Blogposts on embeddings analysis

AI infused societies

Reviewing recent papers on the topic of retraining models one often hear on possible misconceptions on algorithms biases. When we speak about algorithms in general or inherited biases we are often referring to these algorithms using statistics behind their algorithmic skin. From these statistical traces we are then taking the set of parameters and what then is used to produce predictions of these algorithms. So, when people refer to the potential bias this may be decreased when taking statistical results keeping in mind potential accuracy, recall and other characteristics of these algorithms.

Another approach is to use the geometrical characterisation of the datasets. For building the proper geometrical characterisation based on the datasets we are training the model on, we need to rely on some strong theoretical foundation (geometrical axiomatisation system, functional analysis are such theory for machine learning for instance), e.g. building proper topological spaces, for example based on the simplicial complexes realisation of the datasets. Just replying on the proximity of points and assuming some metrics may not provide sufficient results as our 'human' bias towards using Euclidean metrics for most of the spaces may again induce some geometrics biases.

As (de Dampierre et al. 2024) mentions in his work on Bunkatopics referring to it as 

"Algorithmically infused societies" (Wagner et al., 2021):

When using common tools for data analysis such as Principal Component Analysis (PCA) (Abdi & Williams, 2010), UMAP (McInnes et al., 2020), and unseen-species models (Chao, 1988), researchers inherit decisions that have partially been made by the creators of these packages.

Scaling in AI

AI tools are growing and are incorporated in many aspects of human lifes.

At the same time when Transformers were introduced, and performance was improved somewhat predictably as one scales up either the amount of compute or the size of the network, a phenomenon which is now referred to as scaling laws [Textbooks is all you need https://arxiv.org/abs/2306.11644]

Yet one can overcome the scaling problems to increase efficiency by curating and coordinating the data on which the models are being trained, relying on quality rather than on quantity seems to be able to do it, but the question remains whether this remains the case always and is statistically significant.

The recent work of Eldan and Li on TinyStories (a high quality dataset synthetically generated to teach English to neural networks) was demonstrating that using the high quality data can dramatically change the shape of the scaling laws, potentially allowing to match the performance of large-scale models with much leaner training/models. Yet these are only evidence based observations [Bubeck et al.]
Sparks of artificial general intelligence: Early experiments with gpt-4]. Can there be any of the theoretical proof that data quality would indeed increase the training?

For that one would need to first agree on quantitative metrics for datasets characterisation. For example the datasets which are labeled with human-generated labels can potentially be a good testbed for that.