From construction technologies such as precast to use of emerging technologies including artificial intelligence, machine learning, and data analytics …

Dubai's Roads and Transport Authority (RTA) has started experimenting the use of artificial intelligence (AI) technologies (machine learning …

A new Google Magenta project (created by an intern!) lets you mix lo-fi, hip-hop music tracks to build a custom music room in your browser, with no musical ability required. Magenta is designed to use Google's machine learning systems for the creation of art and music, and the Lo-Fi Player is a fun example of what it can do. When you open Lo-Fi Player, you're taken to a pixellated virtual "room" where you click different objects -- a clock, a cat, or a piano, for instance-- in the room to change the different tracks, like the bass line and the melody. "The view outside the window relates to the background sound in the track, and you can change both the visual and the music by clicking on the window," Lo-Fi Player creator Vibert Thio wrote in a blog post. Thio writes that the team chose the format of a music-generating room rather than a composition tool or musical instrument because it's "a popular genre with a relatively simple music structure."

Convolutional and recurrent neural networks have been widely employed to achieve state-of-the-art performance on classification tasks. However, it has also been noted that these networks can be manipulated adversarially with relative ease, by carefully crafted additive perturbations to the input. Though several experimentally established prior works exist on crafting and defending against attacks, it is also desirable to have theoretical guarantees on the existence of adversarial examples and robustness margins of the network to such examples. We provide both in this paper. We focus specifically on recurrent architectures and draw inspiration from dynamical systems theory to naturally cast this as a control problem, allowing us to dynamically compute adversarial perturbations at each timestep of the input sequence, thus resembling a feedback controller. Illustrative examples are provided to supplement the theoretical discussions.

Whether human, or human-designed Artificial Intelligence, each represents a response to an era's big challenge. His ship leaked and was nearing the end of its useful life. He had but a few crude instruments to suggest, without much certainty, where he might be. And, with only wind and his sails to propel the craft, he had no way to fight through the howling storms of the North Atlantic. The voyage, which should have taken two to three weeks, ultimately took more than two months.

Following on from yesterday's post, today I looked at more lessons in Introduction to Machine Learning from the Google News Initiative. The separation of machine learning into three different approaches -- supervised learning, unsupervised learning, and reinforcement learning -- is standard (Lesson 3). In keeping with the course's focus on journalism applications of ML, the example given for supervised learning is The Atlanta Journal-Constitution's deservedly famous investigative story about sex abuse of patients by doctors. Supervised learning was used to sort more than 100,000 disciplinary reports on doctors. The example of unsupervised learning is one I hadn't seen before.

Irina Rish, now a renowned expert in the field of artificial intelligence, first became drawn to the topic as a teenager in the former Soviet republic of Uzbekistan. At 14, she was fascinated by the notion that machines might have their own thought processes. "I was interested in math in school and I was looking at how you improve problem solving and how you come up with algorithms," Rish said in a phone interview Friday afternoon. "I didn't know the word yet (algorithm) but that's essentially what it was. How do you solve tough problems?"

TL;DR This paper uses a technique called "iterative nullspace projection (INLP)" in order to debias word embedding representations. They do this by training linear classifiers to predict a certain attribute that they don't want the word embeddings to contain (e.g., race or gender information) and then project the weights of this classifier into the nullspace of the weight matrix. They perform this process iteratively in order to increasingly remove bias components. I personally found the paper really interesting because they took a bit of a different approach to typical NLP papers. One thing that I do find that needs work, however, is that their method does not fully address the "bias by neighbors" issue.

An artificial intelligence (AI) technology made by a firm co-founded by billionaire Elon Musk has won praise for its ability to generate coherent stories, novels and even computer code but it remains blind to racism or sexism. GPT-3, as Californian company OpenAI's latest AI language model is known, is capable of completing a dialogue between two people, continuing a series of questions and answers or finishing a Shakespeare-style poem. Start a sentence or text and it completes it for you, basing its response on the gigantic amount of information it has been fed. This could come in useful for customer service, lawyers needing to sum up a legal precedent or for authors in need of inspiration. While the technology is not new and has not yet learnt to reason like a human mind, OpenAI's latest offering has won praise for the way its text resembles human writing.

This is a problem if we want AIs to be trustworthy. That's why Diffbot takes a different approach. It is building an AI that reads every page on the entire public web, in multiple languages, and extracts as many facts from those pages as it can. Like GPT-3, Diffbot's system learns by vacuuming up vast amounts of human-written text found online. But instead of using that data to train a language model, Diffbot turns what it reads into a series of three-part factoids that relate one thing to another: subject, verb, object.