Training deep neural networks in reduced precision promises significant efficiency gains. However, the paper notes that all existing methods that scale to standard computer vision benchmarks still retain floating-point arithmetic in at least one stage, typically for the weights. That lingering floating-point component limits the potential memory savings.

The new work introduces an integer-only training scheme that removes this last floating-point stage. By keeping all operations in integer form, the method aims to cut memory usage more aggressively than prior reduced-precision approaches.

The abstract does not provide experimental details, but it states the method scales to standard computer vision benchmarks. That suggests the integer-only approach does not sacrifice the ability to handle common vision tasks, though specific accuracy numbers are not given in the source.