构建大型可伸缩的机器学习算法框架

1 .Building parallel machine learning algorithms: scaling out and up William Benton
 willb@redhat.com • @willb 

2 .Motivation 

3 .Motivation 

4 .Motivation 

5 .Motivation 

6 .Forecast Introducing our case study: self-organizing maps Parallel implementations for partitioned collections (in particular, RDDs) Beyond the RDD: data frames and ML pipelines Practical considerations and key takeaways 

7 .Introducing our case study 

8 .

9 .

10 .Training self-organizing maps 

11 .Training self-organizing maps 

12 .Training self-organizing maps 

13 .Training self-organizing maps 

14 .Training self-organizing maps while t < maxupdates: random.shuffle(examples) for ex in examples: t = t + 1 if t == maxupdates: break bestMatch = closest(somt, ex) for (unit, wt) in neighborhood(bestMatch, sigma(t)): somt+1[unit] = somt[unit] + (ex - somt[unit]) * alpha(t) * wt 

15 .Training self-organizing maps process the training while t < maxupdates: set in random order random.shuffle(examples) for ex in examples: t = t + 1 if t == maxupdates: break bestMatch = closest(somt, ex) for (unit, wt) in neighborhood(bestMatch, sigma(t)): somt+1[unit] = somt[unit] + (ex - somt[unit]) * alpha(t) * wt 

16 .Training self-organizing maps process the training while t < maxupdates: set in random order random.shuffle(examples) for ex in examples: the neighborhood size controls t = t + 1 how much of the map around if t == maxupdates: the BMU is affected break bestMatch = closest(somt, ex) for (unit, wt) in neighborhood(bestMatch, sigma(t)): somt+1[unit] = somt[unit] + (ex - somt[unit]) * alpha(t) * wt 

17 .Training self-organizing maps process the training while t < maxupdates: set in random order the learning rate controls random.shuffle(examples) how much closer to the for ex in examples: example each unit gets the neighborhood size controls t = t + 1 how much of the map around if t == maxupdates: the BMU is affected break bestMatch = closest(somt, ex) for (unit, wt) in neighborhood(bestMatch, sigma(t)): somt+1[unit] = somt[unit] + (ex - somt[unit]) * alpha(t) * wt 

18 .Parallel implementations for partitioned collections 

19 .Historical aside: Amdahl’s Law 1 lim So = sp —> ∞ 1 - p 

20 .What forces serial execution? 

21 .What forces serial execution? 

22 .What forces serial execution? state[t+1] = combine(state[t], x) 

23 .What forces serial execution? state[t+1] = combine(state[t], x) 

24 .What forces serial execution? f1: (T, T) => T f2: (T, U) => T 

25 .What forces serial execution? f1: (T, T) => T f2: (T, U) => T 

26 .What forces serial execution? f1: (T, T) => T f2: (T, U) => T 

27 .How can we fix these? a⊕b=b⊕a (a ⊕ b) ⊕ c = a ⊕ (b ⊕ c) 

28 .How can we fix these? a⊕b=b⊕a (a ⊕ b) ⊕ c = a ⊕ (b ⊕ c) 

29 .How can we fix these? a⊕b=b⊕a (a ⊕ b) ⊕ c = a ⊕ (b ⊕ c)