Last night was the annual Queens’ Computer Science dinner and we completely sold out of tickets for the second year in a row.
Ben and Matt did an excellent job as this year’s organisers – everything ran smoothly, and they had even arranged for some live music as we moved from the main dining hall to Old Hall after dinner (for more food).
I’d like to thank the individual alumni who sponsored the dinner and also the corporate sponsors: Palantir, Jane Street, Ocado, Coherent Graphics, and Encore. Your generous support is what makes the dinner possible.
We were lucky this year to have Siraj Khaliq as our guest of honour. Siraj studied Computer Science at Queens’ in 1997. He was one of the first 300 employees at Google and then went on to do his own startup The Climate Company which was acquired by Monsanto for $1.1 billion. Siraj gave a great talk about his experiences and had advice for those of us thinking about startups today.
One of his pieces of advice was to not try to build the ‘best’ thing when ‘good enough’ is good enough. I think this is worth remembering – I can think of numerous occasions in research where we’ve had an idea, let it get shot down by some corner case and then seen papers from people who’ve done the same thing, not worried about these cases, but still found something interesting.
I was really pleased to see so many graduates who came back and joined us. It was interesting to hear how everyone is getting on.
Here’s a shot of a few of us at the drinks reception:
Drinks reception in the Old Senior Combination Room
And here’s a picture of me with Professor Alan Mycroft and Dr Anil Madhavapeddy. Alan and Anil were chosen by the Queens’ undergrads as their favourite lecturers this year.
Thanks everyone for coming, thanks to Ben and Matt for their hard work organising, see you all next year.
Advait (our AI supervisor) gave a good talk last week on probabilistic datastructures. He showed how you could save orders of magnitude of space if you were prepared to have an almost right answer (with probability bounds).
He motivated the talk using the example of MildyInnappropriateCatAppreciationSociety.com which is trying to store unique pictures of cats. You can use a bloom filter to do the same job as a hash table here. If you had 40Mb of raw cat pictures then you could use a 0.6Mb bloom filter an an efficient way of looking up whether you already had a certain picture with 4% error. If you wanted an exact answer with (e.g.) a hash table it would take you 4Mb. Obviously in reality you would have petabytes of cat pictures – then the savings start to add up.
We then looked at cardinality estimation – this is the problem of estimating how many unique cat pictures we have. For our 40Mb example above you can do this in 125kb using a linear counter, or in 2kb using a loglog counter.
The next question was counting the frequency of each cat picture.
One might say we are trying to graph the long tail of cats.
A count-min sketch can do this in just 48kb of space (with 4% error).
We went in to each of the above in detail to understand how they worked and Advait had a quiz on each one to see how they would store some example data.
Advait is using probabilistic algorithms in his research on visualising large datasets and so to close he showed us a sneak preview of his upcoming publication on this.
The talk was presented really clearly and confidently with lots of audience interaction. All the talk of cat databases gave us a real world (!) scenario for the algorithms. And it was nice to hear a bit about his research too.
On Wednesday last week we played Robocode.
For those of you who’ve not seen it before its basically a simulator for battles between robot tanks. You play the game by writing Java code to control your tank and try to destroy everyone else.
We had three teams, each of which was tasked with writing two robots. Then, after 2 hours, we had a tournament to see which robot was best.
Here’s a video of Sid2 fighting Larry. Sid2 (you can guess who wrote that) was built by copying and pasting example code off the web (this turned out to be a very good strategy for building a good robot in 2 hours). Larry was a more customised effort featuring a colour changing technique (for no particular benefit in battle):
After beating Sid2 we thought Larry was going to emerge the winner but then ProbablyTerribleRobot came on the scene:
Here’s a video of all of our robots in melee. Special mention to Robotina and FirstRobot for their very novel movement patterns…