oBWhFajs4-HSdPryIiPmw.png" /></figure><h3>Is your hardware too limited for your big ideas?</h3><p>If you ever worked on a Deep Learning project, you are most likely familiar with the biggest bottleneck of Deep Learning: <strong>Hardware!</strong> And just like me, and everybody else, you surely have faced or are facing issues because your current hardware isn’t that performant. Well, guess what? There is a way to get <strong>FREE </strong>access to very powerful hardware From <strong>Google</strong>, thanks to their program “<a href="https://sites.research.google/trc/"><strong>TPU Research Cloud</strong></a><strong> (</strong><a href="https://sites.research.google/trc/"><strong>TRC</strong></a><strong>)</strong>”. Read on fellow developers!</p><p>As you may know, the most recent advances in Deep Learning and Machine Learning have made it so far that it became more challenging to use simple hardware such as personal computers (unless you own a monster) for Deep Learning projects. Meaning that your current CPU or GPU will be your bottleneck as they will take an enormous amount of time to train models and finetune them.</p><p>As a result, you may need more robust and performant hardware, such as Cloud VMs that you can get from Google (<a href="https://cloud.google.com/">GCP</a>), Microsoft (<a href="https://azure.microsoft.com/fr-fr/">Azure</a>) and Amazon (<a href="https://aws.amazon.com/">AWS</a>), and other similar platforms. However, those are sometimes very expensive, especially when your project is self-funded. In fact, students and young professionals who are working in companies, or on their personal projects, are struggling to get access to that kind of hardware.</p><p><strong>LUCKY FOR US! </strong>Google gave us something amazing called<strong> “</strong><a href="https://sites.research.google/trc/"><strong>TPU Research Cloud</strong></a><strong>” </strong>which is a grant that focuses specifically on Deep Learning researchers, who are facing the aforementioned issues, giving them free access to the currently most performant hardware available: The <a href="https://en.wikipedia.org/wiki/Tensor_Processing_Unit"><strong>TPU </strong>(<strong>Tensor Processing Unit</strong>)</a>. This unit is hundreds of times faster than GPUs, and we’re not getting 1 but multiple TPUs <strong>FOR FREE</strong>! All they ask of you is to thank and mention them when your project gets published, as a way to get more people to use their products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/801/0*C2D73w-tx_pwdLBH.png" /><figcaption><a href="https://cloud.google.com/tpu">Source</a></figcaption></figure><p>To apply, you simply visit their <a href="https://sites.research.google/trc/"><strong>website</strong></a><strong>,</strong> click on “apply now<strong>”</strong> and fill the <a href="https://docs.google.com/forms/d/e/1FAIpQLSeBXCs4vatyQUcePgRKh_ZiKhEODXkkoeqAKzFa_d-oSVp3iw/viewform"><strong>form</strong></a>. Your project will then get reviewed by the Google Team (the more solid your project, the better chances it has to be approved). Once approved, you will get an email with a link to activate the TPU API grant on <a href="https://cloud.google.com/gcp/">Google Cloud Platform (GCP)</a>, giving you 30-day FREE access to a cocktail of V2 and V3 TPUs. To activate the API, you will need to have an international credit card with a valid billing address. It won’t be used for any payments, but it is just a way for them to make sure that you are sincerely interested. In case you don’t have one, you can ask one of your friends or family members.</p><p>Keep in mind that in order to actually use the TPUs, you will need a Google VM and Cloud Storage, that you can also get in the Google Cloud Platform. These services are not included in the TRC grant though (Only access to TPUs is free), so you will need to pay for them. But, don’t worry! There is also a FREE TIER, where new customers get $300 in free credits to spend on Google Cloud. These free credits are valid for 1 year from the day you redeem them. Also note that $300 is more than enough for simple projects, and with adequate use, they can last you months!</p><p>So if you are a student, a researcher, or a Deep Learning enthusiast, and you don’t have that much money to spare, The TPU Research Cloud is a godsend. Even better is that as everything is on the Cloud, you can use whatever current computer you have to work on it, with the possibility of even linking Google Colab to your GCP project. In both cases most of the work will be done on the GCP, so you will have to get used to using the platform, finding documentation on Google and Tensorflow.</p><p>I personally used TPU Research Cloud during my Master to pre-train BERT models (you can imagine the workload) and I’m more than satisfied. It’s a great opportunity.</p><p>That being said, go ahead and check it for yourself! And good luck with your projects :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7657cf02811e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-developer-experts/tensorflow-research-cloud-7657cf02811e">TPU Research Cloud</a> was originally published in <a href="https://medium.com/google-developer-experts">Google Developer Experts</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>