How To Survive A Practiced Motorcar Learning Production Manager
There are a lot of interesting meetups at Seattle, together with I travail to attend i every duo weeks. Ruben Lozano Aguilera was the speaker for this meetup on October 17. Ruben is a production managing director at Google Cloud, together with earlier that he was a production managing director at Amazon.
Machine learning turns information + answers into rules.
For instance ML is a proficient lucifer for the "search" problem. Search requires complex logic, for which it is non slow to develop rules. It scales upwards actually fast inwards damage of novel keywords, combinations together with content. It requires personalization depending on the context, together with has unopen to real-time adaptation element equally well.
Another of import betoken is that the occupation should convey existing examples of actual answers. When you lot bootstrap from a proficient plenty dataset, you lot tin give the axe scale further, because information -> predictions -> client sense -> to a greater extent than traffic -> to a greater extent than data.
Some pop ML problems are ranking, recommendation, classification, regression, clustering, together with anomaly detection.
Don't purpose ML when your problem:
The information requires farther consideration: Can you lot purpose data? Is it available, accessible, together with sufficient? Is high quality? relevant, fresh, representative, unbiased? Is it appropriate to purpose the data: privacy, safety concerns?
For the following, tin give the axe you lot purpose ML or not?
The people come upwards from 2 domains:
The ML lifecycle involves four phases: problem, data, features, together with model.
To formulate the problem, you lot demand to clarify what to solve, found measurable goals, together with decide what to predict.
For the information phase, you lot demand to
For the features phase, you lot demand to consider scaling, decomposition, aggregation, together with discard whatever features that are non relevant.
Finally, for the model phase, you lot outset dissever the information gear upwards into preparation information together with seek data, could last 70+30 or 90+10. Then comes the model preparation (using whatever algorithm you lot are using), which produces the ML model. You hence seek this output ML model with the seek data.
To productize your model, you lot should integrate the ML solution with existing software together with range off it running over time. At this betoken considerations virtually the deployment environment, information storage, safety together with privacy, monitoring & maintenance come upwards inwards to play. Some slap-up ML solutions cannot last productized due to high implementation costs or inability to last tested inwards practice.
The production managing director is really much involved inwards the outset 2 phases: formulating the occupation together with selecting the data. Product managing director is also involved inwards characteristic alternative but non much involved with the concluding model phase.
Since the presentation didn't elevate whatever deep learning specific problems/tasks/observations, I asked Ruben virtually what significance deep learning had on the projects he worked on. Turns out, he didn't purpose any. He said that simpler ML models were plenty for the tasks they undertook hence he never needed a deep-learning solution. He also said that deep-learning was really expensive upwards until a duo years ago, together with that was also a factor.
With TensorFlow, Google is supposedly using deep-learning a lot, probable to a greater extent than for icon together with vocalism processing. But, is in that place a written report virtually the prominence of deep-learning purpose alongside ML solutions inwards the industry?
2) How do you lot troubleshoot issues with productizing ML?
As nosotros covered to a higher house in that place are many things that tin give the axe become wrong, such as unanticipated bias inwards your data, inwards your method, conclusions. How do you lot depository fiscal establishment check for these? Ruben answered they brainstorm together with intend really deeply virtually what could become wrong, together with seat these issues. It seems similar this needs to a greater extent than processes together with tool support. Having seen how TLA+ specifications together with model checking create wonders for checking problems with distributed/concurrent systems, I am wondering if similar blueprint score tool back upwards could last developed for ML solutions.
3) How do nosotros learn/teach empathy?
Ruben was a slap-up speaker. He used beautifully designed slides. After all he is a production managing director together with sympathizes with the users/audience. In Q&A session he mentioned that empathy is the most of import science for a ML production manager. I believe empathizing with your audience also goes a long means inwards populace speaking. How do nosotros learn/teach empathy? This is hence basic that you lot expect/hope nosotros larn this equally kids. But it looks similar nosotros range off forgetting virtually this together with neglect to empathize. Also, in that place is e'er levels to things. How do nosotros acquire improve at this?
4) Is ML/DL likewise application-coupled?
I convey a unopen to agreement of ML/DL domain, since I started learning virtually it inwards 2016. I am withal amazed at how tightly application-coupled is the ML/DL work. On i paw this is good, this makes ML/DL really practical together with really applicable. On the other hand, this makes it harder to written report the principles together with systematize knowledge.
What is ML?
Programming transforms information + rules into answers.Machine learning turns information + answers into rules.
When should you lot purpose ML?
Use ML if the problem:- handles complex logic
- scales upwards actually fast
- requires specialized personalization
- adapts inwards real-time
For instance ML is a proficient lucifer for the "search" problem. Search requires complex logic, for which it is non slow to develop rules. It scales upwards actually fast inwards damage of novel keywords, combinations together with content. It requires personalization depending on the context, together with has unopen to real-time adaptation element equally well.
