I don't think so. A game would use SLI connected GPUs as one big GPU for rendering. Deep learning algorithms can distribute the workload among all GPUs directly through the PCI bus.
The memory is mirrored between the GPUs, it doesn't stack up with SLI. This is because the video memory is used mostly for textures, shades, etc. And it is better to have this data locally from GPU perspective. So it doesn't help in case of training for deep learning.
First one has data dependency problems those should be addressed. Also you should check SLI and available bus for connecting these GPUs. Also you should pay attention to word size (16 or 32 bits). anyway you also need more Memory available on your GPU.