BCI-Based Windows Cursor Control: EEG Signal Processing & Machine Learning

Digital Signal Processing (DSP) involves converting analog signals to digital form through sampling, where the sampling frequency must be at least twice the maximum signal frequency (Nyquist rate) to prevent aliasing; DSP uses Z-transform for system analysis and stability determination, DFT/FFT for frequency domain analysis, and filter design methods including IIR (recursive) and FIR (non-recursive) filters with techniques like windowing and bilinear transformation.

Bandpass filters allow signals to pass undistorted within specific frequency ranges while attenuating frequencies outside. For example, 440 Hz to 880 Hz keeps only that octave. Special cases include high-pass filters (keeping frequencies above threshold) and low-pass filters (keeping frequencies below threshold). This frequency-domain thinking is fundamental - systems manipulate signals by adjusting weights of different frequency components in the signal's linear combination. MATLAB's bandpass function accepts two-element vectors defining lower and upper passband frequencies. Frequencies can be specified in normalized units (0 to π radians/sample) or in Hz with sampling rate provided. This is a general-purpose filter for any desired frequency range.

DSP (Digital Signal Processor) is a chip that manipulates audio signals in both time and frequency domains. Early DSPs were expensive with fixed parameters, but modern DSPs are affordable (around $100) and offer extensive customization. Bose achieves its distinctive sound through DSP, not just physical drivers. Speakers have varying efficiency across frequencies, and DSP optimizes each speaker's performance by sending only the frequencies it handles most efficiently. Audio measurements use decibels where 0 dB means input equals output. High-fidelity equipment must maintain flat response with no more than -3 dB attenuation at 20 Hz and 20 kHz. Filters divide frequencies: Low Pass (LPF) allows low frequencies through, High Pass (HPF) allows high frequencies, and Band Pass (BPF) allows a specific range. Filter order determines slope: first-order (6 dB/octave), second-order (12 dB/octave), with higher orders providing steeper attenuation but more phase shift. DSP can manipulate signals in the time domain through phase inversion (180-degree shift) and time delay (3ms per meter). Testing requires a computer with high-quality sound card and separate headphone/microphone outputs. A frequency sweep from 5 Hz to 48 kHz verifies equipment capability.

Digital Signal Processing (DSP) is the art and science of modifying time series data for analysis or enhancement, with applications in cell phone communications, medical imaging, aircraft tracking, and video streaming. The signal processing pipeline converts analog phenomena through transducers to voltage signals, digitizes via ADC, processes through DSP chips built on filter banks, and converts back via DAC. Mathematical foundations include Fourier series (representing signals as sums of sine/cosine functions), Euler's formula (compact exponential representation), and Fourier transform (continuous frequency representation for non-periodic signals). The Laplace transform projects signals to the s-plane, with Fourier transform being a subset when σ = 0. The Z-transform is the discrete equivalent, projecting to the z-plane and reducing to Discrete-Time Fourier Transform on the unit circle. The Discrete Fourier Transform (DFT) converts continuous signals to discrete samples: X(k) = Σx(n)e^(-j2πkn/N) for k = 0 to N-1. Aliasing occurs when sampling rate is insufficient, causing high and low frequency components to mirror. Spectral leakage occurs when signals cannot be represented by integer cycles within the sampling window. Window functions mitigate leakage by tapering edge values. Filters are devices that remove unwanted frequency components: low-pass, high-pass, band-pass, and notch filters. They can be analog (hardware circuits) or digital (software-based). Ideal filters (brick-wall) cannot be realized; practical filters approximate ideal behavior with passband, stopband, and transition band. The cutoff frequency is typically defined at the -3dB point. The moving average filter is a simple FIR filter computing the average of N consecutive samples.

Digital Signal Processing (DSP) is the branch of engineering and mathematics that deals with processing digital signals, which are discrete in both time and amplitude, as opposed to analog signals that are continuous in both domains; the conversion from analog to digital involves two key processes: sampling (discretizing time using the Nyquist Criterion requiring sampling rate ≥ twice the maximum frequency to avoid aliasing) and quantizing (discretizing amplitude), and DSP finds applications in audio processing, image enhancement, and communication systems like cell phones and Wi-Fi, where filters are used to selectively allow or block specific frequency ranges (passband and stopband).

