1 of 5

AI For Behavioral Discovery

Team Members: Adarsh Narayanan, Ben Yu

Advisors: Dr. Martin, Dr. Howard

2 of 5

CNN’s and AlexNet

  • Went over CNN basics and AlexNet Architecture
    • 5 Conv2d and Maxpool2d layers with 3 Fully Connected layers.
  • When sending multiple frames to the model, adds as channel dimension.
    • Multiple kernel layers filter through the multi-frames, then summate for the feature layer.

​

3 of 5

Learning Curve Accuracies

1 frame

4 frame

After increasing the distortion on the simulation dataset …

~83% on Evaluation Set

~70% on Evaluation Set

~73% on Evaluation Set

~82.5% on Evaluation Set

4 of 5

Conclusions

  • Overfitting (Expected)
  • Class bias
  • Slight Class imbalance.
  • Model trains better on 1st fold

1st Fold

2nd Fold

5 of 5

Plans for Next Week

  • Plot Saliency/Grad-Cam
    • Understand focused features
  • Continue training/updating simulation data.
  • Benchmark/Verify if AlexNet model is learning motion
    • Temporal understanding.