Michael's Wiki

FIXME Somehow similar to r-cnns?

YOLO

YOLO methods apply a direct discriminative classification approach to object detection.

Large-Scale Unsupervised Learning

Attempts to replicate human brain behavior and structure in neural networks are heavily inspired by neuroscience. These efforts are characterized by an emphasis on very large networks (inspired by the massive $~10^{11}$ neurons in a human brain) and unsupervised processing of larger unorganized data sets1).

Whereas neural networks are all historically inspired by biology, these methods are more closely derived from existing biological neuroscience concepts.

Local Receptive Fields

In autoencoders, each feature is restricted to connect to a small subset of the lower layer. Inspired by biological neurons, local receptive fields support sparse coding and help enable many unsupervised methods and utilization of greater parallelization and GPU processing2) e.g. text classification3).

Pooling

FIXME

Local Contrast Normalization

FIXME

Feature Learning

Feature Learning refers to a family of techniques which learn to replicate, approximate, or otherwise represent an input distribution as efficiently as possible. There are analogies and historic influence from compression, PCA, and Sparse Coding. Feature learning techniques include: