Welcome to our blog!
We are Group 7 of ENGR 103 at Drexel University. Our project is a Biomedical Engineering research-based project. It is inspired by brain scan images from Dr. Hualou Liang's lab.
Our group hopes to:
- Analyze the signals from the brain of a subject
- Develop an algorithm to differentiate between sets of certain stimuli
- Interpret MEG data with MATLAB
We are Group 7 of ENGR 103 at Drexel University. Our project is a Biomedical Engineering research-based project. It is inspired by brain scan images from Dr. Hualou Liang's lab.
Our group hopes to:
- Analyze the signals from the brain of a subject
- Develop an algorithm to differentiate between sets of certain stimuli
- Interpret MEG data with MATLAB
Abstract:
The objective of this lab was to develop a system that can differentiate between a few predetermined sets of visual stimuli to a certain extent of accuracy. The basic idea behind this project was to analyze the signals from the brain of a subject when the subject views a particular object and use them to identify patterns in the data. These patterns ‘train the decoder’ to recognize similar patterns in test data. A decoder was developed using certain algorithms based on regularized multinomial logistic regression [1]. MEG labelled and unlabeled data was used from a research paper provided in class [1]. The labelled data was extracted, and plotted for visual representation and understanding. The data was then analyzed using different features like mean, standard deviation etc. Based on the findings, an algorithm was developed that utilizes some of these features to identify the similar patterns and features in the unlabeled test data. The algorithm was able to predict which set of the visual stimuli was being viewed by the subject based only on the MEG signals recorded.
The objective of this lab was to develop a system that can differentiate between a few predetermined sets of visual stimuli to a certain extent of accuracy. The basic idea behind this project was to analyze the signals from the brain of a subject when the subject views a particular object and use them to identify patterns in the data. These patterns ‘train the decoder’ to recognize similar patterns in test data. A decoder was developed using certain algorithms based on regularized multinomial logistic regression [1]. MEG labelled and unlabeled data was used from a research paper provided in class [1]. The labelled data was extracted, and plotted for visual representation and understanding. The data was then analyzed using different features like mean, standard deviation etc. Based on the findings, an algorithm was developed that utilizes some of these features to identify the similar patterns and features in the unlabeled test data. The algorithm was able to predict which set of the visual stimuli was being viewed by the subject based only on the MEG signals recorded.
Introduction:
Humans have always been curious about the brain and its processes. One of the more popular functions that have been researched in the past is the ability to ‘read minds’ or predict the stimuli based on the changes observed in the brain activity. MEG is one method to measure such brain activity.
Magnetoencephalography (MEG) is a noninvasive technique for investigating neuronal activity in the living human brain. The time resolution of the method is better than 1 ms and the spatial discrimination is, under favorable circumstances, 2-3 mm for sources in the cerebral cortex. In MEG studies, the weak 10 fT-1 pT magnetic fields produced by electric currents flowing in neurons are measured with multichannel SQUID (superconducting quantum interference device) gradiometers. The sites in the cerebral cortex that are activated by a stimulus can be found from the detected magnetic field distribution, provided that appropriate assumptions about the source render the solution of the inverse problem unique. [1]
Overview:
In this lab, the object was to develop a system that can differentiate between a few pre-determined sets of visual stimuli to a certain extent of accuracy. The basic idea behind this project was to analyze the signals from the brain of a subject when they viewed a particular object and use the recorded data to identify patterns in brain activity. These patterns ‘train the decoder’ to recognize similar patterns in test data. A decoder was developed using certain algorithms based on regularized multinomial logistic regression [2]. Using MEG labelled and unlabeled data from the research paper provided in class, the decoder was trained [2]. The data used was only for the specific sets of visual stimuli that have been analyzed by the authors of the paper. However, one cannot obtain new MEG signals and data independently for this project. Hence, the ‘mind reader’ was limited in capacity to specific visual stimuli in that sense. Since there was a large amount of data that needs to be analyzed statistically in multiple ways, MATLAB was used for this project. The first step was to extract the labelled data, and plot it for visual representation and understanding. The next step was to analyze the data using different features like mean, standard deviation etc. Based on the findings, an algorithm was developed that utilizes some of those features to identify the similar patterns and attributes in the unlabeled test data. The algorithm is able to predict which set the visual stimuli was viewed by the subject just based on the MEG signals and brain activity.
