CPS 296.1 Bayesian games and their use in auctions Vincent Conitzer [email protected].
What We Did Not Cover - cs.duke.edu · What We Did Not Cover 1 Much More Detail 2 Statistical...
Transcript of What We Did Not Cover - cs.duke.edu · What We Did Not Cover 1 Much More Detail 2 Statistical...
What We Did Not Cover
COMPSCI 371D — Machine Learning
COMPSCI 371D — Machine Learning What We Did Not Cover 1 / 11
What We Did Not Cover
1 Much More Detail
2 Statistical Machine Learning
3 Other Supervised Techniques
4 Reducing the Burden of Labeling
5 Unsupervised Methods
6 Addressing Multiple Learning Tasks Together
7 Prediction over Time
COMPSCI 371D — Machine Learning What We Did Not Cover 2 / 11
Much More Detail
Much More Detail
• Computationally efficient training algorithms:Optimization techniques• Deep learning architectures for special problems:
Image motion analysis, video analysis, ...
COMPSCI 371D — Machine Learning What We Did Not Cover 3 / 11
Much More Detail
Beyond Discriminative Neural Networks
• Abstraction for its own sake: Auto-encoders• A game-theoretical technique to draw from a distribution:
Generative Adversarial Networks
Which image is fake?
COMPSCI 371D — Machine Learning What We Did Not Cover 4 / 11
Much More Detail
Generative Adversarial Networks
Sampler
Generator
Discriminator
noise
Discriminator Loss
Generator Loss
real image
fake image
real-image database
random switch
• Discriminator guesses if input is real or fake• Discriminator loss penalizes wrong predictions• Generator loss penalizes correct predictions• After training keep only the generator
COMPSCI 371D — Machine Learning What We Did Not Cover 5 / 11
Statistical Machine Learning
Statistical Machine Learning
• How to measure the size of H: Vapnik-Chervonenkisdimension, Rademacher complexity• How large must T be to get an h that is within ε of a
performance target with probability greater than 1− δ:Probably Approximately Correct (PAC) learning• H is learnable if there exists a size of T that is large enough
for this goal to be achieved• Which Hs are learnable?• How large must S be to get a performance measure
accurate within ε: Concentration bounds, statisticalestimation theory, PAC-like techniques
COMPSCI 371D — Machine Learning What We Did Not Cover 6 / 11
Other Supervised Techniques
Other Supervised Techniques
• Support Vector Machines: How to maximize free spacearound decision boundariesMaximizes the margin, i.e., the distance between theboundary and the nearest training samples• Boosting: How to use many bad predictors to make one
good oneSimilar in principle to ensemble predictors, differentassumptions and techniques• Learning to rank
Example: Learning a better Google
COMPSCI 371D — Machine Learning What We Did Not Cover 7 / 11
Reducing the Burden of Labeling
Reducing the Burden of Labeling• Semi-supervised methods: Build models of the data x to
leverage sparse labels y• Domain adaptation: Train a classifier on source-domain
labeled data (x, y) and target-domain unlableled data x sothat is works well in the target domain
Vegas Building Detector
Vegas Image
Vegas Building Detector
Shanghai Image
Shanghai Building Detector
Shanghai Image
TRAIN
ADAPT
https://en.wikipedia.org
COMPSCI 371D — Machine Learning What We Did Not Cover 8 / 11
Unsupervised Methods
Unsupervised Methods• Dimensionality reduction:
Compressing X ⊆ Rd to X ′ ⊆ Rd ′with d ′ � d
• Principal or Independent Component Analysis (PCA, ICA)• Manifold learning, GANs
• Clustering:• K -means• Expectation-Maximization• Agglomerative methods• Splitting methods
COMPSCI 371D — Machine Learning What We Did Not Cover 9 / 11
Addressing Multiple Learning Tasks Together
Addressing Multiple Learning Tasks Together
• Multi-task learning: How to learn representations that arecommon to different but related prediction tasks
COMPSCI 371D — Machine Learning What We Did Not Cover 10 / 11
Prediction over Time
Prediction over Time• State-space methods
• Time series analysis• Stochastic state estimation• System identification
• Recurrent neural networks• Reinforcement learning: Actions over time
Learning policies underlying observed sequences
COMPSCI 371D — Machine Learning What We Did Not Cover 11 / 11