Another of import betoken is that the occupation should convey existing examples of actual answers. When you lot bootstrap from a proficient plenty dataset, you lot tin give the axe scale further, because information -> predictions -> client sense -> to a greater extent than traffic -> to a greater extent than data.
Some pop ML problems are ranking, recommendation, classification, regression, clustering, together with anomaly detection.
Don't purpose ML when your problem:
- can last solved past times uncomplicated rules
- does non suit to novel data
- requires 100% accuracy
- requires total interpretability/why-provenance
The information requires farther consideration: Can you lot purpose data? Is it available, accessible, together with sufficient? Is high quality? relevant, fresh, representative, unbiased? Is it appropriate to purpose the data: privacy, safety concerns?
For the following, tin give the axe you lot purpose ML or not?
- What clothe items should last protected past times copyright? No. This is unsafe financially, you lot demand to acquire 100% accuracy.
- Which resumes should nosotros prioritize to interview for our candidate pipeline? No, this may last based on biased data.
- What products should last alone sold to Hispanics inwards the US? No. This is discriminatory together with creepy.
- Which sellers convey the greatest revenue potential? Yes.
- Where should Amazon construct side past times side caput quarters? No. This is non a repeatable problem; in that place is simply i label: Seattle.
- Which search queries should nosotros reach for the Amazon fresh store? Yes.
What is the ML lifecycle?
For productizing ML, you lot demand people, processes, together with tools/systems.The people come upwards from 2 domains:
- Math, statistics: ml scientist, applied scientist, resarch scientist, information scientist
- Software, programming: occupation concern word engineer, information engineer, software engineer, dev manager, technical programme manager
The ML lifecycle involves four phases: problem, data, features, together with model.
To formulate the problem, you lot demand to clarify what to solve, found measurable goals, together with decide what to predict.
For the information phase, you lot demand to
- select data: available, missing data, discarding information (data cleaning)
- preprocess data: formatting, cleaning, sampling
For the features phase, you lot demand to consider scaling, decomposition, aggregation, together with discard whatever features that are non relevant.
Finally, for the model phase, you lot outset dissever the information gear upwards into preparation information together with seek data, could last 70+30 or 90+10. Then comes the model preparation (using whatever algorithm you lot are using), which produces the ML model. You hence seek this output ML model with the seek data.
To productize your model, you lot should integrate the ML solution with existing software together with range off it running over time. At this betoken considerations virtually the deployment environment, information storage, safety together with privacy, monitoring & maintenance come upwards inwards to play. Some slap-up ML solutions cannot last productized due to high implementation costs or inability to last tested inwards practice.
The production managing director is really much involved inwards the outset 2 phases: formulating the occupation together with selecting the data. Product managing director is also involved inwards characteristic alternative but non much involved with the concluding model phase.
MAD questions
1) Umm, deep learning?Since the presentation didn't elevate whatever deep learning specific problems/tasks/observations, I asked Ruben virtually what significance deep learning had on the projects he worked on. Turns out, he didn't purpose any. He said that simpler ML models were plenty for the tasks they undertook hence he never needed a deep-learning solution. He also said that deep-learning was really expensive upwards until a duo years ago, together with that was also a factor.
With TensorFlow, Google is supposedly using deep-learning a lot, probable to a greater extent than for icon together with vocalism processing. But, is in that place a written report virtually the prominence of deep-learning purpose alongside ML solutions inwards the industry?
2) How do you lot troubleshoot issues with productizing ML?
As nosotros covered to a higher house in that place are many things that tin give the axe become wrong, such as unanticipated bias inwards your data, inwards your method, conclusions. How do you lot depository fiscal establishment check for these? Ruben answered they brainstorm together with intend really deeply virtually what could become wrong, together with seat these issues. It seems similar this needs to a greater extent than processes together with tool support. Having seen how TLA+ specifications together with model checking create wonders for checking problems with distributed/concurrent systems, I am wondering if similar blueprint score tool back upwards could last developed for ML solutions.
3) How do nosotros learn/teach empathy?
Ruben was a slap-up speaker. He used beautifully designed slides. After all he is a production managing director together with sympathizes with the users/audience. In Q&A session he mentioned that empathy is the most of import science for a ML production manager. I believe empathizing with your audience also goes a long means inwards populace speaking. How do nosotros learn/teach empathy? This is hence basic that you lot expect/hope nosotros larn this equally kids. But it looks similar nosotros range off forgetting virtually this together with neglect to empathize. Also, in that place is e'er levels to things. How do nosotros acquire improve at this?
4) Is ML/DL likewise application-coupled?
I convey a unopen to agreement of ML/DL domain, since I started learning virtually it inwards 2016. I am withal amazed at how tightly application-coupled is the ML/DL work. On i paw this is good, this makes ML/DL really practical together with really applicable. On the other hand, this makes it harder to written report the principles together with systematize knowledge.
0 Response to "How To Survive A Practiced Motorcar Learning Production Manager"
Post a Comment