This Neural Network was trained to recognize distinguishing features rather than patterns
Image recognition is a part of computer vision that detects, identifies & processes an image in real life. Advancements in Machine Learning and the use of high-speed data services are fueling the growth of this technology. This is leading to the widespread adoption of image recognition across different industries. The biggest example of this is the use of CCTV surveillance cameras used for facial recognition.
Despite having come a long way, these neural networks still need to be trained on extensive libraries of images & patterns to be able to distinguish between different sets of objects. While we humans identify a new image by looking at it as a collection of recognizable features, a neural network simply looks for pixel patterns across the entire image.
In this particular example, we might identify a species of birds by features like the contour of its beak, the colors of its plume, and the shape of its feet. But if the same bird was being looked at by a neural network it would scan the pixel patterns across the bird and its background without differentiating between objects. This makes neural networks vulnerable to making mistakes in correctly identifying objects and perhaps the reason for their biggest criticism.
Researchers from Duke University and MIT Lincoln Laboratory have now trained a neural network called the prototypical part network (ProtoPNet) — which has the ability to recognize distinguishing features across bird species. Of course, this meant training of the neural network by showing it many similar images of each species and having it identify features of the images that looked similar within species yet different across them.
As an example, this training enabled ProtoPNet to learn that a cardinal’s distinguishing feature is its black mask against the red feathers. Once the algorithm was trained on sufficient data, it was presented with a new image of a bird. The network then searched for recognizable features using collective evidence from before to make a prediction about the species of the bird.
The testing of this trained image recognition algorithm demonstrated that the added feature of interpretability didn’t affect its accuracy. On the two tasks of bird and car model identification, researchers found that the network neared and in some cases exceeded the results achieved by some of the existing state-of-the-art non-interpretable algorithms.
Apart from the accuracy, the network’s ability to explain how they arrived at a certain conclusion is of paramount importance in high stakes environments like hospitals where, for example, these algorithms might help doctors in identifying a tumor. This would not only help humans trust these algorithms but be able to easily identify when their deductions are wrong.
Medium | Twitter | LinkedIn | StockTwits | Telegram
