<?xml version="1.0" encoding="UTF-8" ?>
<rss version="2.0">
    <channel>
      <title>Tiger103 ˚₊‧🐯.𖥔 ݁</title>
      <link>https://blogs.juha-ahmad.de</link>
      <description>Last 10 notes on Tiger103 ˚₊‧🐯.𖥔 ݁</description>
      <generator>Quartz -- quartz.jzhao.xyz</generator>
      <item>
    <title>Lecture 1.1: Introduction and Decision Theory</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/01-1-Introduction-and-Decision-Theory</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/01-1-Introduction-and-Decision-Theory</guid>
    <description><![CDATA[ Machine learning as inductive inference, inductive bias, the statistical setup (loss, true risk, Bayes risk, Bayes classifier), the regression function, the explicit Bayes classifier for the 0-1 loss, scoring functions and surrogate losses, the L2 regression function, and Bayesian decision theory with priors, likelihoods, posteriors and costs, including how to read decision boundaries off a diagram. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 1.2: Learning from Finite Samples</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/01-2-Learning-from-Finite-Samples</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/01-2-Learning-from-Finite-Samples</guid>
    <description><![CDATA[ The finite sample setting: i.i.d. training data, convergence of random variables, (universal) consistency, the plug-in classifier, empirical risk and the law of large numbers, empirical risk minimization and why it can fail, estimation and approximation error in the classical and the modern regime, the bias-variance decomposition, and the No-Free-Lunch theorem. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 2: Perceptron Mistake Bound</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/02-Perceptron-Mistake-Bound</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/02-Perceptron-Mistake-Bound</guid>
    <description><![CDATA[ Warmup for learning theory: the perceptron as SGD on the linear loss, distance to a hyperplane and the margin, Novikoff's mistake bound R²/ρ², the leave-one-out generalization bound with a 1/n rate, and the non-separable case. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 3: Statistical Learning Theory</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/03-Statistical-Learning-Theory</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/03-Statistical-Learning-Theory</guid>
    <description><![CDATA[ Bounding the complexity of the function class: consistency with respect to F, Hoeffding and McDiarmid, the error of a fixed function, uniform convergence (Vapnik-Chervonenkis), finite classes and the union bound, partitioning estimators, shattering coefficient and symmetrization, growth function, VC dimension and the Sauer-Shelah lemma, VC dimension of linear, margin and neural network classifiers, Rademacher complexity, and the limits of this approach. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 4: Algorithmic Stability</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/04-Algorithmic-Stability</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/04-Algorithmic-Stability</guid>
    <description><![CDATA[ Stability as an alternative to capacity bounds: average and uniform stability, the expected generalization gap equals the average stability, the uniform stability bound via McDiarmid (useful iff β_n = o(1/√n)), stability plus approximate ERM gives consistency, ERM with strongly convex Lipschitz losses is stable (4L²/(μn)), and stability of GD and SGD on smooth losses. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 5: Regularization</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/05-Regularization</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/05-Regularization</guid>
    <description><![CDATA[ Explicit regularization: the regularized risk R_n(f) + λΩ(f), nested function classes and consistency with λ_n → 0; linear least squares (full rank, generalized inverse, n vs. d); ridge regression (Tikhonov) with its closed form, shrinkage via the SVD and Stein's paradox; Tikhonov regularization makes ERM strongly convex and stable; sparsity, p-norms and the lasso; excess risk bounds and rates for OLS, L2, L1 and L0 regularization. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 6: Aggregation, Bagging and Boosting</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/06-Aggregation,-Bagging-and-Boosting</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/06-Aggregation,-Bagging-and-Boosting</guid>
    <description><![CDATA[ Ensemble methods: bootstrap and bagging (variance ρσ² + (1−ρ)σ²/B), spatial decision trees and random forests with their consistency, weak and strong PAC learners, AdaBoost with the training error bound exp(−2γ²T), VC and margin bounds on the test error, the equivalence of weak and strong learnability, and gradient boosting (XGBoost). ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 7: Data, Measurement and Validity</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/07-Data,-Measurement-and-Validity</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/07-Data,-Measurement-and-Validity</guid>
    <description><![CDATA[ Beyond algorithms: target construct vs. measurement (US census, health insurance, liver transplants), harms in data (bias, sampling bias, privacy, copyright), good data practices and datasheets; features and representations (explicit feature maps, basis functions, kernels, random Fourier and ReLU features, representation learning); benchmarking and the four notions of validity (statistical, internal, external, construct), recreated test sets, and validity threats in foundation model research. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 8: Overparameterized Learning</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/08-Overparameterized-Learning</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/08-Overparameterized-Learning</guid>
    <description><![CDATA[ Learning in the over-parameterized interpolation regime: puzzling results (random labels), double descent, implicit regularization (GD converges to the minimum norm solution, logistic GD to the max-margin direction), benign overfitting (constant-0 toy setup, spiky-smooth structure, isotropic and aligned linear models, four regimes), relation to classical theory, why large models (universal approximation, hardness, robust interpolation), loss landscape, neural tangent kernel. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item><item>
    <title>Lecture 9.1: Fairness</title>
    <link>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/09-1-Fairness</link>
    <guid>https://blogs.juha-ahmad.de/Statistical-Machine-Learning/09-1-Fairness</guid>
    <description><![CDATA[ Machine learning in society, part 1: the COMPAS debate and other examples, sources of unfairness, group fairness criteria (demographic parity, equalized odds, equal opportunity, predictive parity, calibration), extreme cases and impossibility results, individual and counterfactual fairness, pre-, in- and post-processing (randomized derived classifiers, ROC curves), tradeoffs and feedback loops. ]]></description>
    <pubDate>Fri, 02 Oct 2026 21:26:29 GMT</pubDate>
  </item>
    </channel>
  </rss>