Google DeepMind Internship
Deep Reinforcement Learning group
Google Research Internship
Large-Scale Supervised Deep Learning for Videos.
Stanford Computer Science Ph.D. student
, Computer Vision, Natural Language Processing. My adviser is Fei-Fei Li.
Google Research Internship
Large-Scale Unsupervised Deep Learning for Videos.
University of British Columbia: Master's Degree
I worked with Michiel van de Panne on learning Compositional Controllers for Physically-simulated Articulate Figures.
University of Toronto: Bachelor's Degree
Double major in Computer Science and Physics.
. I am currently a 5th year Computer Science PhD student at Stanford, working with Fei-Fei Li
. My research centers around Deep Learning and its applications in Computer Vision, Natural Language Processing and their intersection. In particular, I am interested in fully end-to-end learning with Convolutional/Recurrent Neural Networks architectures and recent advances in Deep Reinforcement Learning and attentional mechanisms. Over the course of my PhD I squeezed in two internships at Google, where I worked on large-scale feature learning over YouTube videos, and this summer I interned at DeepMind and worked on Deep Reinforcement Learning. Earlier this year, together with Fei-Fei, I designed and taught a new Stanford undergraduate-level class on Convolutional Neural Networks for Visual Recognition (CS231n)
. The class was the first Deep Learning course offering at Stanford, had 150 students, and received very positive reviews. We're looking forward to offering it again this Winter quarter!
On a side for fun I maintain a blog (karpathy.github.io
). I am also sometimes jokingly referred to as the
reference human for ImageNet (post
:)), and I create / maintain those nice-looking conference proceedings LDA visualization pages each year (NIPS 2014 example
DenseCap: Fully Convolutional Localization Networks for Dense Captioning
Efficiently identify and caption all the things in an image with a single forward pass of a network. Our model is fully differentiable and trained end-to-end without any pipelines. The model is also very efficient (processes a 720x600 image in only 240ms), and evaluation on a large-scale dataset of 94,000 images and 4,100,000 region captions shows that it outperforms baselines based on previous approaches. Torch 7 code and data coming soon.
Justin Johnson*, Andrej Karpathy*, Li Fei-Fei
Visualizing and Understanding Recurrent Networks
We study both qualitatively and quantitatively
the performance improvements of Recurrent Networks in Language Modeling tasks compared to finite-horizon models. Our analysis sheds light on the source of improvements
, and identifies areas for further potential gains. Among some fun results we find LSTM cells that keep track of long-range dependencies such as line lengths, quotes and brackets.
Andrej Karpathy*, Justin Johnson*, Li Fei-Fei
Deep Visual-Semantic Alignments for Generating Image Descriptions
We present a model that generates natural language descriptions of full images and their regions. For generating sentences about a given image region we describe a Multimodal Recurrent Neural Network architecture. For inferring the latent alignments between segments of sentences and regions of images we describe a model based on a novel combination of Convolutional Neural Networks over image regions, bidirectional Recurrent Neural Networks over sentences, and a structured objective that aligns the two modalities through a multimodal embedding. This work was also featured in a recent New York Times article
Andrej Karpathy, Li Fei-Fei
CVPR 2015 (Oral)
ImageNet Large Scale Visual Recognition Challenge
Everything you wanted to know about ILSVRC: data collection, results, trends over the years, current computer vision accuracy, even a stab at computer vision vs. human vision accuracy -- all here!
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, Li Fei-Fei
Deep Fragment Embeddings for Bidirectional Image-Sentence Mapping
We train a multi-modal embedding to associate fragments of images (objects) and sentences (noun and verb phrases) with a structured, max-margin objective. Our model enables efficient and interpretible retrieval of images from sentence descriptions (and vice versa).
Andrej Karpathy, Armand Joulin, Li Fei-Fei
Large-Scale Video Classification with Convolutional Neural Networks
We introduce Sports-1M: a dataset of 1.1 million YouTube videos with 487 classes of Sport. This dataset allowed us to train large Convolutional Neural Networks that learn spatio-temporal features from video rather than single, static images.
Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, Li Fei-Fei
CVPR 2014 (Oral)
Grounded Compositional Semantics for Finding and Describing Images with Sentences
Our model learns to associate images and sentences in a common
We use a Recursive Neural Network to compute representation for sentences and a Convolutional Neural Network for images. We then learn a model that associates images and sentences through a structured, max-margin objective.
Richard Socher, Andrej Karpathy, Quoc V. Le, Christopher D. Manning, Andrew Y. Ng
Emergence of Object-Selective Features in Unsupervised Feature Learning
We introduce an unsupervised feature learning algorithm that is trained explicitly with k-means for simple cells and a form of agglomerative clustering for complex cells. When trained on a large dataset of YouTube frames, the algorithm automatically discovers semantic concepts, such as faces.
