https://chandrashaker.in/tags/advanced-ann/ Recent content in Advanced-Ann on Chandras Edu | AI, US Stocks, Courses & Jobs India Hugo en-in © 2026 Chandra Shaker · [Privacy Policy](/privacy-policy/) · [Disclaimer](/disclaimer/) · [Contact](/contact/) Mon, 01 Jun 2026 00:00:00 +0000 https://chandrashaker.in/courses/paiml/unit4/advanced-error-functions/ Mon, 01 Jun 2026 00:00:00 +0000 https://chandrashaker.in/courses/paiml/unit4/advanced-error-functions/ <h2 id="unit-4-navigation">Unit 4 Navigation</h2> <p><strong>← <a href="https://chandrashaker.in/courses/paiml/unit4/face-recognition/">An Example: Face Recognition</a></strong>  |  <strong>Next → <a href="https://chandrashaker.in/courses/paiml/unit4/recurrent-networks/">Recurrent Networks</a></strong></p> <hr> <h2 id="why-modify-the-error-function">Why Modify the Error Function?</h2> <p>Standard Backpropagation minimises:</p> <pre tabindex="0"><code>E(w) = (1/2) × Σ_d Σ_k (t_kd − o_kd)² </code></pre><p>This works for many cases. But we may also want to:</p> <ul> <li><strong>Reduce overfitting</strong> by penalising large weights</li> <li><strong>Enforce invariance</strong> by matching how the output varies with inputs</li> <li><strong>Output probabilities</strong> rather than arbitrary real values</li> <li><strong>Enforce symmetry</strong> across equivalent inputs</li> </ul> <p>Each new objective leads to a <strong>different error function E</strong>, and hence a <strong>different gradient descent update rule</strong>.</p>