“Exponentially Faster Language Modeling”, Peter Belcak, Roger Wattenhofer2023-11-15 (NN sparsity, MoE NN)⁠:

[Warning: apparently irreproducible & may be fake] Language models only really need to use an exponential fraction of their neurons for individual inferences.

As proof, we present UltraFastBERT, a BERT variant that uses 0.3% of its neurons during inference while performing on par with similar BERT models. UltraFastBERT selectively engages just 12⁄4095 neurons for each layer inference. This is achieved by replacing feedforward networks with fast feedforward networks (FFFs). While no truly efficient implementation currently exists to unlock the full acceleration potential of conditional neural execution, we provide high-level CPU code achieving 78× speedup over the optimized baseline feedforward implementation, and a PyTorch implementation delivering 40× speedup over the equivalent batched feedforward inference.

We publish our training code, benchmarking setup, and model weights on Github.