Linear Discriminant Analysis (LDA) estimates class priors and density functions, assuming classes are normally distributed with identical covariance matrices. Fisher's Discriminant Analysis finds optimal projections maximizing class separation. Logistic regression directly models class probabilities by modeling the log-odds ratio as a linear function of features. Support Vector Machines (SVM) find the maximum margin hyperplane that separates classes, maximizing distance to nearest data points (support vectors). The hyperparameter C controls the trade-off between maximizing the margin and allowing some misclassifications (soft margin), making SVM more robust to outliers.

Supervised learning involves training algorithms on labeled data where each data point has an associated label. Linear Discriminant Analysis (LDA), developed over 100 years ago, is a fundamental supervised learning technique that projects multi-dimensional data onto an optimal line (or hyperplane) to separate different classes. The core challenge is identifying the best projection direction that maximizes separation between class distributions while minimizing overlap. This projection transforms complex classification problems into simpler one-dimensional discrimination tasks, enabling effective classification decisions based on where new data points fall relative to the decision boundary.

Support Vector Machines (SVMs) are supervised machine learning algorithms that classify data by finding the optimal hyperplane—a decision boundary that separates data points of different classes in the feature space. The key principle is maximizing the margin, which is the distance between the hyperplane and the nearest data points (called support vectors) from either class. SVMs can handle both linearly separable and non-linearly separable data through kernel functions that map data into higher-dimensional spaces, enabling effective classification even when classes overlap in the original feature space.

This section covers fundamental supervised learning classification algorithms. Naive Bayes assumes conditional independence of features given class labels, using Gaussian distribution for continuous data, Bernoulli for binary data, and Multinomial for text classification with word frequencies. K-Nearest Neighbors uses Euclidean distance and K represents the number of neighbors considered. Decision Trees use Information Gain or Gini Index for node splitting. Random Forest combines multiple decision trees to improve performance. Ensemble learning improves accuracy by combining multiple weak learners rather than using a single strong learner. Support Vector Machine finds the optimal hyperplane that maximizes the margin between classes.

The supervised classification pipeline consists of: (1) labeled training data with feedback, (2) feature extraction to transform raw data into numerical matrices, (3) model selection, and (4) model training. Raw models (neural networks, random forests) have no inherent value—their value comes only after training on specific data. Feature extraction converts diverse data types (text, dates, numbers) into unified numerical representations. Models require matrices with consistent dimensions where no single dimension has inherent importance. The true value lies in data preparation and model selection, not in the model architecture itself.

CSP can be formulated and solved through multiple mathematical approaches. First, as an optimization problem: maximize W^T * Σ_c * W (projected variance for one class) subject to W^T * Σ_avg * W = 1 (normalization constraint). Second, as a generalized eigenvalue problem: jointly diagonalize the covariance matrices of the two conditions, yielding eigenvectors as spatial filters and eigenvalues representing captured variance. Third, geometrically through whitening and rotation: apply PCA to the average covariance matrix to whiten the data, then rotate using PCA on individual class distributions. Each approach provides the same set of spatial filters, with the generalized eigenvalue method being computationally most efficient.

The covariance matrix can be calculated using matrix operations by first computing the deviation matrix (original data minus the mean vector), then multiplying the transpose of the deviation matrix by the deviation matrix itself; the diagonal elements represent the variances of each variable, while the off-diagonal elements indicate the covariance between variables, with positive values showing variables that tend to vary together and negative values showing inverse relationships.

This section covers the mathematical foundations required to construct spatial covariance matrices. Key assumptions include isotropy (covariance depends only on distance, not direction) and second-order stationarity (covariance depends only on distance and not on location in space). The covariance matrix is constructed using a distance matrix D containing all pairwise distances between points and a correlation function. The matrix elements follow σ² × ρ(d; θ), where σ² is the variance component and ρ(d; θ) is the correlation component depending on distance d and parameters θ. Not all functions producing values between -1 and +1 are valid correlation functions—they must yield positive definite covariance matrices. Common valid functions include exponential and Gaussian correlation functions.

This lecture explains how to transform the Common Spatial Patterns (CSP) method into a convex optimization problem by reformulating the CSP prediction function as an inner product between a weight matrix and a covariance matrix, plus a bias term; this reformulation reveals that CSP can be viewed as a linear discriminant analysis problem where the data is the vectorized covariance matrix, but the high dimensionality (number of channels squared) makes direct LDA impractical, motivating the need for large-scale learning methods.

The covariance matrix is a square matrix that summarizes the variance and covariance of a dataset, where the diagonal elements represent the variance of each variable (measuring spread in that direction) and the off-diagonal elements represent the covariance between variables (measuring how they change together); for a weighted dataset, the mean, variance, and covariance are calculated by weighting each point's contribution by its respective weight before averaging.