|April 22 · Issue #81 · View online |
Welcome to another week in deep learning.
We learn, this week, why RNN’s and LSTM are going the way of the dinosaurs and how to handle bias in text embeddings.
| Artificial Intelligence — The Revolution Hasn’t Happened Yet |
A thought-provoking piece by one of the giants in machine learning, Michael Jordan, on the state of progress in human imitative A and the challenges to engineering complex systems that make sensible use of data flows in our lives.
| Researchers Teach AI to Think Like Dogs and Find Out What They Know About the World - The Verge |
This verge article explains a recent paper that describes how to model a visually intelligent agent on the behavior of dogs and how these results suggest that animals could provide a new source of training data for AI systems.
| AI in Banking: The Reality Behind the Hype – Financial Times |
As if to underline Michael Jordan’s point from the above article we get a look behind the curtain of the banking industry, where the promise of human imitative AI fell short of grandiose predictions. As one banking executive puts it in the article: “The problems we have solved are very narrow,” she adds. “The misconception is that humans and machines can perform at the same level. There’s still a long way to go and many challenges we need to solve before a machine can operate [at a level] even near the human mind.”
| AI in the UK |
A report commissioned by The Select Committee on Artificial Intelligence on the state of AI in the UK.
| Personalized Hey Siri |
In this blog post Apple engineers detailed how moving from a linear discriminative technique (LDA) to deep neural networks enabled them to add speaker recognition to the iPhone’s “Hey Siri” feature.
JupyterCon focuses on real-world practices and how to successfully implement interactive computation in your workflow and projects. You’ll discover the best practices for collaborative and reproducible data science; new use cases, and the expertise you need to transform your workflow with Jupyter tools.
| The fall of RNN / LSTM |
We fell for Recurrent neural networks (RNN), Long-short term memory (LSTM), and all their variants. Now it is time to drop them! After ResNet and Attention it became clear that LSTMs were just a clever bypass technique, see this post
| Instance Embedding: Segmentation Without Proposals |
A thorough review of three papers in the field of instance segmentation that outline approaches that differ from the mainstream proposal-based Faster-RCNN.
| Google Developers Blog: Text Embedding Models Contain Bias |
Man is to doctor as woman is to nurse, this biased semantic relationship is encoded when training word embeddings on a conventional data set such as the Google News corpus. This blogpost uses The Word Embedding Association Test (WEAT) to explore some of these problematic associations.
| Hallucinogenic Deep Reinforcement Learning using Python and Keras |
| Jupyter: Tips, Tricks, Best Practices with Sample Code for Productivity Boost |
A short guide including tips and tricks on making the best of your Jupyter notebooks.
| Introducing TensorFlow Probability |
TensorFlow Probability is a probabilistic programming toolbox for machine learning researchers and practitioners to quickly and reliably build sophisticated models that leverage state-of-the-art hardware.
| ModelDepot |
ModelDepot uses Tensorflow.js to train Tiny YOLO
model to detect objects from your webcam.
| Paper Repro: Deep Neuroevolution |
This thoughtful post on reproducing the seminal Uber paper “Deep Neuroevolution” is easy to follow and includes instruction on how to reproduce the reproduction!
| Research Blog: Looking to Listen: Audio-Visual Speech Separation |
In “Looking to Listen at the Cocktail Party”, these Google authors present a deep learning audio-visual model for isolating a single speech signal from a mixture of sounds such as other voices and background noise. They are able to computationally produce videos in which speech of specific people is enhanced while all other sounds are suppressed.
| Imagine This! Scripts to Compositions to Videos |
Fascinating paper describing a model that is capable of generating realistic scenes by learning from labelled video. The model is dubbed CRAFT forComposition, Retrieval, and Fusion Network which explicitly predicts a temporal-layout of mentioned entities (characters and objects), retrieves spatio-temporal entity segments from a video database and fuses them to generate scene videos.
| Spherical CNNs |
Convolutional Neural Networks (CNNs) have become the method of choice for learning problems involving 2D planar images. However, a number of problems of recent interest have created a demand for models that can analyze spherical images. Examples include omnidirectional vision for drones, robots, and autonomous cars, molecular regression problems, and global weather and climate modelling.
This paper introduces the building blocks for constructing spherical CNNs proposing a definition for the spherical cross-correlation that is both expressive and rotation-equivariant.
| Adversarial Attacks Against Medical Deep Learning Systems |
The discovery of adversarial examples has raised concerns about the practical deployment of deep learning systems. In this paper, the authors argue that the field of medicine may be uniquely susceptible to adversarial attacks, both in terms of monetary incentives and technical vulnerability. To this end, they outline the healthcare economy and the incentives it creates for fraud, extend adversarial attacks to three popular medical imaging tasks, and provide concrete examples of how and why such attacks could be realistically carried out.