Undergraduate Teaching 2023-24

Not logged in. More information may be available... Login via Raven / direct.

Engineering Tripos Part IIB, 4G10: Brain machine interface, 2023-24

Leader

Dr Y Ahmadian

Lecturer

Dr Y Ahmadian

Lecturer

Prof G Hennequin

Lecturer

Prof G Malliaras

Timing and Structure

15 Lectures. Assessment 100% coursework

Prerequisites

3M1, 3G3, 3F2, 3F8 useful

Objectives

As specific objectives, by the end of the course students should be able to:

  • exposure to modern technologies for interfacing with the nervous systems.
  • an in-depth understanding of a range of signal processing and probabilistic machine learning techniques, taught here in the concrete context of a real-world application: neural data analysis and brain machine interfaces.
  • an understanding of the theoretical and practical challenges involved in the design and operation of a BMI
  • exposure to the neural basis of motor control in primates
  • an appreciation of how BMIs can be used both in clinical applications and in basic neuroscience research into biological learning and neural representations.
  • hands-on experience with neural data analysis as part of the coursework

Content

This course will provide a hands-on introduction to a key set of information engineering tools in the context of brain machine interfaces (BMI), an exciting and fast developing bioengineering technology. Following introductory lectures covering an overview of relevant neural circuits and recording and stimulation technology used in BMI, the bulk of the lectures will introduce various modelling, data analysis, and decoding techniques in the context of motor-oriented BMI and motor cortex.

Introduction (1L)

  • History, overview and goals of BMI; motivating example.
  • Overview of neurons, and sensory- and motor-related circuits involved in BMI.

Hardware technology (3L)

  • Introduction to different recording & stimulation technologies (spinal, intracortical, superficial, deep brain).
  • Interface technology: motor prosthetics, sensory prosthetics, other clinical applications.

Modeling and data analysis (11L)

  • Modeling population activity (4L) ​
    • Statistical methods (probabilistic PCA, factor analysis, Gaussian Process Factor Analysis)
    • Dynamical methods (linear dynamical systems, recurrent nonlinear networks and variational autoencoders)
  • Decoding methods (2L):
    • Kalman filter, and a nonlinear/non-gaussian extension of it.
  • The full loop of BMI (1L):
    • An example model integrating both control and decoding.
  • Using BMI to learn about learning in the brain (2L).
  • Challenges of BMI: instability of neural representations (2L).

Conclusion (1L)

  • A recent impressive application as a motivating conclusion.

Coursework

Assessment will be 100% coursework: two long programming-based exercises involving (1) implementation of methods learned in lectures, (2) guided development of their extensions, and (3) analysis of real datasets with those methods.

 

Coursework Format Due date & marks

Coursework activity #1: rotational dynamics in the motor cortex

 

An influential theory holds that the motor cortex forms a dynamical system that autonomously generates its own dynamic activity patterns, which are sent as motor commands to the spinal cord to shape body movements.  But what form do these dynamics take? Many animal movements, which are evolutionarily well-preserved, are of a periodic nature: walking, running, chewing, the peristalsis of the gut, etc. This has lead to the hypothesis that the neural dynamics underlying these preserved patterns are even responsible for generating movements that are not overtly periodic. This coursework explores a dynamical dimensionality reduction method developed to uncover periodic, rotational dynamics in the motor cortex. 

 

Learning objectives:

  • implement a method for dynamical dimensionality reduction of neual recordings data,

  • use linear algebra to solve a constrained least squares minimisation problem,

  • hands-on experience with neural data analysis, and learning to interpret the patterns revealed by the analysis,

  • apply and gain an appreciation for the use of statistical controls to rule out falsely inferred structure due to overfitting.

Individual report

Anonymously marked

Posted Tue week 4

Due Wed week 6

[30/60]
 

Individual report

Anonymously marked

Posted Fri week 8

Due Sat two weeks later

[30/60]

 

Booklists

Please refer to the Booklist for Part IIB Courses for references to this module, this can be found on the associated Moodle course.

Examination Guidelines

Please refer to Form & conduct of the examinations.

