Machine Learning Meetings and Events
Group Meetings: Group meetings are held Mondays from 11am to Noon (talk starts 11:10am) in D.L. Pratt 290C unless otherwise noted. Meetings are coordinated by Hugo Larochelle.
Tea Talks: Tea talks are held every Wednesday at 4:00pm in D.L. Pratt 290C. Talks should be simple, accessible, and not exceed 15 minutes. Speakers bring snacks, make tea, and provide a copy of the presented paper.
Group Meeting Nov 24, 2008: Deep Belief Nets for Visual Object Recognition: Part II
- Speaker: Vinod Nair
- Abstract:
NOTE: This is an updated version of the tea-time talk I gave a few weeks ago. Since then we’ve beaten the state-of-the-art results for NORB. This time I’ll try to better explain the details of what we did using proper slides, diagrams, equations etc. The new abstract is given below.
This talk introduces a new Deep Belief Net model and evaluates it on a 3D object recognition task. The three main novel contributions are: 1) a top-level model that represents the joint distribution as a third-order Boltzmann machine, 2) a hybrid, generative/discriminative algorithm for learning the top-level model, and 3) a regularizer that encourages sparse binary latent representations when learning the lower layers of the model. We evaluate performance on the NORB database, which contains stereo-pair images of five object classes under different lighting conditions and viewpoints. Our model, using no prior knowledge about invariances, achieves 6.5% error on the test set, which is close to the best published result for NORB (5.9%) using a convolutional neural net that has some knowledge of translation invariance wired in. It substantially outperforms shallow models such as SVMs (11.6%). DBNs are especially suited for semi-supervised learning, and to demonstrate this we consider a modified version of the NORB recognition task in which additional unlabeled images are created by applying small pixel translations to the images in the database. With the extra unlabeled data (and the same amount of labeled data as before), our model achieves 5.3% error, making it the current best result for NORB.
This is joint work with Geoff Hinton.