Method and system for personalising machine learning models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models, particularly conditional neural networks, are inefficient for on-device deployment due to increased model size and latency caused by concatenating conditioning vectors, making them unsuitable for resource-constrained devices like smartphones and IoT devices.
Innovation Solution
A method involving a first neural network that transforms weights of a second neural network using conditioning vectors during training, eliminating the need for concatenation, resulting in personalized weights that reduce model size and latency, suitable for resource-constrained devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conditioning vectors are concatenated to input features or used for modulation in conditional neural networks, then personalized processing capability is improved, but model size increases dramatically
Solution Approach 1:
The patent segments the conditional processing into two separate networks: a first neural network that processes the conditioning vector to generate personalized weights, and a second neural network that uses these weights to process input data. This segmentation avoids concatenating conditioning vectors with input features, thereby preventing model size inflation while maintaining personalized processing capability.
Solution Approach 2:
The patent applies preliminary action by pre-processing the conditioning vector through the first neural network to generate personalized weights before the actual data processing occurs. This preliminary transformation of the conditioning vector into weight parameters eliminates the need for runtime concatenation operations that would increase model size.
2Adaptability or versatility
If conditioning vectors are concatenated or used for modulation, then personalized processing is achieved, but latency increases
Solution Approach 1:
By segmenting the conditional processing into weight generation (first network) and data processing (second network), the patent enables efficient inference where the conditioning vector is transformed into weights once, and then the second network processes input data using these pre-computed weights without additional concatenation overhead, reducing latency.
Solution Approach 2:
The first neural network performs preliminary transformation of the conditioning vector into personalized weights before inference. This preliminary action allows the second network to directly use these weights during inference without performing additional concatenation or modulation operations, thereby reducing computational latency.
3Productivity
If large neural networks with many parameters are used for complex tasks, then processing capability is improved, but computational resources required increase
Solution Approach 1:
The patent introduces dynamics by making the neural network weights adaptive and personalized based on the conditioning vector. The first neural network dynamically generates personalized weights for the second network based on individual user characteristics, allowing the system to achieve high processing capability for complex tasks while using smaller, more efficient network architectures that adapt to specific users rather than requiring large fixed-size networks.
Data Source
AI summary
Broadly speaking, embodiments of the present techniques provide a method and system for personalising machine learning models on resource-constrained devices by using conditional neural networks. In particular, the present techniques allow for resource-efficient use of a conditioning vector by incorporating the conditioning vector into weights learned during training. This reduces the computational resources required at inference time.