Existing Solutions:
During the recent years, supervised classification has become increasingly important methodology in analyzing functional neuroimaging data within functional magnetic resonance imaging (fMRI) as well as in electroencephalography (EEG) and magnetoencephalography (MEG) with earliest papers tracing back to the early 1990s. Supervised learning is the machine learning task of inferring a function from labeled training data(Mohri).The training data consist of a set of training examples. A supervised learning algorithm analyzes the training data and produces an inferred function, which can be used for mapping new examples. An optimal scenario will allow for the algorithm to correctly determine the class labels for unseen instances.Currently data from neuroimaging data is sorted using such algorithms.
Project Objectives:
The objective of this project was to determine the manner in which the brain activity can most accurately be decoded and predict the image that the subject was shown. To accomplish this task, the signals were taken and compared to each other using selected features (standard deviation, variance, mean). Once the datum was accurately described, and algorithm was utilized to find trends in images of the same class. This algorithm, once trained, returns a table showing the predicted image class against the actual image class, and a percentage of accuracy in its predictions.
Technical Activities:
The major tasks of this project are listed below chronologically. Developing the algorithm was the expected major technical challenge since it required knowledge related to extracting and analyzing large amounts of data using MATLAB. The project was divided into individual tasks with deadlines that were completed over the ten week period. This including sorting, filtering and analyzing the data using various methods. Other tasks such as proper documentation and citation have been distributed amongst the group members. We also maintain a group blog post to update the class on our progress.
Understanding the research paper:
This project was completed using the information found in the research paper ‘Mind reading with regularized multinomial logistic regression’ which was provided to the class as reference [1].
Referenced the data provided and used the methods of data filtration mentioned.
Familiarize the team with MATLAB:
Since this project required some basic knowledge about coding in MATLAB, a couple of weeks were dedicated to familiarizing members with MATLAB coding.
Extract and access the data:
Downloading the data into a readable format was the first step of the project. The code had to be modified to extract and isolate the required parts of the data. (For example: Activity for each node for each entry).
Analyze the data and evaluate the features:
A visual representation of the data was created initially for the understanding of the signals. The next step was to calculate the value of different features for the data in MATLAB.
Develop the algorithm:
After studying the value for different features for different data sets, the next step was to decide on specific features to include in the algorithm. These features helped to distinguish data from one set of visual stimuli from the other set of visual stimuli.
Test the algorithm:
‘Training’ the decoder:The first step was to pass labelled data through the decoder, containing the modified algorithm, to determine the patterns in various features for different sets of visual stimuli (MEG data).
Prediction/ 'Mind reading': Next, the unlabeled data was passed through the decoder and the decoder was able to predict the set of visual stimuli the data belongs to, based on the patterns encountered in the labelled data.
Revise/Alter the algorithm:
The last step was to revise the algorithm and alter it to get a high percentage of successful predictions.
Project Timeline:
Below is a timeline for the project, keeping in mind the effort required for each part and the time constrain of completing it in ten weeks. It is illustrated in Table 1 below.
Facilities and Resources:
No resources or facilities are needed in this project besides MATLAB software and provided data.
Expertise:
This project required the team to have some basic coding skills. The team had to be familiar with few basic functions used in MATLAB.
Results:
MATLAB was used to determine that unfiltered data was most accurate. It was also found that the matter in how to group by classes is vital - grouping data helped substantially, because it gave insight on how the classes (stimuli) were sorted. Also, untreated data worked better than regular data giving better results- using untreated data removed any bias or trends that may have skew the data.
Future Work:
An interesting topic for future work is possible filtering operations of the data. The current algorithm does not apply any preprocessing steps besides the usual calibration and sample rate related operations. The performance could be improved with denoising either in the time domain (within-channel smoothing) or in the spatial domain (between-channel smoothing).[2]
Overview:
In this lab, the object was to develop a system that can differentiate between a few pre-determined sets of visual stimuli to a certain extent of accuracy. The basic idea behind this project was to analyze the signals from the brain of a subject when they viewed a particular object and use the recorded data to identify patterns in brain activity. These patterns ‘train the decoder’ to recognize similar patterns in test data. A decoder was developed using certain algorithms based on regularized multinomial logistic regression [2]. Using MEG labelled and unlabeled data from the research paper provided in class, the decoder was trained [2]. The data used was only for the specific sets of visual stimuli that have been analyzed by the authors of the paper. However, one cannot obtain new MEG signals and data independently for this project. Hence, the ‘mind reader’ was limited in capacity to specific visual stimuli in that sense. Since there was a large amount of data that needs to be analyzed statistically in multiple ways, MATLAB was used for this project. The first step was to extract the labelled data, and plot it for visual representation and understanding. The next step was to analyze the data using different features like mean, standard deviation etc. Based on the findings, an algorithm was developed that utilizes some of those features to identify the similar patterns and attributes in the unlabeled test data. The algorithm is able to predict which set the visual stimuli was viewed by the subject just based on the MEG signals and brain activity.
