Google I/O 2019 was more interesting for me to watch. There are eye catching innovations displayed in the conference. Among them TCAV (stands for Testing with Concept Activation Vectors ) grabbed my attention. Because Machine Learning engineers often keep believing on the algorithm itself,and are not concerned about the internal functioning of a neural network. In some cases bias play a major role in manipulating the results. In some other cases its hard to understand what weights a neural network has taken to display its conclusions. So it is helpful for you when your Neural Network shows the things that are responsible for drawing the conclusions.
Been Kim, a research scientist at Google Brain created this TCAV. Lets see how it works with a neural network which identifies a Zebra. TCAV performs a sensitivity testing on the neural network so that it can identify the factors that have influenced a neural network to draw its conclusion. A person can understand the importance of these factors by the TCAV test score. from the below figure we could see that how stripes for Zebra played a major role in identifying it.
This would bring positiveness to the usage of Machine Learning in the fields like Medicine where it is important to study the internal factors that a neural network uses to predict a disease.
Well its your turn to have TCAV in your projects. Bye !
Video by Quanta Magazine :
Been Kim research paper
1 Comment
Good