Continual Image Classifier Learning for Hard Images on Mobile Clients
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image classification methods face challenges in accurately classifying hard images due to environmental variables, require extensive training data, incur high computing costs, and suffer from catastrophic forgetting, especially on mobile devices with limited resources.
Innovation Solution
A method utilizing a continual learning server that employs an Adversarial Autoencoder to process unclassifiable images, generating training data sets for updating a classifier on a client device by selectively expanding or adding neural network layers, reducing computing power requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional continual learning is applied to a neural network on a server to accurately classify hard images, then classification accuracy is improved, but computing power requirements and training costs increase significantly
Solution Approach 1:
The system segments the continual learning process into two parts: complex model training and parameter generation is performed on a server with high computing power, while only lightweight parameter updates are transmitted to and applied on client devices with limited resources. This division allows accurate classification to be achieved without requiring full training capacity on resource-constrained devices.
Solution Approach 2:
The server acts as an intermediary that processes hard images through the AAE model to generate optimized parameter updates. These parameters serve as a mediator that transfers the benefits of extensive training from the server to client devices, enabling accurate hard image classification on clients without requiring them to perform the computationally intensive training themselves.
2Measurement precision
If network capacity is expanded to accommodate more training data for hard image classification, then classification performance is improved, but device size and weight increase making it difficult to mount on mobile devices
Solution Approach 1:
The system extracts only the essential parameter updates from the full training process and transmits them to client devices. Instead of transferring entire trained models or large amounts of training data that would increase device size, only the necessary parameter changes are extracted and applied, maintaining performance improvement while minimizing device size impact.
Solution Approach 2:
The solution transitions from a local approach where each device must store and process extensive training data locally (increasing device size) to a cloud-based approach where training data and model processing occur on remote servers. The parameter updates serve as a compact representation that bridges these dimensions, enabling performance improvement without local storage expansion.
3Adaptability or versatility
If conventional continual learning updates the entire classifier model, then new classification tasks are learned, but previously learned information is lost due to catastrophic forgetting
Solution Approach 1:
The system incorporates feedback mechanisms where the server evaluates the impact of parameter updates on both new and existing classification tasks. By monitoring performance on previously learned tasks during the parameter update process, the system can adjust update strategies to minimize catastrophic forgetting while still adapting to new tasks.
Solution Approach 2:
The server performs preliminary evaluation and optimization of parameter updates before transmitting them to client devices. This preliminary action includes assessing potential impacts on previously learned tasks and adjusting the parameter updates accordingly, preventing catastrophic forgetting before it occurs on the client side while enabling adaptation to new tasks.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for performing continual learning on a classifier, in a client, capable of classifying images by using a continual learning server is provided. The method includes steps of: a continual learning server (a) inputting first hard images from a first classifier of a client into an Adversarial Autoencoder, to allow an encoder to output latent vectors from the first hard images, allow a decoder to output reconstructed images from the latent vectors, and allow a discriminator and a second classifier to output attribute and classification information to determine second hard images to be stored in a first training data set, and generating augmented images to be stored in a second training data set by adjusting the latent vectors of the reconstructed images determined not as the second hard images; (b) continual learning a third classifier corresponding to the first classifier; and (c) transmitting updated parameters to the client.