FINAL PROJECT TOPICS

You can pick one of the following topics, modify one topic, or propose an entirely different topic. Send email with your proposal to yann and sumit ( [ a t ] cs.nyu.edu) with you topic proposal.

• Adaptive Signal Processing: train a convolutional network (a Time-Delay Neural Net) to reconstruct an analog signal from multiple distorted digital versions of the signal. (talk to YLC for details).
• Biological Image Analysis: train a convolutional network to detect and classify organisms in microscopic images. Organism must be classified as healthy or sick from pollution depending on their shape (talk to YLC for details).
• Finance: use various learning methods to predict the likelihood that a company will go bankrupt (real data available. talk to YLC for details).
• Astronomy: Learn a parameterized 3D model of galaxies. Use that model to infer the 3D structure of a galaxy from a multi-spectral 2D image. A technique similar to that described in Blanz and Vetter for human faces may be used.

  Blanz, V. and T. Vetter: A Morphable Model for the Synthesis of 3D Faces. SIGGRAPH'99 Conference Proceedings, 187-194 (1999). available here.

• Unsupervised image modeling: train an unsupervised energy-based model on video sequences so that it produces a set of local feature detectors from which the image can be reconstructed (ask YLC for details).
• Oversampling photos or videos: take of bunch of high-resolution/hihg-quality photos or video frames. Subsample them (e.g. make them half the original size) and compress them (aggressively) with JPEG. Train a neural net (e.g. a convolutional net) to reconstruct 2x2 patches of the high-qual/high-res image from the corresponding pixel and its surroundings (e.g. 11x11 pixels) in the compressed/low-res image.
• Pruned Mean-Square SVM: build an SVM-like architecture: one layer of gaussian units (with one unit per training sample), followed by a linear classifier (or one per class for multiclass classification). Train the weights with means squared error (and/or logistic regression). Prune the weights using the "Optimal Brain Damage" technique (see http://books.nips.cc/papers/files/nips02/0598.djvu ). Retrain, Re-prune... Test this method on the MNIST dataset of handwritten digits ( http://yann.lecun.com/exdb/mnist/ ).
• Siamese Nets for Handwriting Recognition: train two identical copies of a neural net (with identical weights) to produce similar output vectors when they are presented with images of a characters from the same category, and to produce different outputs when they are presented with characters from two different cetegories. Use the MNIST dataset available from here.
• first come, first served. More project topics to come....

Other project ideas that have been proposed in the past: training an edge detector from manually annotated images, training a machine to predict keystrokes, training a system to predict the orientation of a digital photos, ....