News in artificial intelligence and machine learning
From 13th thru 29th Feb. Referred by a friend? Sign up here. Want to share? Give it a tweet.
Technology news, trends and opinions
DeepMind made the news by launching their Health division along with two early-stage projects, Streams and Hark. The former focuses on acute kidney injury, which progresses within 7 days, has multifactorial causes but is preventable in 25% of cases. The Streams app sends test results and physiological readings to physicians while on the go. Hark, on the other hand, is a clinical task manager app. Early evidence suggests that users find it more effective, efficient and less distracting than incumbent pagers.
At MWC last week, Mark Zuckerberg vouched for the proliferation of VR and the work of FB's Connectivity Lab applying computer vision on DigitalGlobe high-res satellite imagery to inform their strategy in delivering internet connectivity to the unplugged world. Of note, the team trained deep CNNs typically used on FB photos on 8k binary labeled (does/doesn't contain a building) satellite images from one country to analyse 21.6M square kilometers from 20 countries. This is an intriguing example of FB/GOOG/AMZN etc. encroaching on territory previously addressed by startups (e.g. Orbital Insight).
Why vertical combined with domain expertise and unique data acquisition strategies beat horizontal approaches for enterprise products underpinned by machine learning.
The XPRIZE organisation and IBM Watson will offer a $5m prize to the best demonstration of human-machine collaboration with cognitive computing by 2020.
The Allen AI Science Challenge finished their 4-month long competition on Kaggle to solve 8th grade multiple choice exam, which tested the natural language understanding abilities of user generated models. The winning score (1478 entries from 302 players) came in at 59.3%. Random guessing would score 25%.
On the topic of machine learning competitions, Numerai uses homomorphic encryption on financial data to create a tournament where data scientists build models to address stock market efficiency.
Here's a nice summary piece in Nature on the enabling macro trends driving the proliferation of AI with infographics to match.
Following the proliferation of 'neural art' created by Google's DeepDream project and others, the company is running a two-day event in SF with the Gray Area Foundation for the Arts to explore the intersection of AI and art.
Research, development and resources
How to code and understand DeepMind's neural stack machine. This in-depth tutorial explores the approaches of the original paper (here), namely using RNNs to learn processes important for machine translation tasks (long-range reorderings and substitutions).
Snips publish NTM-Lasagne: A library for Neural Turing Machines in Lasagne. The library is benchmarked against and reproduces the copy, repeat copy and associative recall tasks from Google DeepMind’s original paper.
Around the World in 60 Days: Getting Deep Speech to work in Mandarin, Baidu Silicon Valley AI Lab
Learning functions across many orders of magnitudes, Google DeepMind. When training deep reinforcement learning models to play Atari games, significant variation in the magnitudes and frequencies of rewards between games made a priori data normalisation using domain knowledge a requirement. Here, the authors present an adaptive normalisation technique to remove this heuristic.
SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <1MB model size, UC Berkeley and Stanford. The motivation behind this work is to create smaller DNN architectures that require less communication between servers during distributed training, less bandwidth between the cloud and edge device and thus be tractable for limited-memory hardware.
PlaNet - Photo Geolocation with Convolutional Neural Networks, Google. Instead of using image retrieval techniques, the authors subdivide the Earth into a set of 26k geographical cells as the target classes and train a CNN using 126M geotagged images. The model outputs a probability distribution over the Earth representing the likelihood that the test image is located in a particular cell.
Venture capital financings and exits
$77m worth of deal making, including:
Digital Reasoning, whose product Synthesys ingests (un)structured data from the web and client data stores to produce a knowledge graphs that can be queried against/used for alerts, raised a Series D from Goldman Sachs, Silver Lake and Nasdaq. The company now has an exclusive tie up with Nasdaq for the surveillance of global capital markets.
Reflektion, a predictive analytics and e-commerce content personalisation company, raises a $18m Series B from Battery ventures.
Kika Tech, which markets a predictive keyboard for Android, raised a $30m Series B. The company claims 130m downloads and 20m daily active users.
Talla, the virtual assistant focused on corporate recruiting, marketing and office management, raises a $4m seed round. The company is led by the founder of Backupify (cloud data backup), which sold to Datto for $90m.
PredictionIO, the popular open source machine learning server, was acquired by Salesforce for an undisclosed sum. The business launched in 2014, raised $2.5m and employed a team of 8.
Anything else catch your eye? Just hit reply! I’m actively looking for entrepreneurs building companies that build/use AI to rethink the way we live and work.