17
Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young Scholar Student, Belmont High School Graduate research mentors: Matt Higger, Fernando Quiviria, PhD Candidate, Northeastern University Professor Deniz Erdogmus, Associate Professor, Northestern University College of Computer Engineering, Cognitive Systems Laboratory

Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Embed Size (px)

Citation preview

Page 1: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School

Eric Lehman, Young Scholar Student, Belmont High School

Graduate research mentors: Matt Higger, Fernando Quiviria, PhD Candidate, Northeastern University

Professor Deniz Erdogmus, Associate Professor, Northestern University College of Computer Engineering, Cognitive Systems Laboratory

Page 2: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

• Help a targeted group of individuals with severe speech and motor impairments who are unable to perform simple tasks or communicate with everyday individuals

Why use brain interfaces?

Image Source: http://i2.cdn.turner.com/cnn/dam/assets/121016060125-orig-ideas-brainwave-wheelchair-00013909-story-top.jpg

Page 3: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Brain Interface

Stimulus User EEG

Classifier Decision

Page 4: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

SSVEP Brain Interface Video

Page 5: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Definitions

• SSVEP: Stands for “Steady State Visually Evoked Potential”. This type of brain signal is a response to looking at repeated intensities of light from 0 to 60 Hz.• EEG: Stands for “electroencephalography”. EEG

data is the measurement of the brain’s electrical activity voltages on the surface of the scalp over a certain period of time. • Iris Dataset: A dataset that contains 3 different

types for flowers, 50 samples each, and 4 different features (sepal length in cm, sepal width in cm, petal length in cm, petal width in cm).  • Classifier: An algorithm that divides data into

different group based on their similarities.

Page 6: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Minimum Mean DistanceClassifier

•An algorithm that classifies multiple types of data.

•When given a test point, the program:

1. calculates the distance from the new data point to the average of training data points.

2. selects the training data point with the shortest distance

3. identifies the new data point in the same group as the closest training point.

6

Page 7: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Minimum Mean Distance Classifier

7

Classification of Iris Flower Dataset Using Minimum Mean Distance Classifier

Ground Truth Class

Estimated Class

I. setosa I. versicolor

I. virginica

I. setosa 1 0 0

I. versicolor 0 0.92 0.14

I. virginica 0 0.08 0.86

iPad
Page 8: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

8

Minimum Mean Distance Classifier

Classification of EEG Data Using Minimum Mean Distance Classifier

Ground Truth Class

Estimated Class

20 Hz 15 Hz 12 Hz

20 Hz 1 0 0

15 Hz 0 1 0

12 Hz 0 0 1

iPad
Page 9: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

k-Nearest Neighbor Classifier

•An algorithm that classifies and divides multiple types of data.

•When given a new test data point, the KNN classifier:

• 1. Calculates the distance from the test data to all training data points

• 2. Selects the k number of training data points that are the closest to the test data point

• 3. Identifies the test data point as the same as the most common class among the k nearest training data points

9

Page 10: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

k-Nearest Neighbor Classifier

10

Classification of Iris Flower Dataset Using Minimum Mean Distance Classifier

Ground Truth Class

Estimated Class

I. setosa I. versicolor

I. virginica

I. setosa 1 0 0

I. versicolor 0 0.94 0.04

I. virginica 0 0.06 0.96

Page 11: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

K-Nearest Neighbor Classifier

11

Classification of EEG Data Using K-Nearest Neighbor Classifier

Ground Truth Class

Estimated Class

20 Hz 15 Hz 12 Hz

20 Hz 1 0 0

15 Hz 0 1 0

12 Hz 0 0 1

Page 12: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

K Fold Cross Validation

•Separates the training set from the test set by segmenting the data into k number of sections

•The classifier will test on one section and train the remaining sections

•Prevents overfitting

12Image Source: http://classes.engr.oregonstate.edu/eecs/winter2011/cs434/notes/knn-4.pdf

Page 13: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

• RSVP Typing system• Uses P300 brain

signal to determine which letter is the acquired target

• SSVEP brain interface• Control robot motions

through looking at a screen

Applications

Image Source: http://www3.ece.neu.edu/~purwar/research/photo_gallery.htm, http://www3.ece.neu.edu/~orhan/

Page 14: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Applications

14

Can classify not just EEG data, but many other types of data!

Iris Flower Dataset

Image source: http://en.wikipedia.org/wiki/Iris_flower_data_set

Page 15: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Future Work

• Perform K-Fold Cross Validation

• Classify unprocessed EEG data using more advanced concepts to determine the most likely decision

• Classify EEG data obtained from other types of stimuli such as tactile sensors

• Help individuals with Locked-in Syndrome to communicate and with others through brain interfaces

Page 16: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Acknowledgements

• Graduate Research Mentors: Matt Higger, Fernando Quivira, PhD Candidates, Northeastern University

• Professor Deniz Erdogmus, Department of Electrical and Computer Engineering, Cognitive Systems Lab, Northeastern University

• Orkan Sezer, Summer intern, Northeastern University

• Center for STEM Education• Young Scholars Program & Team• Claire Duggan - Director• Kassi Stein, Jake Holstein, Chi Tse - YSP

Coordinators

Page 17: Minimum Mean Distance and k-Nearest Neighbor Classifiers for Signal Processing Kun Yi Li, Young Scholar Student, Quincy High School Eric Lehman, Young

Questions?