Download presentation

Presentation is loading. Please wait.

Published byEaston Crawford Modified over 3 years ago

1
Test: CNN vs. AMM Data: Four sets of Jail Break data from ARL/Penn State Total Negative 88 Total Positive 69 Total 157 Two sets of five tests on all four data sets July 15, 2013 James P. LaRue Jadco Signals

2
y=double(y); yy=y(:,:,1); yy1=yy(1:3:end,:); yy2=yy(2:3:end,:); yy3=yy(3:3:end,:); Y1=yy1+yy2+yy3; yy1=Y1(:,1:2:end); yy2=Y1(:,1:2:end); Y2=yy1+yy2; KeepN(:,:,k)=Y2(5:32,3:end); Convert 120x60 JPEGS into 28x28

3
Examples of: NOT PRESENT Examples of: PRESENT

4
Accuracy Train with random pick 25/157 Epoch-One training cycle of 25 1.00 0.82 0.07 0.94 1.00 0.78 0.18 0.97 0 0.79 1.00 0.95 0.72 0.73 0.98 0.97 1.00 0.80 0.86 0.95 1.00 0.93 0 0.72 Not Present Present CNN AMM 1 2 3 4 5 10 Number of Epochs Mean = 0.78 0.81 0.51 0.92 Variance = 0.16 0.01 0.23 0.01

5
Time to execute 157 decisions After training CNN AMM 0.3423 0.020 0.3407 0.0171 0.3435 0.0176 0.3427 0.0176 0.3400 0.0176 0.3296 0.0167 meantime 0.3398 0.0179 1.05 3.05 4.10 2.10 2.75 4.85 3.15 2.88 6.03 4.20 2.75 6.95 5.24 2.86 8.10 10.50 2.89 13.39 1 2 3 4 5 10 Number of Epochs CNN AMM AMM+CNN Time in Minutes Train with random pick 25/157 Epoch-One training cycle of 25 Training Time

6
Accuracy Train with random pick 20/157 Epoch-One training cycle of 20 1.00 0..79 0.15 0.94 0.96 0.79 0.37 0.97 1.0 0.79 0.14 0.95 0.97 0.79 0.75 0.97 0.93 0.71 1.00 0.97 1.00 0.90 0 0.76 Not Present Present CNN AMM 1 2 3 4 5 10 Number of Epochs Mean = 0.97 0.80 0.40 0.92 Variance = 0.00 0.00 0.15 0.00

7
1The CNN has to train first. Then AMM can train. Thus, Total train time for AMM Includes CNN train time. 2 Compare total time to accuracy of CNN vs AMM. Example: from accuracy slide Across all Epochs: AMM scored 86%, CNN scored 65% 3 AMM is always consistent across Epochs. CNN is not consistent. Similar results can be shown for other tests. Note several CNN ‘all or nothing’ classifications. 4.AMM execution time recorded with single shot matrix. CNN always needs to run through all its steps, whereas multiple AMM matrices can be compressed into one matrix. 5.Note: For MNIST data set, CNN required 60 minutes to attain 80% accuracy, while AMM required 6 CNN minutes to reach 78%. 6.Conclusion: If time is not a factor, then use logic based CNN. However, if time is a factor, then use intuitive based AMM, knowing the decision will be robust. Discussion

Similar presentations

OK

Actigraphy Assessment of Mother’s Sleep : 6, 12 & 18 weeks postpartum Introduction: Janelle MacKenzie, Kerry Armstrong & Simon Smith Method: Results: Conclusion:

Actigraphy Assessment of Mother’s Sleep : 6, 12 & 18 weeks postpartum Introduction: Janelle MacKenzie, Kerry Armstrong & Simon Smith Method: Results: Conclusion:

© 2018 SlidePlayer.com Inc.

All rights reserved.

To make this website work, we log user data and share it with processors. To use this website, you must agree to our Privacy Policy, including cookie policy.

Ads by Google