|June 1 · Issue #42 · View online |
Happy reading and hacking.
As always if you enjoy receiving this newsletter you can help us out by sharing it with friends and colleagues (or even family).
| ICML Accepted Papers Institution Stats |
Andrej Karpathy analyses the institutional statistics of ICML papers with some interesting results. As it turns out about 20-25% of all papers have industry involvement with Google/DeepMind leading the pack.
| Under The Hood Of Google’s TPU2 Machine Learning Clusters |
Google unveiled its second-generation TensorFlow Processing Unit (TPU2) at Google I/O last week, however, it provided very little additional information. This post gleans quite a few interesting details from the photos that were released of the chip.
| Apple Is Working on a Dedicated Chip to Power AI on Devices |
Given the central importance of AI for Apple to stay competitive the move to own this critical part of the stack seems natural. For more infos we’ll have to wait for WWDC in June.
| A16Z AI Playbook |
Cool resource by Andreesen Horowitz on what is possible with AI for both technical and non-technical folks.
| AlphaGo, in context by Andrej Karpathy |
Great overview of how Alpha Go combines different (fairly standard) machine learning techniques to achieve extraordinary results. Salient point: While AlphaGo does not introduce fundamental breakthroughs in AI algorithmically it does epitomize Alphabet’s AI power.
| Convolutional Methods for Text |
An accessible introduction to using convolutional neural networks for traditional NLP tasks.
| CycleGAN |
An in-depth explanation of the CycleGAN paper including a walkthrough.
| OpenAI Baselines: High Quality Implementations of Reinforcement Learning Algorithms |
OpenAi dropped an educational value bomb; a repository full of implementations of reinforcement learning algorithms with performance on par with published results. See Github
for the code.
| Real-time object detection with YOLO |
Implementing the YOLO object detection neural network in Metal on iOS.
| Deep-Image-Analogy: Visual Attribute Transfer through Deep Image Analogy. |
Source code for the paper ‘Visual Attribute Transfer through Deep Image Analogy’.
| pix2code: Generating Code from a Graphical User Interface Screenshot |
pix2code leverages Deep Learning techniques to automatically generate code given a graphical user interface screenshot as input. The model is able to generate code targeting three different platforms (i.e. iOS, Android and web-based technologies) from a single input image with over 77% of accuracy. See paper
| The Behance Artistic Media Dataset |
- Automatically-labeled binary attribute scores for over 2.5 million images across 20 attributes each.
- 393,000 crowdsourced binary attribute labels for individual images
- Short image descriptions/captions for 74,000 images from the crowd
- Image URLs for all images mentioned above
| Mapillary - Crowdsourced Street Photos |
- 25,000 Images
- 100 Categories
- 60 Instance-wise Categories
- 6 Continents
- Variety of Weather, Season, Time of Day, Camera, and Viewpoint
| On-the-fly Operation Batching in Dynamic Computation Graphs |
The authors of this paper present an algorithm, and its implementation in the DyNet toolkit, for automatically batching operations. Developers simply write minibatch computations as aggregations of single instance computations, and the batching algorithm seamlessly executes them, on the fly, using computationally efficient batched operations.
| The Marginal Value of Adaptive Gradient Methods in Machine Learning |
Adaptive optimization methods, which perform local optimization with a metric constructed from the history of iterates, are becoming increasingly popular for training deep neural networks. The authors construct an illustrative binary classification problem where the data is linearly separable, GD and SGD achieve zero test error, and AdaGrad, Adam, and RMSProp attain test errors arbitrarily close to half.