In the closing weeks of 2016, Google published an article which quietly sailed under most people's radar. Which is a shame, because the article may just be the most astonishing thing about machine learning that I read last year. Don't feel bad if you missed it. Not only was the article competing with the pre-Christmas rush most of us were navigating, it was also tucked away on Google's Research Blog beneath the geektastic headline Zero-Shot Translation with Google's Multilingual Neural Machine Translation System. It doesn't exactly scream must read, does it?

Machine learning was undoubtedly one of the hottest topics in radiology last year, with a steady stream of academic research papers highlighting how machine learning, particularly deep learning, can outperform traditional algorithms or manual processes in certain use-cases. Investment in machine learning start-ups also continued, with several companies attracting early stage funding. To date, more than $100m has been invested in start-ups that are developing AI solutions for radiology. Furthermore, commercial activity gained pace, with at least 20 companies exhibiting AI-based products at the RSNA conference towards the end of the year, although most were prototypes and only a handful had regulatory clearance. Whilst the enthusiasm for machine learning is certainly justified, it inevitably raises expectations, potentially to unrealistic levels.

The images above are examples of the three possible classes I needed to predict: no traffic light (left), red traffic light (center) and green traffic light (right). The challenge required the solution to be based on Convolutional Neural Networks, a very popular method used in image recognition with deep neural networks. The submissions were scored based on the model's accuracy along with the model's size (in megabytes). Smaller models got higher scores. In addition, the minimum accuracy required to win was 95%.

Maintaining data reliability is a resource-intensive uphill task for many organizations. Companies often spend too much effort on data reviews and cleanup, but seldom seem to catch up. Most of the time, teams don't even know what the issues are, how to look for them, and how to solve them. They just know that the data is dirty, and like a sitting on a ticking time bomb, we wait for the disaster to happen. The issues are often only illuminated when the data is put to operational use and trips up the end user or the customer with wrong information.

It seems pretty clear that 2017 is going to be the year that Amazon's Alexa virtual assistant really begins to hit its stride, going from a fast-growing niche product to a mainstream must-have. While Microsoft and especially Google have made their competitive strategies clear -- even Samsung and Baidu have started to make rumbles in the market -- there's one elephant that, notably, isn't in the room yet: Apple, the most valuable company in the world and a notorious latecomer to any new product category. While Apple recently built Siri into the Apple TV, the company is said to be working on a direct competitor to the Amazon Echo -- one that would apparently be more advanced than anything we've seen, down to a possible facial-recognition camera so it knows who's talking. That device, if and when it comes out, would bring Apple's 5-year-old Siri assistant head-to-head with Alexa. And the clock is ticking.

Many important real-world datasets come in the form of graphs or networks: social networks, knowledge graphs, protein-interaction networks, the World Wide Web, etc. (just to name a few). Yet, until recently, very little attention has been devoted to the generalization of neural network models to such structured datasets. In the last couple of years, a number of papers re-visited this problem of generalizing neural networks to work on arbitrarily structured graphs (Bruna et al., ICLR 2014; Henaff et al., 2015; Duvenaud et al., NIPS 2015; Li et al., ICLR 2016; Defferrard et al., NIPS 2016; Kipf & Welling, 2016), some of them now achieving very promising results in domains that have previously been dominated by, e.g., kernel-based methods, graph-based regularization techniques and others. In this post, I will give a brief overview of recent developments in this field and point out strengths and drawbacks of various approaches. I wrote a short comment on Ferenc's review here (at the very end of this post).

POSB, one of Singapore's oldest banks and part of the DBS Banking Group, has launched an online virtual assistant, POSB digibank Virtual Assistant. It is powered by the KAI conversational bot/artificial intelligence (AI) platform from a New York-based fintech start-up, Kasisto. POSB's chatbot is available on Facebook Messenger and can answer questions relating to account balances, utility bill payments and fund transfer requests. It will also be rolled out to the WhatsApp and WeChat messaging platforms. Kasisto already supplies it flagship platform to DBS's subsidiary in India, DBS digibank. "We know that our customers are spending time conversing on their favourite mobile messaging apps, and we are immersing ourselves in the customer journey by making it easier and more convenient for them to engage us," explains Jeremy Soo, head of consumer banking group, Singapore, DBS Bank.

Most college students relax over their winter break, eating good food and de-stressing from the previous semester. But when I first talked with Joshua Browder, a Stanford University sophomore, he was busy finishing up schoolwork. I was lucky he had time to talk in the middle of his busy schedule. He's the founder of DoNotPay, a chatbot that helps overturn traffic tickets, and in a few days would be flying to London to meet with government officials about using his technology. Then off to Munich to speak at an international design and innovation conference.

Web users searching for photos and cops looking for suspects in video already benefit from software that understands the content of images. Chris Gibson says it can also make it easier to find treatments for diseases not targeted by existing drugs. "By combining robotics and machine vision, we can work at large scale on hundreds of diseases simultaneously, using a small number of people," says Gibson, who is CEO and cofounder of the 40-person startup Recursion Pharmaceuticals. Recursion uses software to read out the results of high-throughput screening, which automates drug testing in cells. That isn't a new idea, but Recursion uses algorithms that inspect cells at an unusual level of detail.

Voice-controlled assistants are having a moment. But there may be an intriguing wrinkle that their makers have to smooth out: users don't seem to be using many apps. The number of Skills--the Amazon name for apps that operate on its Alexa smart assistant software--available for the company's Echo smart speaker have risen significantly in the past six months, from 950 last May to over 8,000 today. But an analysis of the way people use Alexa and and Google's Assistant platforms shows that third-party apps aren't too well used, nor particularly sticky. The analysis, which was carried out voice software start-up Voice Labs, shows that most app don't get any user reviews, which suggests they're not very popular.