Systems, apparatuses, and methods for deep learning of feature detectors with sparse coding
Detailed herein are embodiments of systems, methods, and apparatuses to be used for feature searching using an entry-based searching structure.
1. An apparatus comprising:
an entry-based search structure to perform a nearest neighbor search of a dictionary of features with entries dependent on features in the dictionary, wherein the entry-based search structure is a content addressable memory (CAM);
physical storage coupled to the entry-based search structure to store the features; and
physical logic to access the entry-based search structure to search for an entry for particular feature stored in the dictionary of features based upon a received input instance and to access the physical storage and retrieve the particular feature, wherein the searchable features of the dictionary and the input instance are non-negative and the search to select a feature with a highest correlation to a residual vector, approximate at least one coefficient of the selected feature using non-negative least squares, and compute a revised residual vector by removing the effect of the selected feature from the residual vector.
2. The apparatus of claim 1 , wherein the physical storage is a random access memory (RAM).
3. The apparatus of claim 2 , wherein the RAM is embedded into a central processing unit of the apparatus.
4. The apparatus of claim 2 , wherein the RAM is external to a central processing unit of the apparatus.
5. The apparatus of claim 1 , wherein each entry of the entry-based search structure is to store an interval value and a value of a feature.
6. The apparatus of claim 1 , wherein the dictionary of features is to describe a plurality of bounding boxes defined by interval values and wherein each bounding box is to contain a single feature.
7. The apparatus of claim 1 , wherein the residual vector comprises data elements of an input vector.
8. The apparatus of claim 1 , wherein the feature with the highest positive correlation with the residual vector is a vector with a smallest dot product with the residual.
9. A method comprising:
selecting a first feature with a highest positive correlation with a residual vector from a content addressable memory (CAM), wherein the feature with the highest positive correlation with the residual vector is a vector with a smallest dot product with the residual;
approximating coefficients of the selected first feature;
generating a revised residual vector by removing its orthogonal projection on a space spanned by the first feature;
storing the selected first feature.
10. The method of claim 9 , wherein the residual vector comprises only non-negative data.
11. The method of claim 9 , wherein the received data vector comprises only non-negative data.
12. The method of claim 9 , wherein the selecting comprises a parallel search of features of the dictionary.
13. The method of claim 9 , further comprising:
normalizing the revised residual vector.
14. The method of claim 9 , further comprising:
initializing a residual vector from a received data vector.
15. The method of claim 9 , further comprising:
selecting a second feature with a highest positive correlation with the revised residual vector;
approximating coefficients of the selected second feature; and
storing the selected second feature.