Learning Tree International just published my article on using Microsoft Cognitive Services to perform face recognition and emotion detection.
Status of Spark MLlib wrappers in SparkR
Wrappers for Spark’s MLlib machine learning library in SparkR have been slow to arrive. However, the future looks bright.
The imminent 2.0 release will bring k-means support to SparkR and the 2.1 release is scheduled to include wrappers for the following machine learning stalwarts
- Alternating Least Squares (ALS)
- Decision Trees
- Gaussian Mixture Models
- Isotonic Regression
- Latent Dirichlet Allocation (LDA)
- Multilayer Perceptron Classifiers
- Random Forests
The Mathematics of Machine Learning
Wale Akinfaderin, a research scientist intern at IBM Research, has written an article on the mathematical background required to fully understand machine learning.
He includes a list of educational resources for those who wish to brush up in certain areas.
Wale correctly stresses that you don’t need this level of knowledge to start making use of machine learning. There are many packaged solutions. However, if you wish to master the topic, some mathematical prowess is inescapable.
First death in a self-driving car
The first death in a self-driving car occurred in May 2016. It was in a Tesla Model S travelling along a Florida highway.
Apparently the car’s Autopilot system got confused by a white truck against a bright sky and failed to brake. When AI systems fail, they sometimes fail big.
Tesla noted in a press release this is the first fatality in 130 million miles, compared to an average of one fatality per 94 million miles of regular driving. While those numbers are on the right side, they don’t seem high enough to convince people to give up control.
Of course, this is both a technological progress and a systems problem. AI in cars is only going to get better. And, as we have more self-driving cars on the road, their predictability, and ability to communicate their intentions, should lead to quantum leaps in safety. Of course there’s always the fear of a system-wide failure that would result in the worst pileup in history.
What I found strange was another point Tesla made in the press release—that the self-driving feature is in beta and only designed to be semi-autonomous. Their definition of semi-autonomous is that you need to keep your hands on the steering wheel. That would seem to suggest that all their customers have been recruited as testers on public roads! Sitting with my hands on the steering wheel while my car drives would, I think, dull my reaction times to an unsafe level.
Machine learning algorithm cheat sheet
The recent explosion of interest in machine learning has resulted in a profusion of algorithms. It can be difficult to know which one is most suited to your problem.
Recognizing this challenge Microsoft have produced a machine learning algorithm cheat sheet. It’s designed to allow you to choose between the algorithms available in Microsoft’s Azure Machine Learning Studio, but, as many of the algorithms are generic, it’s applicable to other machine learning toolkits.
By following a path from your general task, and desirable features of the model, you end up at a suggested algorithm. For example, if you are looking to predict values and need accuracy and fast training, then a decision forest regression is suggested.
Of course, selection of an appropriate machine learning algorithm is a non-trivial task. A cheat-sheet is no replacement for a trained data scientist. And, the cheat sheet has obvious weaknesses. For example, a neural network regression is classified as an accurate approach to predicting values, but with a long training time. Given that, why would I ever pick a neural network over a decision forest when faced with a regression problem?!
However, if you are new to machine learning, and are bewildered by the choices, it’s a good place to start.