3 ms·

I'm not so sure when it comes to machine learning. Those techniquss represent such a fundamentally different approach to pattern recognition that it enables the

• johannes1234321 8y ago
  True. But many many many problems can be solved without that hammer as well.
• omginternets 8y ago
  This is undeniably true, but consider that 99% of classification problems encountered in the industry can be solved with a logistic (if not linear) regression.
• gaius 8y ago
  Those techniquss represent such a fundamentally different approach to pattern recognition You are kidding right? With the exception of a few modern breakthroughs like GAN the vast majority of ML is 70s and 80s ideas that Moore’s Law has just taken mainstream.
  • omginternets 8y ago
    The fact that these approaches are now tractable represents a paradigm-shift in what we can do with computers. That's what the parent-poster meant.
  • ghostcluster 8y ago
    70s and 80s ideas that couldn't be implemented at scale due to the lack of massively parallelized GPU architectures with trillions of transistors to simulate the neurel net layers. It was only in the mid-late 90s that simple handwriting / character recognition techniques were able to be used in somewhat real-time.
  • sanxiyn 8y ago
    > With the exception of a few modern breakthroughs like GAN And ReLU (2011). And Glorot initialzation (2010). And He initialization (2015). And Dropout (2014). And Batch Normalization (2015). And Adam optimizer (2014). And distillation (2015). And... None of these were known or used in 70s and 80s and performance difference is enormous.