4 ms·
In many cases this is true but it is important to note that for deep learning, the training was the hard part that was missing for a long time. Better weight in
by phreeza 9y ago
In many cases this is true but it is important to note that for deep learning, the training was the hard part that was missing for a long time. Better weight initialization and momentum methods were what really made deep networks work (not just GPUs, as some people tend to believe).
- ska 9y agoGPUs were necessary but not sufficient, true. But the real change (at least for supervised learning) was not in the polishing on training methods (an incremental improvement), it was the availability of big enough data sets. No amount of modeling can make up for insufficient data.