- Transformer_COSCO
- Multiple Multi_dimensional Knapsack problem using RL and Transformers
- Self-supervised Heterogeneous Graph Neural Network
- TGR: Task Scheduling using RL
- Gpu Cloudsim Plus
- A Novel Incremental Learning Method for Text Classification
- Font Recognition in OCR
Transformer_COSCO
Multiple Multi_dimensional Knapsack problem using RL and Transformers
HeCo: Self-supervised Heterogeneous Graph Neural Network
TGR: Task Scheduling using RL
Gpu Cloudsim Plus
Learn#: A Novel Incremental Learning Method for Text Classification
Paper implementation
This implementation was aimed at improving my experience with incremental text classification tasks and exploring the utilization of more than four AI models together.
The motivation behind this implementation stems from the limitations of text classification using deep learning. Deep learning models are typically trained in a batch learning setting where the entire training data is used. However, in a continual learning setting, it becomes cumbersome to store previous text data as it requires significant memory resources. Additionally, obtaining a sufficient number of labeled samples for text classification
with deep learning at the beginning is often challenging due to the high labeling cost or infrequent occurrence of relevant data.
Moreover, the performance of deep learning models heavily relies on the distribution of the data samples. When the distribution of newly collected data differs from that of the previous datasets, it may result in overfitting and affect the model’s accuracy. To address these challenges, I explored the use of incremental learning.
Incremental learning offers several benefits. Firstly, when new samples are added, there is no need to retrain the model with all the previous samples, reducing the computational burden. This makes the learning process more efficient and scalable.
The implementation consists of four key components. The first component is the Student model, which utilizes deep learning techniques to solve classification tasks.
The second component is a reinforcement learning (RL) module. RL is employed to selects the adaptive predictions from the Student models and feeds the results to the Teacher model.
The third component is the Teacher model, which provides the text categories with XGBoost.
Finally, the implementation incorporates a discriminator model, calculates the similarity of
each model, and then eliminates the similar Student models, thereby keeping the number of Student models steady
Through this implementation, I gained valuable insights into incremental learning, specifically in the context of text classification. It not only allowed me to overcome the limitations of deep learning in continual learning settings but also offered the opportunity to explore the synergistic utilization of multiple AI models.