1 of 11

Strike (with) a Pose:

Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects

Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen

2 of 11

Motivation

  • Artificial neural networks are increasingly common components of computer vision systems
  • While artificial neural networks are extremely powerful and useful, we still don’t know how these models “think”

2/11

(Bansal et al., 2018)

(Taigman et al., 2014)

(De Fauw et al., 2018)

3 of 11

Questions you will be able to answer after this talk

  • How do artificial neural networks work?
  • How are artificial neural networks trained for computer vision tasks?
  • Do artificial neural networks understand what they are looking at?

1.

2.

3.

3/11

4 of 11

How do artificial neural networks work?

(Honkela, ‎2001)

(Nielsen, 2015)

(Ng/Coursera)

input = observed “features”/things you measured

output = “label”, i.e., the thing you’re trying to predict

nodes/neurons = computational unit

weights/connections (wis) = what the neural network is trying to learn

Training Algorithm

  • Collect a dataset of many (input, output) pairs
  • Randomly initialize neural network weights
  • Until satisfied
    • Generate a prediction for each input
    • Compare predictions to outputs to get errors
    • Sum up all of the errors
    • Get partial derivatives (just calculus!)
    • Move weights a little “downhill”

In practice, use differentiable activation functions like logistic or rectifier

4/11

5 of 11

How are artificial neural networks

trained for computer vision tasks?

(Stanford CS231n)

Step #1: Acquire labeled dataset

(Krizhevsky, Sutskever, and Hinton, 2012)

Step #2: Train a Convolutional Neural Network

(Stanford UFLDL)

5/11

6 of 11

Do artificial neural networks understand

what they are looking at?

3D School Bus

3D School Bus

Real School Bus

6/11

7 of 11

Do artificial neural networks understand

what they are looking at?

#1: Collected 30 3D objects found in a traffic environment

7/11

8 of 11

Do artificial neural networks understand

what they are looking at?

​

#2: Randomly sampled object poses and recorded neural network predictions

Median Misclassification Rate = 97%!

8/11

9 of 11

Do artificial neural networks understand

what they are looking at?

#3: Altered individual parameters of correctly recognized poses

9/11

10 of 11

Do artificial neural networks understand

what they are looking at?

#4: Used black box optimization to discover specific failures

10/11

11 of 11

Do artificial neural networks understand

what they are looking at?

​

Links

  • No!
  • The standard approach to training artificial neural networks for computer vision tasks seems to produce incredibly naive models
  • We’re currently experimenting with approaches that may enable 3D reasoning in artificial neural networks

11/11