Existing Solutions:
During the recent years, supervised classification has become increasingly important methodology in analyzing functional neuroimaging data within functional magnetic resonance imaging (fMRI) as well as in electroencephalography (EEG) and magnetoencephalography (MEG) with earliest papers tracing back to the early 1990s. Supervised learning is the machine learning task of inferring a function from labeled training data(Mohri).The training data consist of a set of training examples. A supervised learning algorithm analyzes the training data and produces an inferred function, which can be used for mapping new examples. An optimal scenario will allow for the algorithm to correctly determine the class labels for unseen instances.Currently data from neuroimaging data is sorted using such algorithms.
Project Objectives:
The objective of this project was to determine the manner in which the brain activity can most accurately be decoded and predict the image that the subject was shown. To accomplish this task, the signals were taken and compared to each other using selected features (standard deviation, variance, mean). Once the datum was accurately described, and algorithm was utilized to find trends in images of the same class. This algorithm, once trained, returns a table showing the predicted image class against the actual image class, and a percentage of accuracy in its predictions.
Technical Activities:
The major tasks of this project are listed below chronologically. Developing the algorithm was the expected major technical challenge since it required knowledge related to extracting and analyzing large amounts of data using MATLAB. The project was divided into individual tasks with deadlines that were completed over the ten week period. This including sorting, filtering and analyzing the data using various methods. Other tasks such as proper documentation and citation have been distributed amongst the group members. We also maintain a group blog post to update the class on our progress.
Understanding the research paper:
This project was completed using the information found in the research paper ‘Mind reading with regularized multinomial logistic regression’ which was provided to the class as reference [1].
Referenced the data provided and used the methods of data filtration mentioned.
Familiarize the team with MATLAB:
Since this project required some basic knowledge about coding in MATLAB, a couple of weeks were dedicated to familiarizing members with MATLAB coding.
Extract and access the data:
Downloading the data into a readable format was the first step of the project. The code had to be modified to extract and isolate the required parts of the data. (For example: Activity for each node for each entry).
Analyze the data and evaluate the features:
A visual representation of the data was created initially for the understanding of the signals. The next step was to calculate the value of different features for the data in MATLAB.
Develop the algorithm:
After studying the value for different features for different data sets, the next step was to decide on specific features to include in the algorithm. These features helped to distinguish data from one set of visual stimuli from the other set of visual stimuli.
Test the algorithm:
‘Training’ the decoder:The first step was to pass labelled data through the decoder, containing the modified algorithm, to determine the patterns in various features for different sets of visual stimuli (MEG data).
Prediction/ 'Mind reading': Next, the unlabeled data was passed through the decoder and the decoder was able to predict the set of visual stimuli the data belongs to, based on the patterns encountered in the labelled data.
Revise/Alter the algorithm:
The last step was to revise the algorithm and alter it to get a high percentage of successful predictions.
Project Timeline:
Below is a timeline for the project, keeping in mind the effort required for each part and the time constrain of completing it in ten weeks. It is illustrated in Table 1 below.
Facilities and Resources:
No resources or facilities are needed in this project besides MATLAB software and provided data.
Expertise:
This project required the team to have some basic coding skills. The team had to be familiar with few basic functions used in MATLAB.
Results:
MATLAB was used to determine that unfiltered data was most accurate. It was also found that the matter in how to group by classes is vital - grouping data helped substantially, because it gave insight on how the classes (stimuli) were sorted. Also, untreated data worked better than regular data giving better results- using untreated data removed any bias or trends that may have skew the data.
Future Work:
An interesting topic for future work is possible filtering operations of the data. The current algorithm does not apply any preprocessing steps besides the usual calibration and sample rate related operations. The performance could be improved with denoising either in the time domain (within-channel smoothing) or in the spatial domain (between-channel smoothing).[2]
References:
[1] Hämäläinen, M., et al. (1993). "Magnetoencephalography theory, instrumentation, and applications to noninvasive studies of the working human brain." Reviews of Modern Physics 65(2): 413-497.
[2] H. Huttunen, T. Manninen, J-P. Kauppi and J. Tohka, "Mind Reading with Regularized Multinomial Logistic Regression," Machine Vision and Applications, pp. 1-15, November 2012. [pdf] [software] [data]
No comments:
Post a Comment