Deep Learning Weekly Issue #137

Apple acquires XNOR.AI, PyTorch 1.4.0, Efficient Transformers, AutoML from Amazon, and more...

Hey folks,

This week in deep learning we bring an acquisition from Apple, a new AutoML tool from Amazon, PyTorch 1.4.0, an autonomous dog training system, and an efficient transformer from Google.

You may also enjoy “A sober look at Bayesian Neural Networks”, a guide to training object detectors with no real data using domain randomization, a sequence model visualization tool from Facebook, an autonomous reinforcement learning library for PyTorch, and a new high-resolution segmentation model architecture.

As always, happy reading and hacking. If you have something you think should be in next week's issue, find us on Twitter: @dl_weekly.

Until next week!


Apple acquires for a reported $200M

A week after news of Snap’s acquisition of AI Factory, it’s reported that Apple has acquired XNOR to bolster their lower power AI projects.

AWS Introduces Open Source AutoML Toolkit ‘AutoGluon’

Amazon open sources an AutoML library with support for hyperparameter tuning, model selection, and more.

PyTorch 1.4.0 Released

Mobile build customization, Distributed model parallel training, Java bindings, and more.


3DPeople Dataset

2 million synthetically generated images of 80 humanoid models including 3D pose annotations, body part segmentation, and more.


Autonomous Dog Training with Companion

Creating an image recognition system that automatically train dogs to help get them ready for adoption.

Reformer: The Efficient Transformer

A transformer model from Google specializing in extremely large context windows of up to 1 million words.

A Sober Look at Bayesian Neural Networks

Asking the question “Do Bayesian Neural Networks make sense?”

Deep Reinforcement Learning With TensorFlow 2.1

Implemented Deep RL techniques in TF2.

Training Object Detectors with No Real Data using Domain Randomization

Solving sim2real transfer for specialized object detectors with no budget

Libraries & Code

[Github] HRNet/HRNet-Semantic-Segmentation

High-resolution representation learning (HRNets) for Semantic Segmentation

VizSeq: A visual analysis toolkit for accelerating text generation research

Facebook releases an open source tool for visualizing text generation models.

[Github] /kheyer/Genomic-ULMFiT

An implementation of ULMFiT for genomics classification using Pytorch and Fastai.

[Github] cpnota/autonomous-learning-library

The Autonomous Learning Library: A PyTorch Library for Building Reinforcement Learning Agents

[Github] google-research/flax

A neural network library for JAX designed for flexibility

Papers & Publications

Advbox: a toolbox to generate adversarial examples that fool neural networks

Abstract: ....Advbox is a toolbox to generate adversarial examples that fool neural networks in PaddlePaddle, PyTorch, Caffe2, MxNet, Keras, TensorFlow and it can benchmark the robustness of machine learning models. Compared to previous work, our platform supports black box attacks on Machine-Learning-as-a-service, as well as more attack scenarios, such as Face Recognition Attack, Stealth T-shirt, and Deepfake Face Detect….

HybridPose: 6D Object Pose Estimation under Hybrid Representations

Abstract: We introduce HybridPose, a novel 6D object pose estimation approach. HybridPose utilizes a hybrid intermediate representation to express different geometric information in the input image, including keypoints, edge vectors, and symmetry correspondences. Compared to a unitary representation, our hybrid representation allows pose regression to exploit more and diverse features when one type of predicted representation is inaccurate (e.g., because of occlusion). HybridPose leverages a robust regression module to filter out outliers in predicted intermediate representation…. Compared to state-of-the-art pose estimation approaches, HybridPose is comparable in running time and is significantly more accurate. For example, on Occlusion Linemod dataset, our method achieves a prediction speed of 30 fps with a mean ADD(-S) accuracy of 79.2%, representing a 67.4% improvement from the current state-of-the-art approach.