EE32b: Projects and Datasets

Project (due on March 10, proposal posting by Feb 21, noon)

The students will complete an experimental project in the general area spanned by the syllabus of this class. Ideas for pojects: A typical project will take 20-40 hours and will either propose and test a new idea, or implement and explore an idea/algorithm from the literature. Teams of two students per project are encouraged.
The students will complete a report on the project. The report is due on March 10. Each student will write her/his report individually. The report will be delivered either on paper or as a web page/site. It will typically consist of the following sections:
  1. Cover page: Name of author, name of team-mate, project title, project abstract (150 words or less summarizing the aim of the project, the main idea(s) and the main findings). Possibly a pretty picture that exemplifies the aims/results of the project.
  2. Introduction: What is the aim/goal of the project, why is this aim important (or why it is not important), what are the possible approaches to achieve the goal, which approach was chosen and why.
  3. Technical approach: Detailed description of the technical ideas.
  4. Datasets used: What do the data look like, how many samples, who collected them, ...
  5. Experiments: For each experiment: what was the aim of the experiment, how was it conducted, what are the main findings. Put plenty of pictures, plots...
  6. Discussion: What does one learn from the experiments? Did the proposed technique work as well as expected?
  7. Conclusions: What are the main questions that are left open? Other experiments you would do next?

Ideas for projects

  1. Digital watermarking (either images or music)

    2. Design and implement and algorithm that allows you to add approximately 100bytes of information (typically text) to either any image or any piece of music (discrete data of course). You should also of course design and implement an algorithm that, given either an image or a sound, is able to recover the watermark.  Of course the process should not make the image/sound different for a human observer.
    The ideal watermark is robust with respect to: cropping, resampling, adding small amounts of noise, compression/decompression and all other typical manipulations that do not alter the image/sound percieavably.
    Variations on the theme: work on other types of data that are typically exchanged over the internet: text files, streaming images/sound ...
     
  2. Analysis and synthesis of birdsong

    3. Analize zebra finch songs coming from the Konishi laboratory.
    ideas for projects: Devise a method for describing each chirp (e.g. as an oriented line in time-frequency). Design and implement a method (manual or automatic, start from manual) for calculating the parameters of each chirp. Perform unsupervised clustering (e.g. using the k-means algorithm - see Matlab command kmeans) to discover how many types of chirps there are. Collect statistics on the frequency and order of chirps and study variability of songs of individual birds and different birds. Write code that produces synthetic songs - see if birds are fooled.
     
  3. Model of the cochlea
  4. Nonlinear dimensionality reduction
  5. Analysis and synthesis of handwritten signatures
  6. Reverse-engineering of music. Write computer program that takes a simple musical piece and extracts the score (notes, durations, tempo, stresses).


Each project team should write up a description of their project and post it to the class newsgroup by Wed, Feb. 21 at noon.
 

Datasets

HANDWRITING (SIGNATURES)
Collected by Mario E. Munich at Caltech.
Each file name is the concatenation of the string `s' and two 3-digit numbers. The first number is the identity of the person who signed and the second number identifies the signature. For example: s022001 is the 2nd signature given by subject 22 (the first signature is s022000).
The format of each file is simple: the first line indicates the number of samples in the signature; the following lines contain (X,Y) pairs for each of the samples in order of time acquired. I.e.:
N
X1 Y1
X2 Y2
...
XN YN
The `y' coordinate is inverted, i.e. it increases top-down rather than bottom-up.
A simple Matlab function for reading and displaying the contents of a signature file is provided in the same directory (display_signature.m).
These signatures were collected at 60 samples per second.
 

HUMAN VOICE
Dataset collected by various people. Used for the Mus Silicium competition (you will find a few samples in a local directory).
The format is `.wav' which Matlab can read. (type `help wavread' in Matlab).

MATLAB SOUNDS
Run the command `which sound' in Matlab. This will give you the path of the file `sound.m'. In the same directory you will find six files that end in `.mat' (chirp.mat, train.mat, ...). Those files contain sounds. You may load them into Matlab; type `help sound' to find out more.

STOCKMARKET
Many sources. For example the NASDAQ web page. Pick a stock, obtain its time-chart and click on it (example).

BIRDSONGS
Songs from zebra finches collected by Anthony Leonardo in the Konishi Lab at Caltech. Load the file `Songdata-20k.mat' into your Matlab. Type `who' to list six vectors whose name starts with `i'. The vectors contain 3 songs each for two birds. They are sampled at 20KHz.

Also check out the Cornell Birdsong web page. It includes a tutorial on spectrum analysis, and birdsong analysis software. You will find lots of interesting data including songs from whales.

ROCK MUSIC
Search the web and download your favorite piece. Take only a 3-10s segment.
If you hate rock then take classical. Any music will do.
You may want to compare the Fourier spectrum of different instruments. E,g. compare a percussion with a wind.