Basic Deep Learning?

174340717_ad3b91846b_z

It was one of those days today when my Paper.Li aggregator sent me off to the USA via Serbia and Adelaide … and then on to Switzerland.

A Darren Burgess tweet

led me to Mladen Jovanović’s blog and to a December 2012 guest post by Carl Valle.

In his introduction to performance monitoring, Carl asks ‘Before we can start claiming to create proprietary metrics and innovative algorithms, are we doing a good job with the basics?’

In his view of the basics, Carl observes:

  • I would argue the real problem with data is not the data itself, it’s who is filtering the information and retelling the story of what is happening.
  • No matter if the team is using  individualized thresholds with IBM smarter planet software or teams cracking the code with special ratios of acceleration and deceleration during practice, people are still tearing ACLs.
  • The real question is how much daily data teams are really getting that is meaningful.
  • Solving the problem with ornamental data is going to be hard and require widely adopted standards and a lot of transparency. What are people really doing objectively and is it working?

Carl’s antidote to these data issues?

  1. Admit a problem exists.
  2. Getting back to fundamentals and reading timeless texts on sport science and training theory.
  3. Basic fitness and power tests can measure simple adaptations or decay of abilities over time.
  4. Having an elegant approach to data collecting and intervention strategies.

I missed Carl’s post first time round and so I am grateful to Darren for the alert. I am pleased too that Mladen invited Carl to write the guest post. February 2014 posts on Mladen’s blog include a post about using R and an interview with Mike McGuigan (AUT, New Zealand).

After finishing these readings, I moved on to a discussion of Deep Learning. I was mindful of Carl’s observation that:

With all of the sport sensors and video cameras collecting terabytes of data, it looks like we must hire experts in data warehousing to handle the “mounds of data” one is collecting during the season.

I think Carl’s step 4 above might help in this regard. The post was from April 2013 and discussed Ray Kurzweil’s work in Artificial Intelligence. A comment on the article took me to Jürgen Schmidhuber’s work and this review paper on Deep Learning.

Carl and Jürgen encouraged me to think about strategic and operational approaches to data.

The coaching part of me is fascinated by first principles These principles resonate with my knowledge discovery interests too.

I am left thinking that if coaches with Carl’s insights meet sensitive data scientists then we might overcome the overselling of emerging technologies and the under utilisation of the opportunities afforded by remote sensing.

I hope this kind of meeting of minds will promote the open sharing of experiences.

Photo Credit

Cables (Steve Mishos, CC BY-NC-SA 2.0)

Leave a Reply