The single artificial neuron with logistic activation is the basic unit of neural networks. Backpropagation computes gradients; steepest descent updates weights. Kernels extend linear methods to nonlinear classification via implicit feature maps.
The augmented data observation $\vec{x}_i = [\vec{a}_i \;\; 1]$ incorporates the bias term as the last weight. Labels $y_i \in \{0, +1\}$ for the neuron formulation.
The logistic function outputs values in $(0,1)$, making it suitable as a probability score. The squared-error objective is minimized by gradient descent.
The derivative of the logistic function has a beautiful self-referential form. It can be computed directly from the activation $z_i$ without re-computing the exponential.
$\eta > 0$ is the learning rate (step size). The update moves in the direction of steepest descent of the objective function $f$.
$\eta > 0$ is the learning rate (step size). Too large: oscillates around minimum. Too small: converges very slowly. The negative of the derivative points "downhill".
The gradient points in the direction of steepest increase. Its negative transpose points in the direction of steepest decrease — the direction we want to move.
The negative log-likelihood objective more heavily penalizes confident wrong predictions. MATLAB's glmfit uses a binomial model with the negative log-likelihood. Both objectives lead to the same logistic activation update rule.
Minsky and Papert (1969) proved entire classes of problems are unsolvable by a single perceptron. This motivated multi-layer networks and nonlinear embeddings.
Kernel functions compute similarity between observations without explicitly constructing a feature map. The Gaussian kernel maps to an infinite-dimensional feature space.
$K$ is symmetric positive semidefinite. Entry $(i,j)$ measures the kernel similarity between observations $i$ and $j$.
The centering matrix $G_m$ removes the mean from the feature space. Kernel PCA applies PCA to the centered Gram matrix $\tilde{K}$ instead of the covariance matrix.