|September 21 · Issue #96 · View online |
Howdy folks and welcome to another exciting week in deep learning!
As always happy reading and hacking!
| Machine Learning Confronts the Elephant in the Room |
News about how easily deep neural networks can be fooled has reached the mainstream and is garnering attention especially in light of autonomous driving advance and high profile set backs. The article, however, also points towards recent research
into how to deal with adversarial attacks and how to make object detection systems more robust.
| AI and the News: An Open Challenge |
This open challenge, which will award up to $750,000 to a range of projects, is seeking fresh and experimental approaches to four specific problems at the intersection of AI and the news:
- Governing the Platforms
- Stopping Bad Actors
- Empowering Journalism
- Reimagining AI and News
| Nvidia Launches the Tesla T4, its Fastest Data Center Inferencing Platform Yet |
Nvidia argues that the T4s are significantly faster than the P4s. For language inferencing, for example, the T4 is 34 times faster than using a CPU and more than 3.5 times faster than the P4. Peak performance for the P4 is 260 TOPS for 4-bit integer operations and 65 TOPS for floating point operations.
| Microsoft Acquires Lobe to Help Bring AI Development Capability to Everyone |
Microsoft acquired Lobe which offers a simple visual interface that promises to let people easily create deep learning powered apps that can understand hand gestures, hear music, read handwriting, and more.
| Implementation of Reinforcement Learning Algorithms. Python, OpenAI Gym, Tensorflow |
This is a veritable treasure trove of reading, tutorials and code to accompany Richard Sutton’s Book and David Silver’s course on Reinforcement Learning compiled by Denny Britz.
| Deep Learning models: Object detection with Go using TensorFlow |
Great tutorial on getting started with Deep Learning in Go and how to integrate with Tensorflow.
| Analysis of Deep Neural Networks for Pixel Processing |
Interesting post in which the authors measured the performance of a number of bounding box detectors with the result that two stage approaches are more performance intensive than one shot approaches while accuracy hardly differs and that there are diminishing returns in accuracy as the number of operations increases on the same architecture.
| Help! I Can’t Reproduce a Machine Learning Project! |
This is a very useful guide on what to do when one finds oneself in the frustration position of not being able to reproduce someone else’s result.
| Neural Processes in PyTorch |
Recently, Deepmind published Neural Processes
at ICML, billed as a deep learning version of Gaussian processes. This illuminating post shows the link between these and VAEs and demonstrates some shortfalls of the method.
| InteriorNet: Mega-scale Multi-sensor Photo-realistic Indoor Scenes Dataset |
InteriorNet is a massive new dataset released by researcher at the Imperial College London and Chinese furnishing-VR startup Kujiale
, it consist of photographs of complex, realistic interiors. and numbers around 1 million CAD models of different types of furniture and furnishing, which over 1,100 professional designers have subsequently used to create around 22 million room layouts.
| ProSR - A Fully Progressive Approach to Single-Image Super-Resolution |
Repository containing an independent implementation of the paper: “A Fully Progressive Approach to Single-Image Super-Resolution” - ProSR is a Single Image Super-Resolution (SISR) method designed upon the principle of multi-scale progressiveness. The architecture resembles an asymmetric pyramidal structure with more layers in the upper levels, to enable high upsampling ratios while remaining efficient. The training procedure implements the paradigm of curriculum learning by gradually increasing the difficulty of the task.
| GitHub - SimonKohl/probabilistic_unet: A U-Net combined with a variational auto-encoder that is able to learn conditional distributions over semantic segmentations. |
| Hamiltonian Descent Methods |
The authors propose an interesting family of optimization methods that achieve linear convergence using first-order gradient information and constant step sizes on a class of convex functions much larger than the smooth and strongly convex ones. This larger class includes functions whose second derivatives may be singular or unbounded at their minima. In sum, these methods expand the class of convex functions on which linear convergence is possible with first-order computation.
| Phrase-Based & Neural Unsupervised Machine Translation |
Machine translation systems achieve near human-level performance on some languages, yet their effectiveness strongly relies on the availability of large amounts of parallel sentences, which hinders their applicability to the majority of language pairs. This work investigates how to learn to translate when having access to only large monolingual corpora in each language.
| CGIntrinsics: Better Intrinsic Image Decomposition through Physically-Based Rendering |
Intrinsic image decomposition is a challenging, long-standing computer vision problem for which ground truth data is very difficult to acquire. The authors explore the use of synthetic data for training CNN-based intrinsic image decomposition models, then applying these learned models to real-world images. To that end, they present CGINTRINSICS , a new, large-scale dataset of physically-based rendered images of scenes with full ground truth decompositions.