Hierarchical Signal Quantization for Similarity Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional hashing methods are inefficient for indexing data such as time series, sound, and images, as they fail to capture similarity between non-identical data points, making it difficult to retrieve similar data from a database.
Innovation Solution
A hierarchical method for representing input signals as compressed binary sequences using a hash function that adapts sensitivity parameters to distinguish between similar and dissimilar signals, allowing for efficient comparison and reconstruction of signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional hashing is used to generate compact binary representations, then search efficiency is improved, but the ability to retrieve similar but non-identical data is lost
Solution Approach 1:
The hash output is segmented into multiple levels or layers, where each level captures different aspects of data similarity. Instead of a single hash value, the system generates a hierarchical structure of hash values that progressively refine the similarity measurement, allowing both efficient filtering and accurate similarity retrieval.
Solution Approach 2:
The patent transitions from conventional single-dimension hashing to multi-dimensional hierarchical hashing. By adding temporal or hierarchical dimensions to the hash structure, the system can represent similarity relationships in multiple layers, enabling retrieval of similar data while maintaining search efficiency.
2Device complexity
If all hash bits provide the same information, then the hashing process is simple, but the ability to distinguish between similar and dissimilar data at different levels is reduced
Solution Approach 1:
Different bits or segments of the hash output are assigned different qualities or weights based on their discriminatory power. Certain hash bits that provide more meaningful similarity information are identified and given higher importance, while less informative bits are de-emphasized, creating a non-uniform information distribution that improves precision without significantly increasing complexity.
Solution Approach 2:
The patent introduces variable parameters into the hashing process, such as different sensitivity thresholds or weighting factors for different hash bits. By dynamically adjusting these parameters based on the data characteristics, the system can optimize the balance between simplicity and precision for different similarity discrimination tasks.
Data Source
AI summary
An inner product if formed from an input signal and a hashing vector, to which a dither scalar is added. A binary digit is produced according to a quantization function, wherein the binary digit is zero or one, and wherein the quantization function is non-monotonic and subject to a sensitivity parameter. The steps are iterated while decreasing the sensitivity parameter until a termination condition is reached.