Adam Coates, Andrej Karpathy, Andrew Ng
Locomotion Skills for Simulated Quadrupeds
We develop an integrated set of gaits and skills for a physics-based simulation of a quadruped. The controllers use a representation based on gait graphs, a dual leg frame model, a flexible spine model, and the extensive use of internal virtual forces applied via the Jacobian transpose.
Stelian Coros, Andrej Karpathy, Benjamin Jones, Lionel Reveret, Michiel van de Panne
Object Discovery in 3D scenes via Shape Analysis
Wouldn't it be great if our robots could drive around our environments and autonomously discovered and learned about objects? In this work we introduce a simple object discovery method that takes as input a scene mesh and outputs a ranked set of segments of the mesh that are likely to constitute objects.
Andrej Karpathy, Stephen Miller, Li Fei-Fei
Curriculum Learning for Motor Skills
My UBC Master's thesis project. My work was on curriculum learning for motor skills. In particular, I was working with a heavily underactuated (single joint) footed acrobot. The acrobot used a devised curriculum to learn a large variety of parameterized motor skill policies, skill connectivites, and also hierarchical skills that depended on previously acquired skills. Almost all of it from scratch. The project was heavily influenced by intuitions about human development and learning (i.e. trial and error learning, the idea of gradually building skill competencies). The ideas in this work were good, but at the time I wasn't savvy enough to formulate them in a mathematically elaborate way. The video is a fun watch!
Andrej Karpathy, Michiel van de Panne
is a Reinforcement Learning library that implements several common RL algorithms supported with fun web demos. The library includes DP,TD,DQN algorithms and sketches of stochastic/deterministic Policy Gradients.
ulogme tracks your active windows / keystroke frequencies / notes throughout the entire day and visualizes the results in beautiful d3js timelines. Check out my blog post introducing the project
to learn more.
Pretty Accepted Papers
Research Lei is an Academic Papers Management and Discovery System. It helps researchers build, maintain, and explore academic literature more efficiently, in the browser. (deprecated since Microsoft Academic Search API was shut down :( )
ScholarOctopus takes ~7000 papers from 34 ML/CV conferences (CVPR / NIPS / ICML / ICCV / ECCV / ICLR / BMVC) between 2006 and 2014 and visualizes them with t-SNE based on bigram tfidf vectors. In general, it should be much easier than it currently is to explore the academic literature, find related papers, etc. This hack is a small step in that direction at least for my bubble of related research.
. Pretty! You can also use tsnejs to embed (almost) arbitrary CSV data in this web interface
for visualizing the SVM. Related, I also write forestjs
which is a Random Forest library. These have been largely deprecated by ConvNetJS.
I'we written an iOS app
that helps people access and remember Rubik's Cube algorithms. I've later also ported it to Android. I also published a 2-4 player iPad game called Loud Snakes
This page was a fun hack. Google was inviting people to become Glass explorers through Twitter (#ifihadclass) and I set out to document the winners of the mysterious process for fun. I didn't expect that it would go on to explode on internet and get me mentions in TechCrunch
, Verge, and many other places.
I think I enjoy writing AIs for games more than I like playing games myself - Over the years I wrote several for World of Warcraft, Farmville, Chess, and Tetris
. On somewhat related note, I also wrote a super-fun Multiplayer Co-op Tetris
Even more various crappy projects I've worked on long time ago.
Hacker's Guide to Neural Networks
is my attempt at explaining Neural Nets from "Hacker's perspective", relying more on code and physical intuitions than mathematics. I wrote this because I felt there were many people (e.g. some software engineers) who were interested in Deep Nets but who lacked the mathematical background to learn the basics through the usual channels.
I helped create the Programming Assignments for Andrew Ng's CS229A (Machine Learning Online Class)
- this was the precursor to Coursera. At UBC I also TA'd CPSC540
(Graduate Probabilistic Machine Learning) and three times UBC's CPSC 121
(Discrete Mathematics), where I taught at tutorials.
I like to go through classes on Coursera and Udacity. I usually look for courses that are taught by very good instructor on topics I know relatively little about. Last year I decided to also finish Genetics and Evolution (statement of accomplishmnet
) and Epigenetics (statement
, + my rough notes
A long time ago I was really into Rubik's Cubes. I learned to solve them in about 17 seconds and then, frustrated by lack of learning resources, created YouTube videos
explaining the Speedcubing methods. These went on to become quite popular with 7 million+ views by now. There's also my cubing page badmephisto.com
that still gets a few thousand views a day. Oh, and a video of me at a Rubik's cube competition
for doing well in undergrad classes, for younglings.
- The Verge articles on NeuralTalk, first here
and then here
. Several inaccuracies, but by now quite used to it.
Still more unsorted misc
- Academic Countdown
was a site I mainted for a while that counted down to upcoming conference deadlines. It took too much time to maintain :) But the code is nice and people can use it to create their own countdowns.