Foundation Model and their reuse

Foundation model is the new winged concept, but it is not really new. In its foundation it is about learning the identity of presented objects e.g. images of different cats. For using it in other applications as it was initially trained on is about dropping the last (inference) layer(s) and further train it. By doing so, the weights getting readjusted. The adjustment follows (often) the PEFT approach. This approach has different flavor of application e.g.LORA (just adapt a few weights to keep the training parameter low).