Adaptive Memory Network for Faster Text QA Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory models, such as those based on long short-term memory (LSTMs), require a large number of parameters and do not generalize well for long sequence inputs, leading to poor inference times and high computation costs due to the need to examine all input facts or their intermediate representations.
Innovation Solution
Adaptive Memory Networks (AMN) dynamically construct a memory network architecture with variable numbers of memory banks weighted by question relevance, allowing the decoder to select only the highest relevance entities for answering, thereby reducing the number of banks used and achieving faster inference times.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional memory models (LSTMs) are used to process long sequence inputs, then the model can capture temporal dependencies, but the inference time increases and computation costs increase due to examining all input facts
Solution Approach 1:
The patent segments the memory structure into multiple memory banks, each storing different aspects or types of facts from the input sequence. This allows the system to organize information in a structured way that enables selective access during inference, reducing the need to examine all input facts while maintaining temporal dependency capture capabilities.
Solution Approach 2:
The patent introduces dynamic memory allocation where the number of memory banks and their contents are adjusted based on the input sequence characteristics. This dynamic adaptation allows the system to optimize the balance between capturing temporal dependencies and reducing inference time by allocating memory resources more efficiently for each specific input.
2Loss of information
If traditional memory models examine all input facts, then comprehensive information is processed, but computation costs increase
Solution Approach 1:
The patent applies local quality by organizing memory banks with different levels of detail or specificity. Not all memory banks store the same type of information, allowing the system to selectively access only the relevant memory banks for a given query, thereby reducing computation costs while maintaining information completeness where needed.
Solution Approach 2:
The patent performs preliminary organization of input facts into memory banks during the encoding phase, structuring the information in advance. This preliminary action enables the decoding phase to quickly access only the necessary information without having to process all input facts, thus reducing computation costs while preserving comprehensive information when required.
3Reliability
If the decoder analyzes all entities in the memory network, then comprehensive answers are generated, but inference time increases
Solution Approach 1:
The patent applies partial action by enabling the decoder to selectively access only the necessary memory banks based on the query and input characteristics. Instead of analyzing all entities in the memory network, the system can focus on specific memory banks that contain the most relevant information, thereby reducing inference time while maintaining answer accuracy through targeted information retrieval.
Data Source
AI summary
A method for performing question answer (QA) tasks that includes entering an input into an encoder portion of an adaptive memory network, wherein the encoder portion parses the input into entities of text for arrangement of memory banks. A bank controller of the adaptive memory network organizes the entities into progressively weighted banks within the arrangement of memory banks. The arrangement of memory banks may be arranged to have an initial memory bank having lowest relevance for lowest relevance entities being closest to the encoder, and a final memory bank having a highest relevance for highest relevance entities being closes to a decoder. The method may continue with inferring an answer for the question answer (QA) task with the decoder analyzing only the highest relevance entities in the final memory bank.