For anyone interested in algorithmic changes that improve efficiency, Microsoft’s recent research around moving from floating point weights to ternary ones (1, 0, -1) was really impressive:
Basically at larger parameter sizes it outperforms FP networks while being a fraction of the memory footprint and bypassing the need for matrix multiplication.
It kind of makes sense that it works too, given past research that the networks are creating a virtualized node topology based on combinations of physical nodes, so with enough nodes to work with there isn’t a loss in functionality and the discrete weights should arrive at optimal thresholds more easily than slight adjustments to FP values.
The next generation of models built on this need to be trained from scratch (this is about pretraining and not quantization after the fact), but it should open the door to new hardware architectures better optimized for networks of ternary weights.
For anyone interested in algorithmic changes that improve efficiency, Microsoft’s recent research around moving from floating point weights to ternary ones (1, 0, -1) was really impressive:
https://arxiv.org/abs/2402.17764
Basically at larger parameter sizes it outperforms FP networks while being a fraction of the memory footprint and bypassing the need for matrix multiplication.
It kind of makes sense that it works too, given past research that the networks are creating a virtualized node topology based on combinations of physical nodes, so with enough nodes to work with there isn’t a loss in functionality and the discrete weights should arrive at optimal thresholds more easily than slight adjustments to FP values.
The next generation of models built on this need to be trained from scratch (this is about pretraining and not quantization after the fact), but it should open the door to new hardware architectures better optimized for networks of ternary weights.