Interesting talk about how ensemble learning was used in the Netflix contest and a short presentation on statistical bootstrapping.

• The basic idea is that multiple models are used and they are fitted to the data they work best with.
• Usually, we compare models and take the best one but what if instead we combine different models and take the best characteristics of each?
• Cross-validation - using a portion of the sample data to check of over-fitting.

Meetup Event

This is one of best examples I’ve seen yet of using visualizations to explain something.

Raft is a distributed consensus algorithm designed to be understandable as opposed to PAXOS which as far as I can tell nobody understands (even people who have implemented it).

Raft Post
Main Site

The main point seemed to be that it’s easier to consider localization at the start of a project (even for one language) than add it later.

Meetup Event

Thoughts, terms, and ideas I’ve come across over the last few months.

• Mobile experiences fill gaps while we wait. Nobody wants to wait while they wait.
  1. Preform actions optimistically
     • Show +1/Like/Comment before the request is even sent. Show an error if it fails.
  2. Adaptively pre-load content
     • Load data before it’s needed.
     • Re-prioritize what to load based on user interest.
  3. Move bits when no one is watching
     • Send contact list while signing in.
     • Start uploading files while the user is filling out the details.
     • Send data as soon as part of it is ready to go and match it up on the server later.
• Code reviews are like having a shared brain.
• Shed load on social media sites by having shorter feeds.
• Achieve loose coupling using notifications, events, signals, etc.
• Project scaling
  • Code standards so everything looks the same
  • Unit tests and design documents so it’s easy to switch implementations and modify.
  • All you need a work priority queue with ideas and bugs.
  • Be transparent.
  • Don’t be date focused! It’s too idealistic.
  • Have short cycles, quick deliverables, and frequent estimates and re-estimates.
• Web crawling techniques
  1. Start at seed page and recursively follow all/subset of links.
  2. Identify a pattern in the URL for pages you want. Ex: resource/id/ and check every valid id in range.
  3. Read the sitemap and choose which links to follow.
• Append only event stores
  • Observations about the world are recorded for perpetuity and the results of observations are calculated on demand.
  • Example: Changes -> Append to transaction log -> DB is a roll-up view of the changes captured in the transaction log.
  • Don’t delete or update anything. Just accrete new knowledge and distil new implications based upon your increasing knowledge.
  • It’s like a warehouse. Shipments come in, shipments go out, and at any point, we can check the current inventory levels.