Posted by: technofutures | October 27, 2008

Cambridge Michaelmas Week 2

It’s been a very busy week, with lots of work on my project, but also a little bit of time for some more interesting experiences: discovering Geocaching, the ‘Murder on Air’ play yesterday…


Now by the end of this paragraph, you will probably think that I am totally crazy, but I blame Hugo for introducing me to this amazingly amusing ‘treasure hunt’-style activity.

Basically, geocaching involves ‘running around’ towns/cities looking for clues that answer a set of questions. All the clues that you find help you to calculate a GPS coordinate which corresponds to the location of a hidden ‘treasure’ somewhere in the local area. I was very excited to see that several ‘caches’ were located in Cambridge, so after lunch at the Graduate Union, I convinced Hugo (who was equally as curious as I…) to join me to try out the ‘Cambridge Camouflage’ cache.

Our Holmesian detective skills would have probably been less put to test had Hugo’s GPS and PDA not threatened failure throughout the duration of our exciting search…  Once we had gathered all the clues (many of which were related to the best college in Cambridge) we calculated the location of the hiding place and went there.

PA210574Hugo behind Trinity pretending that he had a working GPS..

Unfortunately, we couldn’t seem to find the hidden object! We looked everywhere around the supposed location, up trees, behind bushes, along riverbanks… but nothing. 40 minutes later, we almost gave up… until the hiding place and its contained object suddenly appeared to us… Very satisfying after an interesting search! We wrote our names on the ‘log of visitors’ contained in the ‘treasure’ as one is expected to do…

PA210576 PA210578

The cache and treasure were found at last!PA210579 I look forward to the next one.. there seem to be quite a few in Cambridge 🙂 If you’re looking for a fun afternoon… I recommend geocaching!

Cambridge, Cambridge and more Cambridge

Many people commented on my new stationery and folders this week, beautifully branded with a gold university logo. I think I’ll keep the source of these lovely items secret for a while longer… 🙂


And if anyone if wondering how my PhotoSynthing is going… Check out this ‘mash-up’ of my synths that I put together here!


‘Murder on Air’

long-MurderOnAir Another highlight of the week was Saturday’s theatre performance that I went to: “Murder on Air”, at the Cambridge Arts Theatre.

The play consisted of a recreation of the atmosphere of an old BBC radio set, commonly used for recording radio-novels. The actors read three of Christie’s radio plays in the style of their original BBC broadcasts – performed on stage in the authentic studio setting with some absolutely stunning live sound effects (how do you create the sound effect of someone being murdered with a nail in the forehead? how do you create the ambience of a large party?)

PA250587 PA250591

PA250593 PA250595

Hugh Fraser, Sue Holderness and Alexander Bermange (at the props)

It was absolutely amazing, in a very unique setting, with a fantastic group of actors and the added pleasure of the two guest stars performing: Sue Holderness and Hugh Fraser. Hugh Fraser (well-known because of his role as Captain Hastings in Christie’s Poirot TV Series) was stunning and I was very, very happy to see him acting in the flesh! And in addition… he was playing Poirot!


A lot of interesting stuff this week again…

Nanotechnology continued on from last week’s re-derivation of Schroedinger’s Equation, exploring topics such as potential wells, field emission, the quantum harmonic oscillator and the origin of electrical resistance. It was particularly interesting as we start to move on now from the revision of material we covered last year, although some aspects were rather hard/abstract e.g. phonons… and seem to be covered at lightning speed!

Photonics went on to describe methods of generating slightly more complex diffraction patterns, including the challenges associated with aspects such as the unwanted symmetries and wavelength dependence. The second lecture looked at a different aspect of optics, namely ‘optical correlation’ i.e. allowing us to compare two images (and input and a reference). Two simple methods were explored: the matched filter (brought back ‘lovely’ memories of last year’s digital filter course….) and the joint transform correlator. It was explained that even though the matched optical filter is much simpler to implement, it requires very good optics to align the input and reference images precisely… not easy to achieve!

Control Systems – I like the pace of these lectures. They are not too fast and the lecturer is very clear. We discussed concepts such as how the effect of disturbances and noise can be reduced in a control system. I think it’s the first time that Sensitivity (1/(1+L(s))) and Complementary Sensitivity (L/(1+L(s))) have been explained to me – it’s really quite simple. Sensitivity affects disturbances and C. sensitivity affects measurement noise! We also discussed the effects  of not having an exact mathematical representation of a control system plant, and how to get rid of non-linearities (the engineer’s nightmare…). I’m very much looking forward to next week’s lectures, deriving the Bode Gain-Phase formula, which is something completely new to me…

Technological Innovation was again very interesting, covering the types of innovation (e.g. using Schumpeterian S-curves), using as a case study process innovation as opposed to ‘product innovation’ which we covered last week. We were also fortunate to have Gerald Avison from TTP group as our guest speaker this week, demonstrating, amongst other things, a new inkjet can printing technique. Really quite impressive how sharp the digital images on the can were.

– Statistical Pattern Processing: Hmm… what can I say… The content seems to be very relevant to my project, and the lecturer is certainly enthusiastic, but boy is it hard! Granted though, it wasn’t as gruesome as last week’s Expectation maximisation. This week introduced us to the concept of single and multi-layer perceptrons, which are another approach to classification of data sets. Fundamentally, they’re not complicated, but I seem to get somewhat lost in probabilities (priors, posteriors, likelihoods…) and funny symbols. The section on Discriminant analysis though was very interesting, given that I had tried to implement that last week in my project…

Best lecture of the week: Not easy… most were equally good. The best was probably the Control Systems lecture on the detailed effect of feedback, not because the content was particularly exciting, but because the lecturing style was very clear and understandable!


Lots of developments made on the project this week. I think I probably went into too much detail last week, so I’ll provide a more concise summary this week:

  • After last week’s concerns, I finally managed to get my canonical correlations algorithm working for four different classes of objects (shoe, tea boxes, biscuit boxes and a lego block)
  • The meeting with my supervisor on Tuesday was very fruitful and he pointed me into lots of possible directions.
  • We discussed the concept of reducing dimensionality in the subspaces that I am training (see last week’s post). From what I understand it’s mainly a delicate balance between computational speed and recognition accuracy.
  • I modified my algorithm to be able to train 1000 classes of objects! Then I supply up to 36 new test images (all of the same object) to my system, and if all works out the algorithm will ‘recognise the object’.
  • In fact it does recognise the object… all too well in fact. I’m concerned that my system won’t generalise very well to images that are obtained from environments that are not as artificial as those from the Amsterdam Library of Object Images.

PrincipalComponents An example of what I’m working with!

  • I then played with several factors which I thought would affect my canonical correlation training system: occlusions, background of images, dimensionality of obtained subspaces… Up to now, my conclusion is that CC are a good system, with good robustness to small variations in the mentioned factors… I could go on for hours with graphs and analyses…
  • We also discussed other types of algorithms which involve mean subtraction, or modelling my image training sets as Gaussians (or Mixtures of Gaussians), and using the Kullback-Leibler Divergence measure to classify test images. I hope to try to implement those in the next few days..

Anyway, I don’t have time to write more… and it feels like a duplication of my log-book!

The most amusing part this week was the use of ‘pi’ as a variable in one of the papers I was reading today i.e. it was not equal to 3.142. According to that paper, summing the values in the ‘pi’ variables was equal to 1. Some people are strange…!


If anyone is interested in some of the matlab code that I have written, leave a comment and I’ll post some of it…

That’s all for this week!



Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s


%d bloggers like this: