1

[zhjcxy]

is this Transformer²？

[codelion]

Hi, adaptive classifiers is an alternative approach for a similar goal that is much simpler.

We use :

• Transformer Embeddings to represent new examples
• Prototype Memory: Maintains class prototypes for quick adaptation to new examples
• Adaptive Neural Layer: Learns refined decision boundaries through continuous training

While transformerr² uses:

• Prompt-based adaptation.
• Classifier-based adaptation. A task classifier trained with SVF identifies the task during inference and selects the appropriate z-vector.
• Few-shot adaptation. Combines multiple pre-trained z-vectors through weighted interpolation. A simple optimization algorithm tunes these weights based on performance on a few-shot evaluation set.

Our focus is only on classification and we work with any existing base classifier, while transformer² is a new architecture.