AI SLIDE: A Breakthrough or a Shaped Insight
March 4, 2020
DarkCyber noted an interesting, although sketchy summary, of a CPU and hash table approach to machine learning. “Deep Learning Rethink Overcomes Major Obstacle in AI Industry” suggests that Amazon and Google are barking up the wrong artificial intelligence method.
The innovation is the use of hash tables for deep learning. The idea is that one looks up an item, perfect for Intel CPUs. The “old” way relies on matrix mathematics, perfect for nVidia graphics chips. In fact, the solution is a search problem, a point in the write up which may annoy the Googlers; to wit:
“You don’t need to train all the neurons on every case,” Medini [a Rice wizard] said. “We thought, ‘If we only want to pick the neurons that are relevant, then it’s a search problem.’ So, algorithmically, the idea was to use locality-sensitive hashing to get away from matrix multiplication.”
The reason this insight is important is that if it proves useful and can flip the opinions of those innovators with tens of thousands of GPUs generating heat is that machine learning becomes less expensive. (How much does it cost to cool lots of GPUs doing math? Answer: A lot.)
The approach is dubbed SLIDE. The acronym is about as slick as relying on an Intel processor: Sub-Linear Deep Learning Engine. Too bad AMD. You have Linus, the YouTube star, as your cheerleader.
Advantages include:
- Cheaper
- Faster
- More efficient training.
Disadvantages revealed include:
- Memory is needed, lots of memory
- Unexpected cache thrashing (data are here, oops, data are not hear, rinse and repeat)
- Access to Intel engineers reduced the inefficiency by 50 percent, but 50 percent of what? Misses, latency, halts, other?
The point of the announcement is to make clear that Amazon and Google are going about machine learning the wrong way. Does anyone at either firm care? Sure, and it will be fun for the researchers to check out their approach, look up what was investigated in the past, and figure out if it is better to switch than fight.
Net net: Seems interesting and definitely a rah rah for Intel. The write up makes no reference to IBM or other machine learning outfits. Marketing or shaped insight? It is too soon to answer this question definitively.
Stephen E Arnold, March 4, 2020
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
