Overview

Within this course you have learnt about classification and optimisation techniques that are inspired by the intelligence seen within nature; you will now use these techniques to further our understanding of natural intelligence. The analysis and classification of large datasets is one of the primary applications of AI/CI methods. For example, deep convolutional neural networks have extensive applications in computer vision and image processing, where they can classify features and objects within many thousand different images. This coursework will test your knowledge, and ability to apply this knowledge, to a classification task that is inspired by current biomedical engineering research. You will create a system to automatically analyse a set of recordings that have been made from the human brain – one of the most complicated structures known to exist. This document details the datasets that you will be working with, the submission details, and the marking criteria.

Stage Gating

This coursework uses a stage gating approach to help you build towards the final submission. There are 4 tasks in total, each of which will allow you to make a submission and receive automated feedback. This gives you the opportunity for multiple points of feedback (on work submitted so far) and feedforward (areas to improve in future submissions) throughout the course.

Recordings

You will be working with datasets that contain recordings made using a simple bipolar electrode inserted within the cortical region of the brain; this is a typical experimental setup that is frequently employed by neuroscientists. The recordings contain several spikes (extracellular action potentials) that are from five different types of neuron (Class 1, 2, 3, 4 & 5). Each neuron produces spikes that have a subtly different morphology, and each neuron can only produce one spike at a time. One of the challenges with this type of recording is that different neurons often fire simultaneously, and some of the spikes will be partially overlapping.

The goal is to process the recordings and automatically find when each spike occurs, and which neuron produced it (often called spike sorting in the literature). This is akin to the MNIST classification problem, except that you also  need to  detect when  in time  each  spike  has  occurred  in  order to  extract them  for  classification.  This information will  enable the  selective  recording  from  individual  neurons, a  critically unmet  need in modern neuroscience. The recordings are time records and Figure 1 illustrates an example of two spikes, both are from the same type of neuron and have approximately the same morphology. The sample rate for all of the recordings is 25 kHz.

3.5

3

2.5

2

1.5

1

0.5

0

-0.5 

57.535

Figure 1 An example of two spikes, both are from the same class of neuron and have approximately the same shape and amplitude.

Dataset Overview

There are four datasets that you will use for this coursework, they are all in the same basic format and are available on Moodle. Each dataset is a MATLAB data file that contains some or all of the following three vectors:

For example, the first element in the Index vector states where (in samples) the first spike occurs in the d vector; the corresponding first element in the Class vector states the class of this spike. The Index and Class vectors can be used to train your algorithms and to assess their performance on unseen data. You should firstly consider how you can detect the spikes, then consider classification as the next step. You should use the Python tools developed in the laboratory sessions to solve this challenge, and your final solution must be written in Python.

Note: You can import a .matfile into Python using the following code snippet:

Datasets

The following datasets are available on Moodle. It is up to you how you wish to split this data into training, testing, and validation datasets.

Tasks

The following tasks form the coursework, and you should attempt them in order. You are free to use whatever CI techniques you wish, alongside any standard signal processing or mathematical tools. You may use CI techniques that have not been taught in the course. Each task will require you submit a .mat file to Moodle in the format specified.

1.    Load dataset D2 into Python and devise a method for finding the individual spikes in the d vector and thus generate the Index vector.

Submission: D2.mat file containing the vectors D and Index.

Feedback: Precision and recall of the spike locations.

Weighting: 10%

2.    Load dataset D2 into Python and detect and classify each spike using any CI technique of your choosing. You may wish to use dataset D1 to help you with training your classifier.

Submission: D2.mat file containing the vectors D, Index and Class.

Feedback: A confusion matrix outlining the performance .

Weighting: 20%