Download to read offline
![NMC Data & NLP Data
NLP data:
Observation: [‘This’, ‘is’, ‘a’,
‘tokenized’, ‘feature’, ‘vector’,
‘used’, ‘for’, ‘machine’, ‘learning’,
‘in’, ‘NLP’]
NMC data:
User: [ ‘segment: Likes Outdoors’,
‘segment: Male 25-35’, ‘location:
New York, NY’]](https://image.slidesharecdn.com/deeplearningatnmcdevinjones-170308155248/75/Deep-learning-at-nmc-devin-jones-19-2048.jpg)
Uploaded byIdo Shilon
Deep learning at nmc devin jones
This document discusses deep learning research at Nielsen's Machine Learning Lab. It begins with an introduction to the speaker's background and an overview of machine learning categories. It then discusses Nielsen's use of deep learning for tasks like look-alike modeling and online ad targeting. The rest of the document covers motivations for deep learning, neural network architectures like convolutional and recurrent networks, and techniques to improve inference speed, such as using GPUs, sparse matrix multiplication, and network trimming.
More Related Content
Similar to Deep learning at nmc devin jones
Deep learning at nmc devin jones
- 1. Deep Learning atNMC Devin Jones Director Machine Learning Lab, Nielsen
- 2. Devin Jones ● MachineLearning & Statistics ○ Research ■ Classification ■ Inference ■ Time Series ○ Application ■ Large scale ■ Streaming Introduction ● Columbia University ○ CS/ML ● Rutgers University ○ Statistics ○ Econ ○ Operations Research ● Ad Tech (7 years)
- 3. ML at NMC Introto Deep Learning Deep Learning Research at NMC Agenda
- 4.
- 5. “ ” Used to buildlarger audiences from smaller audience segments to create reach for advertisers. In theory, they reflect similar characteristics to a benchmark set of characteristics the original audience segment represents, such as in-market kitchen-appliance shoppers. adage.com The ML Challenge at NMC Look Alike Modeling
- 6. The ML Challengeat NMC: An Example
- 7. Supervised Classification forOnline Ad Targeting ML at NMC Data Algorithm Model
- 8. Supervised What? Machine Learninghas two main categories: Supervised Learning Unsupervised Learning
- 9. Supervised What? Machine Learninghas two main categories: Supervised Learning: Inferences on Labeled Data Unsupervised Learning: Inferences on Unlabeled Data
- 10. Supervised vs UnsupervisedLearning Supervised: Spam or Ham? Unsupervised: Clustering Wikipedia Articles
- 11.
- 12. The quality ofdata for a model will influence the model’s success At NMC, we have access to high dimensional, sparse data: The Feature Set & Scale Models are trained in batches of 100,000 to 100,000,000 users depending on the purpose ~4,000 Segments ~200 Publishers User Agent Geographic Info (zip code) + + + Resulting in over 100k features to choose from
- 13. To date, wehave implemented these algorithms in our real time scoring engine: We score billions of events per day using these models and our ML infrastructure ML Algorithms at NMC Binary Linear Model kNN Multinomial Linear Models Online Learning for Linear Models Random Forest And of course… Deep Learning
- 14.
- 15. Topics Motivation Intro toDL NN Architecture GPU vs CPU
- 16.
- 17. 2 1 Recent Successin Deep Learning NMC data is similar to Natural Language Processing (NLP) data Certain ad targeting problems can be framed as expressive, hierarchical relationships MOTIVATION 3
- 18. Deep Learning: RecentSuccess ▪ AlphaGo defeats all world top professional Go players ▪ Image and Speech recognition exceed human abilities ▪ AI in consumer products: Amazon Echo Google Home Autonomous Driving All of these recent AI breakthroughs are based on Deep Neural Networks!
- 19. NMC Data &NLP Data NLP data: Observation: [‘This’, ‘is’, ‘a’, ‘tokenized’, ‘feature’, ‘vector’, ‘used’, ‘for’, ‘machine’, ‘learning’, ‘in’, ‘NLP’] NMC data: User: [ ‘segment: Likes Outdoors’, ‘segment: Male 25-35’, ‘location: New York, NY’]
- 20. Deep Learning: Somedefinitions
- 21.
- 22.
- 23.
- 24. Neural Network: Neuron Livesin NYC? = Yes Orders from Dominos? Works in ad tech? 0.5 = 0.01= 0.7 = = 1.2= No = Yes
- 25. A Deep NeuralNetwork
- 26. Neural Network Phases TrainingInference
- 27.
- 28. NMC high-level architecture MachineLearning (Training) (Inference)
- 29. Definition Summary ● Training ●Inference ○ Matrix Multiplication ● Nodes ● Layers ● Network
- 30.
- 31. DNN Architecture Image Processing:: Convolutional Networks Speech Recognition :: Recurrent Networks AlphaGo :: Reinforcement Learning
- 32. An Architecture Example:Conv Nets
- 33.
- 34.
- 35. Figure 2. Convergenceof neural network model with forward shortcut (Residual Net) Figure 1. Convergence of neural network model without forward shortcut(regular net) Residual Network Convergence
- 36.
- 37. Category Segment City Prosperity World-ClassHealth Uptown Elite Penthouse Chic Metro High-Flyers Prestige Positions Premium Fortunes Diamond Days Alpha Families Bank of Mum and Dad Empty-Nest Adventure Multi-level Hierarchical Classification
- 38. C1 C2 C3 S1S3 S4 S5 S6S2 Multi-level Hierarchical Classification
- 39. C1 C2 C3 S1S3 S4 S5 S6S2 Naive Approach Multi-level Hierarchical Classification
- 40. DNN For Multi-LevelHierarchical Classification
- 41.
- 42. Batch Size &Processing Time
- 43. We are notbatching matrix algebra operations NMC Serving operates on 1 request at a time!! GPU vs CPU
- 44.
- 45. Inference on aLayer ~ Matrix Multiplication Input Hidden
- 46.
- 47.
- 48. WEAK CONNECTIONS most connectionsin deep neural network are very weak and can be removed TRIMMING LOW ACCURACY IMPACT the trimming has very little impact on the accuracy COMPRESSED DATA the trimming models can be described by sparse matrices, and thus the data in models are highly compressed
- 49. Neural Network WithoutTrimming
- 50.
- 51. Model Model File Size(MB) Trimming Threshold Accuracy Scoring Time (ms) Not trimmed 108 0.0 13.29 10.0 Trimmed 2.7 0.001 13.30 0.22 Trimming: Space, Time & Performance inference improvement, in CPU time and storage 50x
- 52. Key Takeaways Architecture: ● ResidualNetworks saved the day ● Leverage expressive power of DNN for your data Inference: ● You might not need a GPU for Deep Learning ● Improvements can be made on Sparse Matrix Algebra libraries ● Use trimming
- 53.