Hamming Space Locality Preserving Neural Hashing for Similarity Search
■ Nearest neighbor search is one of the most common information retrieval techniques in fields such as image retrieval, face recognition, question answering, text search and more. In large retrieval databases the search in the feature representation space often requires significant computation and memory resources, and imposes a performance bottleneck. As data volumes heavily grow and content search becomes an increasingly required task, methods for fast Approximate Nearest Neighbor (ANN) search, which trades off a slight loss in accuracy for large performance gains, have become the focus of extensive research [10, 2].
■ One main group of methods for ANN is based on binary hashing, which maps data points in the original feature vector representation space into binary codes in the hamming space [2], for compact representation and fast search. We present a neural network method for learning data-dependent, binary hashing that preserve local similarity, by introducing a novel combination of a loss function and sampling method. For clarity, we keep our focus on a hash code ranking strategy, that searches a distinct representation for each data point in the database, rather than hash table lookup, that maps data points into buckets to reduce the number of distance computations [10]. As opposed to previous studies in this field, which report improvement in accuracy only over small code sizes (up to 128 bits), we present results on both small and large code lengths (768 bits), offering flexibility in choice of strategy and resources vs. accuracy. No pre-computations or graphs are required for the training process.
■ The full paper will review the effects of model architecture on accuracy vs. inference speed, present use-cases on million and billion-scale datasets using a dedicated hardware accelerator1 and review further findings on datasets, performance metrics, effects of data distribution and method comparison.
|
|
|
|
[ image retrieval ][ face recognition ][ question answering ][ text search ] |
|
|
FAQ |
|
|
|
|
Please see the document for details |
|
|
|
|
|
|
|
|
|
|
|
English Chinese Chinese and English Japanese |
|
|
2020/9/1 |
|
|
|
|
|
|
|
|
295 KB |
- +1 Like
- Add to Favorites
Recommend
All reproduced articles on this site are for the purpose of conveying more information and clearly indicate the source. If media or individuals who do not want to be reproduced can contact us, which will be deleted.
Integrated Circuits
Discrete Components
Connectors & Structural Components
Assembly UnitModules & Accessories
Power Supplies & Power Modules
Electronic Materials
Instrumentation & Test Kit
Electrical Tools & Materials
Mechatronics
Processing & Customization
