- How neural end-to-end models differ
- Components of neural psychometric models
How neural end-to-end models differ
Earlier methods we have considered for scoring either take a prompting approach, which uses LLMs without updating its parameters to score responses, or taken frozen embeddings as input to classical machine learning (ML) models.
In this section we consider models that are neural end-to-end. The key distinction we make here is that end-to-end transformer architectures means that all parameters are differentiable from input to output and optimized by gradient descent.
This means all model components are trained to work together, which often leads to better performance. The models are built around transformers (I.e., encoders or decoders) and add special heads for predicting item characteristics and test scores.
Models can be more complex than earlier models discussed because they may involve pre-training or at least fine-tuning or RLHF. Models are usually built from a standard set of modular components.
Components of neural psychometric models
The following components are mixed and matched depending on task demands.
- Multi-head self-attention are included in end-to-end neural transformer models. Self attention is the hallmark of transformer models .
- Next are multilayer perceptrons (MLPs, aka feed forward networks: FFNs). The MLP first expands the representation to a higher-dimensional space, applies a non-linearity, and then contracts it back to the original dimension.
- Normalization and residual connections, discussed earlier in the How LLM’s learn section, are added to stabilize deep learning training.
- Sometimes models include adapters and LoRA or QLoRA for parameter efficient fine-tuning (PEFT) discussed in part 2.
Next page
Scoring (3 of 3): Neural contrastive pairwise regression (NCPR)
Previous page
Return home
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).