AI4All Day 7: Complex Neural Networks, Soft Robots, Transporting Atoms
Nidhi Parthasarathy, Wednesday, July 6th 2022
Learning about CNNs
Day 7 began with Jean Benoit Delbrouck teaching us about neural networks and deep learning. He discussed traditional pattern recognition and fixed hand-crafted feature extraction. He went over an example with cars, and the difference between traditional pattern recognition and deep learning. He discussed how a neural network is a branch of perceptrons, the smallest unit of neural networks with different weights added to inputs to produce outputs. We looked at neural networks at playground.tensorflow.org.
Jean also provided some fascinating historical context on the development of this area — about Rosenblatt’s initial work on perceptrons, followed by Marvin Minsky’s book on what perceptrons could not do, and the subsequent reduction in research in this area (“AI winter”) until recently when it picked up again. Finally, he talked about convolutional neural networks (CNNs), which are computing systems used to address visual imagery problems and consist of shared weight architecture of convolution kernels that shift over the input features.
He also discussed convolutional filters and how they help with reduced processing, and how they use pooling when going over the entire image with kernels for improved efficiency.
After the break, Jean went over how to create a model in real time and walked us through the code. Some of the concepts were getting a bit hard to understand, but the research mentors assured us that we were learning about more advanced concepts and we weren’t expected to understand all of it in our first introduction to it.
Learning about Soft Robots
In the afternoon, Allison Okamura talked to us about soft robots for humanity. She started the lecture by introducing herself as a professor in Stanford and gave some background information about where she studied and worked at. She works on collaborative haptics and robots in medicine (CHARM lab) and is in the department of mechanical engineering at Stanford.
She explained that soft robotics was a field of robotics where robots were soft by material or structure (flexible and stretchable materials). The benefits of these robots is their adaptability, safety, and cost efficiency. These robots tend to be inspired by biology, for example, pollen tubes, nerve cells, and even vines. She also talked about concentric tube robots which are a series of hollow, precurved tubes that fit concentrically.
She also talked about 3D printing and how it helps with soft robotics. Apart from these, she talked about the many applications of soft robots in archeology, search and rescue, inspection, and agriculture. She also talked about the capabilities in compliance and manipulation. Finally she talked about modularity and haptic illusions.
Overall, this lecture was very interesting and it made me very curious. Where else can we use soft robotics?
“The Future of Atoms Transportation”
Next, we had a presentation by Pedro Gonzalez from the Kiwibot company (“the future of atoms transportation”). The kiwibot is a small robot that is used for delivery.
Pedro began by explaining the “last mile delivery” problem. He explained that last mile delivery is expensive and unsustainable. For example, a 2 pound burrito is transported in a 2 ton car. It is also not environmentally sustainable as urban last mile delivery can add 25 million tons of CO2 emitted annually and increase traffic congestion by over 21%.
Kiwibot is their solution to this problem. The kiwibot is a human friendly, autonomous robot that saves up to 66% in delivery costs. Furthermore, the company has active participation in policy making, has a very nice API to connect with customers and partners through their favorite apps, and is very good at city mapping, data collection, and advertising.
Pedro then talked about how Kiwibot came to life. He explained that they are a Colombian company founded in 2017 and accelerated by the UC Berkeley SkyDeck program. He explained how driving robots remotely isn’t safe, and robotic delivery scales better with autonomy and how AI is used to achieve this autonomy. Overall, the presentation was really interesting and really showed us a glimpse of the future of autonomous delivery.
Virtual Social
We ended the day with our usual social. Lots of fun talking and playing games with our new friends.
Read on for day 8.
