Why we don't let model to learn by itself? I mean, speaking in the context of RL, the agent mainly relies on interaction to learn any model of its surrounding rather than using labelling and some notation of supervised learning. I hope I'm not talking off the topic you asked, however, I feel the internal model (say latent model) of the RL agent, regardless of the method we are using for training, can be counted as World Model. Although, it might not be in the form of an explicit dynamical model. Looking forward to see others point of view.

The aim of WORLD MODEL research extends beyond having large models memorize correct answers. Instead, it focuses on creating models that possess a deeper understanding of the world and can apply generalized knowledge to perform tasks intelligently. This approach involves capturing underlying patterns, learning from diverse datasets, and cultivating a more comprehensive grasp of contextual information, steering away from simple memorization of specific responses.

For me there are several techniques that come to mind to train models with limited labeled data, these are some ideas:

1. Active Learning: This approach involves iteratively selecting the most informative data points for labeling, prioritizing uncertain predictions by the model. By focusing labeling efforts on these points, the model's learning efficiency is enhanced[1].

2. Semi-Supervised Learning: This technique combines labeled and unlabeled data for training, leveraging a broader dataset effectively. For instance, in tasks like facial recognition, models can learn general features from a large pool of unlabeled data and fine-tune with a smaller set of labeled data[2]. Added support for semi-supervised learning can come from an attention mechanism [7].

3. Self-Supervised Learning: By utilizing self-supervised learning methods, it is possible to reduce the need for a large volume of labeled examples while maintaining performance levels in tasks like medical image analysis[4]. Like Semi-Supervised learning Attention Mechanism can aid with self learning [8]

4. Human-in-the-loop: combine human-computer-interaction for helping with labeling of when self learning in ambiguous for the learning model.

5. Synthetic data: from the labeled data already collected, automatically extend each labeled set with synthetic data versions [9]. However, enough data is need enough for each label is required so the full distribution of the set is captured in the synthesised data

These strategies offer viable ways to optimise the use of labeled data in machine learning tasks, allowing for more efficient model training and improved performance without requiring extensive amounts of labeled data upfront.

Citations:

[1] 5 Ways to Improve The Quality of Labeled Data https://encord.com/blog/improve-quality-of-labeled-data-guide/

[2] How to Build Good AI Solutions When Data Is Scarce https://sloanreview.mit.edu/article/how-to-build-good-ai-solutions-when-data-is-scarce/

[3] Improving Human-Labeled Data through Dynamic Automatic Conflict Resolution https://machinelearning.apple.com/research/improving-human-labeled-data

[4] Self-supervised Learning as a Means to Reduce the Need for Labeled Data in Medical Image Analysis - arXiv https://arxiv.org/pdf/2206.00344.pdf

[5] What Is Data Labelling and How to Do It Efficiently [2023] https://www.v7labs.com/blog/data-labeling-guide

[6]Human-in-the-loop machine learning: a state of the art - Artificial Intelligence Review https://link.springer.com/article/10.1007/s10462-022-10246-w

[7]Semi-Automated Data Labeling http://proceedings.mlr.press/v133/desmond21a/desmond21a.pdf

[8] [PDF] Self-supervised Learning as a Means to Reduce the Need for Labeled Data in Medical Image Analysis - arXiv https://arxiv.org/pdf/2206.00344.pdf

[9]Synthetic Data for Machine Learning: its Nature, Types, and https://www.altexsoft.com/blog/synthetic-data-generation/