News in AI: when is ML worth it and for what, particle physics, healthcare
…10 days from 23rd November 2015. I'll be publishing an analysis and commentary on financings and exits for H2 2015 soon, as well as sharing content from a talk I gave last night at Re.Work's investing in deep learning dinner in London.
Technology news, trends and opinions
When is Machine Learning Worth it? A pragmatic piece by Ferenc Huszar who presents three scenarios in which ML techniques have differing value add: 1) some problems can be solved with simple heuristics, 2) others require ML to even come close to being solved (e.g. speech/object recognition), while a last category 3) see iterative ML improvements as a means to scale performance of an existing system (e.g. trading). Certainly the first camp is where you often hear companies pitch a cursory inclusion of ML in their product, while the second/third categories are where the meat of value creation lies.
Where are the Opportunities for Machine Learning Startups. Points to search, healthcare, cybersecurity and tools to improve the efficiency with which knowledge professional complete repetitive tasks. I’m with her on all of these points, but still on the lookout for the mammoth opportunity within healthcare. In the not too distant future, I’m betting we’ll be non-invasively monitoring physiology and lifestyle in real time (nature and nurture) to detect anomalies from a healthy state, produce more accurate diagnoses and and predict treatment response to improve healthcare outcomes.
More movements by large technology companies open sourcing parts of their AI intellectual property to gain developer mindshare: IBM open sourced a library called SystemML (specialising in algorithm customisability and automatic optimisation). This follows Google’s TensorFlow (discussed last week), Microsoft’s Distributed Machine Learning Toolkit, Intel’s Trusted Analytics Platform (focused at enterprise).
Artificial Intelligence Aims to Make Wikipedia Friendlier and Better. Here’s a good example of using ML to aid content editors to maintain the Wikipedia corpus by scoring the quality of edits and identifying mistakes.
Artificial Intelligence Called in to Tackle LHC Data Deluge. The two largest Large Hadron Collider experiments, which discovered the Higgs boson in 2012, produce hundreds of millions of collisions per second. Here, machine learning is used to help decide which collisions are of interest based on existing knowledge of prior events. A hot topic of discussion was whether deep learning could be applied to particle physics to discover new particles.
Machine Intelligence in the Real World. A discussion of market entry points for technologists working in the space. It’s a helpful framework when assessing new opportunities. Given the pace at which technology goes open source or is productised and provided for next to nothing, data and talent truly moves the needle. We see a renewed focus on building a solution to an unsolved/poorly served high-value, persistent problem for consumers or businesses vs. generalist services.
Research, development and resources
A Roadmap towards Machine Intelligence, Facebook AI Research and University of Trento. The authors present the general characteristics (namely communication and learning) that define intelligence machines and propose that we should be focusing on modelling all aspects of intelligence holistically within a single system. They present a simulated environment to aid in this regard.
Towards Principled Unsupervised Learning, Google Brain and Google DeepMind. Unlike supervised learning where the goal of a system is to minimise training error using labeled examples, unsupervised learning often deals with insufficient labelled examples. This makes it difficult for unsupervised cost functions to know which of the many possible supervised tasks the author cares about. Here, a new unsupervised cost function is presented and tested on speech recognition to prove efficacy on training functions without the use of input-output examples.
A Century of Portraits: A Visual Historical Record of American High School Yearbooks, UC Berkeley and Brown University. By mining a vast database of high-school yearbook photos, a machine-vision algorithm reveals the change in hairstyles, clothing, and even smiles over the last century (MIT Tech Review).
Doctor AI: Predicting Clinical Events via Recurrent Neural Networks, Georgia Institute of Technology. The authors use electronic health records (diagnosis, procedures and medications) of 250k patients over 8 years to prove that RNNs can be used to predict future medical events and the timing of these events. Moreover, an RNN trained on data from one hospital can be used to improve predictions for another hospital with insufficient patient records.
The Royal Society in London is beginning a project on machine learning engaging policymakers, academia, industry and the wider public. It will focus on applications in the next 5–10 years (project scope). They publish solid video and interactive content on the industry here.
An Introduction to Machine Learning Theory and Its Applications and Deep Learning. For those of you keen on palatable introductory content, here’s a visual tutorial with examples (includes some maths) walking you through key concepts.
Venture capital financings and exits
Progessa, a Canadian sub/near-prime consumer lending company, raised an $11.4m Series A. The business focuses on modelling loan eligibility, a popular application of data science and ML.
Mirador Financial, on the other hand, raised $7.2m Series A from Core Innovation Capital for their SME-focused lending product.
Anything else catch your eye? Drop me a line on @nathanbenaich. I'm actively looking for entrepreneurs building companies that build/use AI to rethink the way we live and work.