Today we have released an upgrade to TensorFlow 2.11 for all built-in neural network blocks in Edge Impulse. This should have little immediate impact on your models (you might see very small variations in accuracy / quantization error), but it sets us up for support for new model types (e.g. new transformer models) / VeLO integration / better JAX integration / etc. This is only for training, the inference libraries have not changed (so you cannot use all 2.11 ops in EI for Linux yet for example).
If you see anything weird (errors, big accuracy drops) on existing models, please ping us in this thread.
Why TF2.11 and not TF2.13? With hundreds of model types, 20+ accelerators, and 220K+ projects it takes a long time to validate / backport fixes / etc. So this has been ongoing for a while.