Artificial intelligence apparatus using a plurality of output layers and method for same
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models face challenges in achieving high accuracy with small amounts of learning data and are prone to overfitting, especially when applied to robot cleaners that need to escape from restraint situations.
Innovation Solution
An AI apparatus with a dual learning model architecture, where a pre-training model is connected to multiple output layers learned through transfer learning, with weights adjusted based on brightness and time information, and the ability to generate new output layers and reset weights to improve classification accuracy and escape route determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If transfer learning is applied to learn only the last classification layer with small amounts of new data, then the time cost and computational quantity are reduced, but the result value becomes overfitted due to being focused on the relevant data
Solution Approach 1:
The patent segments the classification layers into two groups: frozen layers (excluding the last classification layer) and a retrained last classification layer. This segmentation allows the model to benefit from transfer learning while preventing overfitting by maintaining the generalization capabilities of the frozen layers and only adapting the final layer to new data.
Solution Approach 2:
The patent applies different treatment to different parts of the neural network: the frozen layers retain their pre-trained weights for maintaining generalization, while only the last classification layer is retrained on new data. This local differentiation optimizes both time efficiency and classification accuracy.
2Measurement precision
If the weight of each classification layer is learned through data in deep learning, then the classification accuracy can be improved, but it takes a lot of time to learn the weight of all the classification layers
Solution Approach 1:
The patent utilizes pre-trained models where the weights of classification layers have already been learned from large datasets beforehand. This preliminary action allows the system to achieve high classification accuracy without retraining all layers, significantly reducing learning time while maintaining performance.
Solution Approach 2:
The patent copies pre-trained weights from existing models and applies them to the new classification task. By copying the learned representations from pre-trained layers and only fine-tuning the final layer, the system achieves accurate classification without the time cost of training all layers from scratch.
Data Source
AI summary
The present disclosure discloses an artificial intelligence apparatus including an input interface configured to obtain input data, a sensing interface configured to obtain environment information, and one or more processors configured to classify an object by inputting the input data obtained from the input interface to an artificial intelligence model, in which the artificial intelligence model uses a first learning model and a second learning model which is connected with the first learning model and includes a plurality of output layers to respectively assign weights to the respective result values output by the plurality of output layers and combine the respective result values to which the weights are assigned to derive the final result.