Loading
An embedding is a list of numbers a model produces to stand for an input (a tile, a gene, a patient), placed so that similar inputs land near each other.
In machine learning an embedding maps complex high-dimensional data into a lower-dimensional vector space (Wikipedia). Foundation models are mostly used as embedding machines: a frozen encoder turns each tile or cell into a vector, mean or attention pooling aggregates them into a slide or sample vector (Wikipedia on pooling layers), and a small model on top does the prediction. Whether the embedding groups samples by tissue or subtype without any labels (cluster purity) is the first check of its quality.
Shares Contrastive learning (InfoNCE), Principal component analysis (PCA) as a feature compressor, Transformer and attention, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.
Shares Zero-shot prediction, Fine-tuning versus frozen features, and LoRA, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.
Shares Linear probe, Principal component analysis (PCA) as a feature compressor, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.
Shares Tokenisation (genes, tiles and sequence as tokens), Zero-shot prediction, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.
Shares Zero-shot prediction, Transformer and attention, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.
Shares Tokenisation (genes, tiles and sequence as tokens), Fine-tuning versus frozen features, and LoRA, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.
Shares Fine-tuning versus frozen features, and LoRA, Transformer and attention, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.
Shares Fine-tuning versus frozen features, and LoRA, Cancer AI vocabulary (CanSim terms map) and the tag cansim-terms.