Limitations of Deep Learning

Now let’s talk about some limitations of deep learning.

There are several limitations in deep learning models. First of all, the models are not scale and rotation invariants, and can easily misclassify images when the object poses are unusual.

Here are some examples from the CVPR 2019 paper “Strike with a Pose: Neural networks are easily fooled by strange poses of familiar objects.” Let’s take a look of those images. At the first row, the first image is correctly classified as school bus. However, if we rotate and show only the bottom of the bus, it will be misclassified as garbage truck, a punching bag, or a snowplow. Similarly, a motor scooter may be misclassified as parachute or bobsled with strange poses; a fire truck may be classified as school bus or fireboat. Although many methods have been proposed to solve those issues, the errors show that the models lack knowledge of our real world.

To add insult to injury, the models can be fooled and cheated intentionally using Generative Adversarial Networks. These techniques are called adversarial attacks. Here is an example from Ian Goodfellow’s paper. By adding some small intentional information, which is not detectable by humans, we can make CNN models misclassify panda into gibbon with high confidence! This phenomenon is very robust, Even if the photos of the adversarial examples can still fool the models. Adversarial attack raises a serious security issue of deep-learning based image recognition models. For example, a hacker can change the direction of a traffic sign to fool autonomous vehicles without being detected by the police.