Engineering Machine Learning technologies
1 like484 views
The document discusses the development of machine learning technologies, emphasizing user focus and ethical AI applications that prioritize social benefits, avoid bias, and uphold human rights. It outlines principles for ethical AI usage, product design, and detailed aspects of speech recognition and natural language processing through modular approaches. Additionally, it highlights the importance of training data availability and incremental product development cycles in AI engineering.
![LSTM
LSTM LSTM
LinearLinear
X1
(k)[t] X2
(k)[t]
FP
h1
(k)[t] h2
(k)[t]
FS
y(k)[t]
Acoustic
Model
Gated Feedback
Output Targets
LSTM
DNN
DNN
tConv
pool
+
nonlin
Linear
LSTM
LSTM
LSTM
DNN
Linear
Multitask
Learning
Improvements](https://image.slidesharecdn.com/productificationmachinelearningtechnologies-190208155017/85/Engineering-Machine-Learning-technologies-21-320.jpg)
Engineering Machine Learning technologies
- 1. Engineering Machine Learning Technologies From ideas to products with ML/AI Andres L. Martinez a.k.a almo Google Cloud DevRel @davilagrau
- 2. Why is this time different?Photo by Ross Findon on Unsplash
- 3. Photo by Quino Al on Unsplash Innovation
- 4. Photo by Finn Hackshaw on Unsplash
- 5. Fast{er} Photo by Josh Calabrese on Unsplash
- 6. Focus on the users
- 7. the user
- 8. Let’s start with the basics Photo by Hello I'm Nik on Unsplash
- 9. Don’t be afraid to launch a product without machine learning
- 10. Don’t be Evil 2018 AI for Good
- 11. Objectives for AI applications ● Be socially beneficial. ● Avoid creating or reinforcing unfair bias. ● Be built and tested for safety. ● Be accountable to people. ● Incorporate privacy design principles. ● Uphold high standards of scientific excellence. ● Be made available for uses that accord with these principles. AI applications we will not pursue: ● Technologies that cause or are likely to cause overall harm. ● Weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people. ● Technologies that gather or use information for surveillance violating internationally accepted norms. ● Technologies whose purpose contravenes widely accepted principles of international law and human rights. AI Principles
- 12. Productification Photo by AbsolutVision on Unsplash Photo by Fancycrave on Unsplash
- 13. First, design and implement metrics
- 14. Product Design
- 17. What does it make it so special? ● Filtering Definition + Signal Reconstruction ● Speech Recognition ● Natural Language Processing
- 19. References Control Engineering Neural Network Adaptive Beamforming for Robust Multichannel Speech Recognition (2016) http://bit.ly/2zQWp08 Speech Recognition State of the Art Speech Recognition With Sequence to Sequence Model (2018) http://bit.ly/2Bb8JIo Natural Language Processing BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018) http://bit.ly/2Qt3Vsn http://bit.ly/2rt7hfW
- 20. Neural {Adaptive} Beamforming Neural Network Adaptive Beamforming for Robust Multichannel Speech Recognition (2016)
- 21. LSTM LSTM LSTM LinearLinear X1 (k)[t] X2 (k)[t] FP h1 (k)[t] h2 (k)[t] FS y(k)[t] Acoustic Model Gated Feedback Output Targets LSTM DNN DNN tConv pool + nonlin Linear LSTM LSTM LSTM DNN Linear Multitask Learning Improvements
- 22. Experimental Setup: simulated Training Setup ● 2000 hours noisy training data ● 3 millions English utterances ● Room Simulator ○ Noise ○ Reverberation ● Noise ○ YouTube ○ Daily Noise ● SNR 0-20 dB (12 dB average) Evaluation Setup ● 200 hours ● 30k English Utterances ● Different target speaker ● Different noise position ● Similar mic array geometry
- 23. Speech Recognition State of the Art Speech Recognition With Sequence to Sequence Model (2018)
- 24. L{isten} A{ttend} S{pell} Encoder Attention Decoder Acoustic Model Pronunciation Model Language Model Queries Queries Queries Queries Project Source Project Source Project Source Project Source Project Source Word piece Model Multi-Headed Attention Improvements
- 25. Experimental Setup: simulated Training Setup ● 12500 hours noisy training data ● 15 millions English utterances ● Room Simulator ○ Noise ○ Reverberation ● Noise ○ YouTube ○ Daily Noise ● SNR 0-30 dB (15 dB average) Evaluation Setup ● 14800 English Utterances ● 15700 Longer sentences
- 26. Natural Language Processing BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018)
- 28. Experimental Setup and benchmark Training Setup ● BooksCorpus 800M ● Wikipedia 2500M Benchmark ● GLUE Dataset ● MNLI ● QQP ● QNLI ● SST-2 ● CoLa ● STS-B ● MRPC ● RTE ● WNLI ● SQuAD v1.1 ● SWAG (Adversarial)
- 29. Summary ● Modular Approach ○ Beamforming / Speech Recognition /NLP ● Moral Risk insignificant or reduced ○ AI Principles easy to apply and verify ● Availability of training data ○ Simulated ● Transfer Learning ● Short/Incremental productification cycle
- 30. Don’t forget to check this out! Best Practices for ML Engineering (Martin Zinkevich) http://bit.ly/2EDsLza
- 31. … and this! AI Principles http://bit.ly/2EDsLza
- 32. Progress of AI and Machine Learning https://goo.gl/FWWBv2 … and this!!!
- 33. Thank you! Questions? Andres L. Martinez a.k.a almo @davilagrau