Adaptive Signal Processing & Machine Intelligence

Instructor: Prof. Danilo Mandic

Term: Spring

Location: Imperial College London, Dept. of EEE

Course Overview

This course introduces models that go beyond classical parametric, stationary signal processing. Students encounter adaptive and nonparametric approaches capable of tracking changes in system parameters in real time — enabling online operation on real-world, non-stationary data. Special emphasis is placed on demonstrating the close links between spectral estimation, adaptive signal processing, and machine intelligence.

Role: Graduate Teaching Assistant (2025, Spring term)

Learning Outcomes

Understand and apply dimensionality reduction via autoregressive and subspace spectral estimation, and tensor decompositions for Big Data
Design gradient-based adaptive filters, complex-valued adaptive filters, recursive least squares, and constrained optimisation schemes for streaming data
Construct and train artificial neural networks, autoencoders, and deep learning architectures; explore links with tensors and Big Data
Analyse practical case studies in communications, acoustics, biomedical engineering, renewable energy, and finance
Implement algorithms in MATLAB and evaluate their performance on real-world signals

Prerequisites

Digital Signal Processing (recommended)
Advanced Signal Processing (recommended)
Basic Probability Theory & Statistics
Programming experience — MATLAB preferred

Textbooks

Course Notes & Problem/Answer Sets — Dr. D. P. Mandic
Statistical Digital Signal Processing and Modeling — M. Hayes
Recurrent Neural Networks for Prediction — D. Mandic & J. Chambers

Assessment

Coursework: 100%

Schedule

Week	Date	Topic	Materials
1		Course Introduction & Motivation Data analytics landscape; biological vs digital systems; AI in healthcare; socio-economic and environmental aspects of AI; sustainability of large-scale computation.
2		Learning from Stochastic Signals Parametric vs nonparametric models; adaptive and nonstationary signal processing; AR models; online estimation.
3		Spectral Estimation & the Fourier Transform Conventional and model-based spectral estimation; role of phase spectrum; machine-learned Fourier transform; STFFT spectrogram; Hilbert-Huang spectra.
4		Adaptive Signal Processing — Foundations LMS algorithm (Widrow 1960); interference cancellation; complex LMS; affine projection and proportionate NLMS; magnitude-only and phase-only LMS.
5		Convergence, Complex & Quaternion Adaptive Filters Convergence speed and interpretability; widely linear CLMS; Quaternion LMS for 3D/4D data; cooperative estimation over sensor networks.
6		Kalman Filtering & Recursive Least Squares Kalman filter — standard, extended, unscented, and particle variants; recursive least squares; real-time estimation of signal statistics.
7		Multi-dimensional Representations & Complex Noncircularity Need for bivariate and multivariate analysis; scatter diagrams; complex circularity as a signal fingerprint; widely linear systems.
8		Artificial Neural Networks From biological to artificial neuron; FIR adaptive filter + nonlinearity; feedforward and recurrent networks; error backpropagation; Hebbian learning.
9		Dimensionality Reduction & Autoencoders Role of nonlinearity and dimensionality; PCA as an autoencoder (bottleneck network); latent space representations; deep learning and reservoir computing.
10		Blind Source Separation & Independent Component Analysis Cocktail party effect; BSS fundamentals; ICA for non-Gaussian sources; complex-valued ICA and the role of noncircularity.
11		Tensors & Big Data Tensor construction and decomposition (CPD); curse of dimensionality; tensors vs matrices; from DNNs to tensor networks; depth efficiency.
12		Case Studies & Review Applications in satellite & mobile communications, ECG/biomedical signal processing, acoustic echo cancellation, renewable energy (wind), and financial signal processing.