Decision Mechanics

Insight. Applied.

Do you really need machine learning?

By

I have a lot of sympathy for the view expressed in the following tweet

Many organizations who dive into machine learning haven’t even started to extract value from their data.

I understand why they want to get on the train. If you’ve not managed to draw value from your data yet why not just shovel it all into these amazing algorithms and let insights flow out the other end. Jump to the latest technology. Makes sense.

Unfortunately that’s not how data science works. The success of any data science project depends on you understanding your business and your data.

Some things you might do before adopting machine learning

  • Work out what data you need to help you manage and improve your business
  • Introduce your managers to your data people
  • Clean and curate your data
  • Put systems in place that give you real-time access to useful simple statistics—averages, totals, trends, etc.
  • Use simple statistical tools—like linear regression—to get deeper insights
  • Make use of decision science techniques (e.g. game theory)—not everything is amenable to running masses of data through algorithms

If you’re ready for it, machine learning is fantastic. However, if you deploy it as a magic bullet, don’t be surprised when you shoot yourself in the head.

Google reveals performance of its custom AI chips

By

chart showing relative efficiency of different kinds of processing units

Google has just release performance data for its Tensor Processing Unit (TPU) custom machine learning chips.

TPUs are

  • 15-20x faster than GPUs and CPUs when running deep learning applications
  • 30-80x more power efficient (operations per Watt)
  • frugal in terms of lines of code required to control them (their deep learning applications are implemented using 100-1500 lines)

Google claim that without these chips they would have to double the number of data centers they run.

Social rating and machine learning

By

I recently blogged about the risks of social rating systems. Machine learning adds another dimension to this.

In her book “Weapons of Math Destruction”, Cathy O’Neil highlights the umaccountability of algorithms used to make decisions that have a significant impact on peoples’ lives. The details of these algorithms are often undocumented—for commercial or security reasons—making it difficult to challenge their conclusions.

Unaccountable big data algorithms serve to amplify the risks posed by social scoring.

Machine learning replacing traders at Goldman Sachs

By

trading dashboard

Traditionally automation has displaced low skilled labor. But the big opportunities for machine learning lie in the areas currently staffed by professionals.

Marty Chavez, deputy CFO at Goldman Sachs recently explained that automation had cost 600 equity traders their jobs. Furthermore, he went on to describe how the same innovation is being applied to areas such as currency trading and even some investment banking operations.

Goldman Sachs predict that they’ll be replacing traders with computer engineers at a ratio of 4:1. Out of the 9000 people employed at the bank around 3000 are engineers.