 
Last modified: 05/10/2023 11:53

Engineering Tripos Part IIB, 4C11: Data-driven and learning based methods in mechanics and materials, 2023-24

Leader

Dr B Liu

Lecturer

Dr A Cicirello

Timing and Structure

Lent term. 13 lectures. Assessment 100% coursework

Prerequisites

3C7 assumed; 3D7 useful

Aims

The aims of the course are to:

  • Introduce the state-of-the-art concepts and theories for deep learning and neural networks.
  • Describe the main methods of constructing learning-based partial differential equation solvers with illustrative examples on Darcy flow and elasticity.
  • Explain the concept and theory of path dependency (memory) and multi-scale modelling, with application of the data-driven methods for discovering and approximating constitutive models for various materials.

Objectives

As specific objectives, by the end of the course students should be able to:

  • Understand the principles of applying data-driven methods to physical problems.
  • Design, implement and train learning-based PDE solvers for stress analysis.
  • Discover non-linear, path-dependent material models from data using deep neural networks.

Content

Mechanics and materials are gradually becoming data-rich due to rapid advances in experimental science and high-performance multiscale computing. There has been a growing interest in the field of solid mechanics for developing data-driven and learning-based methods to characterize, understand, model, and design material/structural systems. With data-driven approaches, it is possible to remove/relax the need for ad hoc constitutive models for describing the material behavior, to achieve fast multi-scale computation for structures as well as to generate optimal designs. This module will introduce a wide spectrum of data-driven/learning based methods that have been developed and used in mechanics and materials, with an emphasis on developing a working understanding of how to apply these methods in practice.

Syllabus

Neural network basics (4L)

  1. Basic concepts in supervised and unsupervised learning.
  2. Fully connected neural network, stochastic gradient descent.
  3. Advanced neural network architectures: convolution neural network, Res-net, U-net.
  4. Python for machine learning and pytorch tutorial.

 

Machine learning for PDEs: Physics Informed Neural Networks and Neural Operators (4L)

  1. Physics informed neural networks for ODE and PDE.
  2. Learning the solution operator of PDE with Neural Operators.
  3. Fourier and Graph Neural Operators.

 

Machine learning for path dependent problems and learning based multi scale modeling (4L))

  1. Machine learning methods for memory and path dependence.
  2. Long Short Term Memory and Transformer networks.
  3. Multiscale modeling and Recurrent Neural Operator.
  4. Generative modeling methods.

 

Data-driven methods in mechanics and beyond - guest lecture (1L)

  1. Neural operators in climate change - the earth 2 project.
  2. Researches in NVDIA.

 

Coursework

Course work 1: Neural network and Pytorch basics

Description: This course work consists of two problems: 

(i) Regression problem: Student will be provided with measured stress-strain data for two unknown elastic materials. Students are asked to build, train and validate a neural network model for approximating the constituitive relationship of the material. They will use basic fully connected neural network. 

(ii) Classification problem: Student will be asked to design, implement and train a neural network classifier that predicts whether a truss structure (Effiel tower) will collapse under certain external pressures. They will investigate the use of both basic fully connected neural network as well as advanced deep Res-net, and assess the netowrks performance.

Format: 1 individual report

Course work 2: Learning based stress analysis

Students will be asked to solve a 2D elasticity problem for a plate with hole under bi-axial loading using Physics Informed Neural Networks. They will also be asked to design and implement a Fourier Neural Operator to learn the solution operator of the Darcy flow problem. 

Format: 1 individual report

Course work 3: Learning based constitutive model for anisotropic solids

Description:  Students will be asked to come up with novel designs of neural network architectures that can represent memory/path-dependency of solid materials. They will be given a micro-mechanical unit-cell problem governed by visco-elasticity, and are expected to train their neural networks to find the homogenized macroscopic constitutive model, together with the hidden internal variables that captures the memory of deformation path at the macroscopic scale. 

Format: 1 individual report

 

Booklists

Please refer to the Booklist for Part IIB Courses for reference to this module, this can be found on the associated Moodle course.

Examination Guidelines

Please refer to Form & conduct of the examinations.

 
Last modified: 05/03/2024 11:43

Pages

Subscribe to CUED undergraduate teaching site RSS