Machinelearning

Gaussian Mixture Models This set takes inspiration from chapter 9.2 of (Bishop 2006). We assume that the reader already knows quite well what is a Gaussian mixture model and we will just restate the models here. We will discuss the problem of estimating the best possible parameters (so, this is a density estimation problem) when the data is generated by a mixture of Gaussians. $$ \mathcal{N}(x \mid \mu, \Sigma) = \frac{1}{\sqrt{ 2\pi }} \frac{1}{\lvert \Sigma \rvert^{1/2} } \exp \left( -\frac{1}{2} (x - \mu)^{T} \Sigma^{-1}(x - \mu) \right) $$Problem statement 🟩 $$ p(z) = \prod_{i = 1}^{k} \pi_{i}^{z_{i}} $$ Because we know that $z$ is a $k$ dimensional vector that has a single digit indicating which Gaussian was chosen. ...

Anomaly detection is a problem in machine learning that is of a big interest in industry. For example a bank needs to identify problems in transactions, doctors need it to see illness, or suspicious behaviors for law (no Orwell here). The main difference between this and classification is that here we have no classes. Setting of the problem Let’s say we have a set $X = \left\{ x_{1}, \dots, x_{n} \right\} \subseteq \mathcal{N} \subseteq \mathcal{X} = \mathbb{R}^{d}$ We say this set is the normal set, and $X$ are our samples but it’s quite complex, so we need an approximation to say whether if a set is normal or not. We need a function $\phi : \mathcal{X} \to \left\{ 0, 1 \right\}$ with $\phi(x) = 1 \iff x \not \in \mathcal{N}$. ...

Robbins-Moro Algorithm The Algorithm $$ w_{n+1} = w_{n} - \alpha_{n} \Delta w_{n} $$For example with $\alpha_{0} > \alpha_{1} > \dots > \alpha_{n} \dots$, and $\alpha_{t} = \frac{1}{t}$ they satisfy the condition (in practice we use a constant $\alpha$, but we lose the convergence guarantee by Robbins Moro). More generally, the Robbins-Moro conditions re: $\sum_{n} \alpha_{n} = \infty$ $\sum_{n} \alpha_{n}^{2} < \infty$ Then the algorithm is guaranteed to converge to the best answer. One nice thing about this, is that we don’t need gradients. But often we use gradient versions (stochastic gradient descent and similar), using auto-grad, see Backpropagation. But learning with gradients brings some drawbacks: ...

Backpropagation is perhaps the most important algorithm of the 21st century. It is used everywhere in machine learning and is also connected to computing marginal distributions. This is why all machine learning scientists and data scientists should understand this algorithm very well. An important observation is that this algorithm is linear: the time complexity is the same as the forward pass. Derivatives are unexpectedly cheap to calculate. This took a lot of time to discover. See colah’s blog. Karpathy has a nice resource for this topic too! ...

The perceptron is a fundamental binary linear classifier introduced by (Rosenblatt 1958). It maps an input vector $\mathbf{x} \in \mathbb{R}^n$ to an output $y \in \{0,1\}$ using a weighted sum followed by a threshold function. The Mathematical Definition Given an input vector $\mathbf{x} = (x_1, x_2, \dots, x_n)$ and a weight vector $\mathbf{w} = (w_1, w_2, \dots, w_n)$, the perceptron computes: $$ z = \mathbf{w}^\top \mathbf{x} + b = \sum_{i=1}^{n} w_i x_i + b $$where $b$ is the bias term. The output is determined by the Heaviside step function: ...

Active Learning concerns methods to decide how to sample the most useful information in a specific domain; how can you select the best sample for an unknown model? Gathering data is very costly, we would like to create some principled manner to choose the best data point to humanly label in order to have the best model. In this setting, we are interested in the concept of usefulness of information. One of our main goals is to reduce uncertainty, thus, Entropy-based (mutual information) methods are often used. For example, we can use active learning to choose what samples needs to be labelled in order to have highest accuracy on the trained model, when labelling is costly. ...

Bayesian Information Criterion (BIC) The Bayesian Information Criterion (BIC) is a model selection criterion that helps compare different statistical models while penalizing model complexity. It is rooted in Bayesian probability theory but is commonly used even in frequentist settings. Mathematically Precise Definition For a statistical model $M$ with $k$ parameters fitted to a dataset $\mathcal{D} = \{x_1, x_2, \dots, x_n\}$, the BIC is defined as: $$ \text{BIC} = -2 \cdot \ln \hat{L} + k \cdot \ln(n) $$where: ...

The beta distribution The beta distribution is a powerful tool for modeling probabilities and proportions between 0 and 1. Here’s a structured intuition to grasp its essence: Core Concept The beta distribution, defined on $[0, 1]$, is parameterized by two shape parameters: α (alpha) and β (beta). These parameters dictate the distribution’s shape, allowing it to flexibly represent beliefs about probabilities, rates, or proportions. Key Intuitions a. “Pseudo-Counts” Interpretation α acts like “successes” and β like “failures” in a hypothetical experiment. Example: If you use Beta(5, 3), it’s as if you’ve observed 5 successes and 3 failures before seeing actual data. After observing x real successes and y real failures, the posterior becomes Beta(α+x, β+y). This makes beta the conjugate prior for the binomial distribution (bernoulli process). b. Shape Flexibility Uniform distribution: When α = β = 1, all values in [0, 1] are equally likely. Bell-shaped: When α, β > 1, the distribution peaks at mode = (α-1)/(α+β-2). Symmetric if α = β (e.g., Beta(5, 5) is centered at 0.5). U-shaped: When α, β < 1, density spikes at 0 and 1 (useful for modeling polarization, meaning we believe the model to only produce values at 0 or 1, not in the middle.). Skewed: If α > β, skewed toward 1; if β > α, skewed toward 0. c. Moments Mean: $α/(α+β)$ – your “expected” probability of success. Variance: $αβ / [(α+β)²(α+β+1)]$ – decreases as α and β grow (more confidence). $$ \text{Mode} = \frac{\alpha - 1}{\alpha + \beta - 2} $$The mathematical model $$ \text{Beta} (x \mid a, b) = \frac{1}{B(a, b)} \cdot x^{a -1 }(1 - x)^{b - 1} $$ Where $B(a, b) = \Gamma(a) \Gamma(b) / \Gamma( + b)$ And $\Gamma(t) = \int_{0}^{\infty}e^{-x}x^{t - 1} \, dx$ ...

Machine learning cannot distinguish between causal and environment features. Shortcut learning Often we observe shortcut learning: the model learns some dataset dependent shortcuts (e.g. the machine that was used to take the X-ray) to make inference, but this is very brittle, and is not usually able to generalize. Shortcut learning happens when there are correlations in the test set between causal and non-causal features. Our object of interest should be the main focus, not the environment around, in most of the cases. For example, a camel in a grass land should still be recognized as a camel, not a cow. One solution could be engineering invariant representations which are independent of the environment. So having a kind of encoder that creates these representations. ...

There is a big difference between the empirical score and the expected score; in the beginning, we had said something about this in Introduction to Advanced Machine Learning. We will develop more methods to better comprehend this fundamental principles. How can we estimate the expected risk of a particular estimator or algorithm? We can use the cross-validation method. This method is used to estimate the expected risk of a model, and it is a fundamental method in machine learning. ...