Ramasubramanian Balasubramanian

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2019-37

May 14, 2019

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-37.pdf

We first explore the behavior of self-assembling wires, ball bearings that form emergent structures, under an electric field. We model this behavior using the laws of physics as well as some heuristics, with the aim to implement this self-organizing behavior on to a group of drones covering a given region optimally, communicating with each other in a peer-to-peer manner, and collating multimedia information demonstrating collaborative intelligence. After looking at how these ball bearings turn 'intelligent' even with simple adaptation rules, we explore frameworks with more theoretically grounded rules: can neural networks collaborate to improve performance? We realize that we first need a way to understand what goes on behind-the-scenes when a neural network makes a particular decision to discuss its behavior, and hence we turn to heatmapping. We use it to figure out how much does each pixel contribute to the final output in an image classification task, or in other words what parts of an image led a network to classify an image into a particular class. We demonstrate that the network looks for characteristic features belonging to each class while trying to classify and how this could lead to errors. We then ask the important question of whether we can use heatmaps obtained from one neural network to filter out background noise from the images by blackening out the pixels that don’t contribute much to the final output, and if so, can these filtered images, when passed through another neural network, lead to a higher classification accuracy. We see that the data processing inequality holds and the accuracy does not increase, but given that the images are simpler now, they need fewer bits to be encoded and hence networks with simpler architectures, having a smaller capacity, are enough for the task without much drop in the accuracy.

Advisors: Kannan Ramchandran and Gerald Friedland

---

BibTeX citation:

@mastersthesis{Balasubramanian:EECS-2019-37,
    Author= {Balasubramanian, Ramasubramanian},
    Title= {On Collaborative Intelligence},
    School= {EECS Department, University of California, Berkeley},
    Year= {2019},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-37.html},
    Number= {UCB/EECS-2019-37},
    Abstract= {We first explore the behavior of self-assembling wires, ball bearings that form emergent structures, under an electric field. We model this behavior using the laws of physics as well as some heuristics, with the aim to implement this self-organizing behavior on to a group of drones covering a given region optimally, communicating with each other in a peer-to-peer manner, and collating multimedia information demonstrating collaborative intelligence. After looking at how these ball bearings turn 'intelligent' even with simple adaptation rules, we explore frameworks with more theoretically grounded rules: can neural networks collaborate to improve performance? We realize that we first need a way to understand what goes on behind-the-scenes when a neural network makes a particular decision to discuss its behavior, and hence we turn to heatmapping. We use it to figure out how much does each pixel contribute to the final output in an image classification task, or in other words what parts of an image led a network to classify an image into a particular class. We demonstrate that the network looks for characteristic features belonging to each class while trying to classify and how this could lead to errors. We then ask the important question of whether we can use heatmaps obtained from one neural network to filter out background noise from the images by blackening out the pixels that don’t contribute much to the final output, and if so, can these filtered images, when passed through another neural network, lead to a higher classification accuracy. We see that the data processing inequality holds and the accuracy does not increase, but given that the images are simpler now, they need fewer bits to be encoded and hence networks with simpler architectures, having a smaller capacity, are enough for the task without much drop in the accuracy.},
}

EndNote citation:

%0 Thesis
%A Balasubramanian, Ramasubramanian 
%T On Collaborative Intelligence
%I EECS Department, University of California, Berkeley
%D 2019
%8 May 14
%@ UCB/EECS-2019-37
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-37.html
%F Balasubramanian:EECS-2019-37