Phillip Kuznetsov

I am currently a Research Assistant in Alyosha Efros's Lab (part of BAIR) and a Machine Learning Consultant through my company Alinea.AI. I completed my undergraduate in EECS at UC Berkeley in Spring of 2018. During that time, I founded Machine Learning at Berkeley - a student organization of 70 members that conducted projects with industry partners, research with labs on campus, and hosted numerous events on campus. Through ML@B, I also taught a course and several workshops on misc. ML topics.

I recently wrote a chapter on Adversarial Machine Learning in the soon-to-be-released "Artificial Intelligence Safety and Security".

Email  /  GitHub  /  LinkedIn Resume

Artificial Intelligence Safety and Security
Chapter on Adversarial Machine Learning
Phillip Kuznetsov, Riley Edmunds, Ted Xiao, Humza Iqbal, Raul Puri, Noah Golmant, Shannon Shih
CRC Press 2018
Book / Chapter

Chapter surveying the field of Adversarial Machine Learning with connections to AI Safety.

Transferability of Adversarial Attacks in Model-Agnostic Meta-Learning
Riley Edmunds, Noah Golmant, Vinay Ramasesh, Phillip Kuznetsov, Piyush Patil, Raul Puri
Deep Learning and Security Workshop (DLSW) in Singapore, 2017

Work demonstrating that Universal Adversarial Attacks transfer between child models of Meta-Learned models than from models that are initialized randomly.

Using TensorFlow to generate images with PixelRNNs
Phillip Kuznetsov, Noah Golmant
O'Reilly Blog

Work demonstrating that Universal Adversarial Attacks transfer between child models of Meta-Learned models than from models that are initialized randomly.

Machine Learning at Berkeley

Founded UC Berkeley's first Machine Learning student organization. President in Spring Semester of 2018.

Alexa Prize
Proceedings  /  GitHub

One of the sponsored teams entered in the inaugural Alexa Prize competition. Given a $100k research grant and unlimited to AWS to build a conversational chatbot.

Extrapolating Texture from Texture Cues
Phillip Kuznetsov, Stefan Palombo, Gabriel Gardner, Rahil Mathur, Michael Luo
Project Website

We attempt to apply textures automatically to non-textured images of 3d renderings using convolutional neural networks as a final step in a graphics pipeline. Our final product is a system in which you can pass an untextured rendering with a texture cue into a trained convolutional neural network that then outputs a fully textured result.

Creative Adversarial Networks

First Open-source Implementation of Creative Adversarial Networks. Adapts the Generative Adversarial Network objective function to try to maximize the entropy for a style-class distribution, while also minimizing the original adversarial objective.

Image Compression by Abstracting out Details

We propose a method to apply a pre-trained Generative Adversarial Networks to image compression. The proposed method removes significant portions of an image while retaining some assistant information, and fills the gaps using generative model inpainting. We use Plug and Play Generative Networks as our inpainting network, and explore several different ablation schemes in order to determine the most useful information present in an image, according to the quality of the reconstruction.

Asynchronous Hebbian Learning
Minecraft Demo  /  GitHub

Asynchronous neural network implementation using Hebbian learning rules. Later adapted to replace local Hebbian rules with RL approach based on DDPG.

Raspberry PI cluster

Multiple node cluster built out of 32 Raspberry PI B+. Networked together using a standard network switch. Bootstrapped a custom power supply originall intended for Christmas lights. Utilized Erlang to communicate between the compute nodes and run the above algorithm across all machines.


Machine Learning Decal