Statistical and machine learning methods for classification of episodic memory

Abstract

Multiple modern methods of statistical feature extraction and machine learning
are applied to classification of encoding and retrieval of episodic memories us-
ing electroencephalogram (EEG) recordings. Raw data, different time-frequency
methods, and multiclass common spatial patterns are used for statistical feature ex-
traction. For each type of feature extraction multiple machine learning algorithms
are tested and compared. Classification accuracies of up to 82 % are reached with
one-dimensional convolutional neural networks on raw data. It is found that more
complex and time-consuming classifiers generally improve the accuracy. However,
the features chosen are the main factor deciding the accuracy. A novel idea for de-
signing an encoding-retrieval classifier is discussed and implemented. In spite of
having multiple different designs, almost all classifier combinations involving the
retrieval data fail to reach significant classification levels.

Popular Abstract

Can we read minds? In some ways, the answer is yes at least when encoding memories. Subjects are shown images of one of three types, and we are able to classify the type with high precision.
The human body is an immensely complex machine and one of the most complex parts isits brain. There are several the tories on how neural mechanisms
work within the field of cognitive neuroscience, and our goal is to advance the understanding of the mechanisms of memory using modern-day engineering methods.
A common way to measure activity in the brain is with the electroencephalogram
(EEG), where electrodes are placed over the scalp to measure voltage. It’s cheap, it’s fast, it’s well-researched, but it’s also extremely noisy, so getting useful information from EEG data is tricky. In our case, test subjects have been shown three different categories of images and we are tasked with devising an
algorithm that can tell the resulting EEG data apart, both upon viewing and recollection.
The idea is that some of the parts of the brain that activate during memorisation should also activate during recollection, in a process called ecphory.
When designing such an algorithm, there are two main parts to consider—feature extraction and classification. Feature extraction is about helping the
computer see things clearer: if the aim is to distinguish circles from squares in an image, a good feature may be “amount of corners”. The classification algorithm takes the features (e.g.four corners) and reduces them to a final guess (square).
We mainly tested three types of features: raw data, timefrequency analysis, and common spatial patterns. Raw data is just what it sounds like. Timefrequency (TF) analysis provides information on which frequencies occur at certain points in time.
This is interesting because previous research has shown that some types of neural mechanisms are encoded within certain frequency bands, such as for example theta waves oscillating at 7–10 Hz. The common spatial patterns (CSP) algorithm
instead uses information about which electrodes show similar results across images of the same type. The patterns that best separate the different types are
used as features. The best classification results came from using raw data (up
to 82 %). We believe that the raw data retains most of the information about when exactly most of the activity in the brain occurred, while TF methods (60–69 %) dilute that information slightly and CSP (48–51 %) disregards it entirely. As
is often the trend in machine learning research today, convolutional neural networks—while taking a long time to train—gave the highest results in terms of
accuracy. However, faster algorithms like support vector machines did not have much lower accuracy. Apart from being of interest to cognitive neuroscientists, any progress in understanding this problem can benefit for example brain-computer interface (BCI) researchers, since we are looking for general features of brain activity. Perhaps the three-class problem can be extended to even more types of images, and we can come even closer to saying that we can read minds.

• author: Basic Knezevic, Damir and Heimerson, Albin
• supervisor: Maria Sandsten
• organization: Mathematical Statistics
• course: FMS820 20181
• year: 2018
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMS-3346-2018
• ISSN: 1404-6342
• other publication id: 2018:E32
• language: English
• id: 8949940
• date added to LUP: 2018-06-15 10:54:06
• date last changed: 2024-09-24 14:32:00