Fundamentals of Deep Learning: Part 6: Advanced Architectures
Fundamentals of Deep Learning: Part 6: Advanced Architectures
Fundamentals of Deep Learning: Part 6: Advanced Architectures
DEEP LEARNING
Part 6: Advanced Architectures
1
Part 1: An Introduction to Deep Learning
Computer Vision
• Optometry
Reinforcement Learning
• Game Theory
• Psychology
Anomaly Detection
• Security
• Medicine
5
FIELDS OF AI
Computer Vision
• Optometry
Reinforcement Learning
• Game Theory
• Psychology
Anomaly Detection
• Security
• Medicine
6
FIELDS OF AI
Computer Vision
• Optometry
Reinforcement Learning
• Game Theory
• Psychology
Anomaly Detection
• Security
• Medicine
7
NATURAL LANGUAGE
PROCESSING
8
FROM WORDS TO NUMBERS
Dictionary
“A dog barked at a cat.”
1. A 8. Cat
2. An 9. Cats
[1, 10, 7, 4, 1, 8] 3. And 10. Dog
4. At 11. Dogs
5. Ate 12. Eat
6. Bark
7. Barked
9
FROM WORDS TO NUMBERS
Inputs Outputs
A A
An An Dictionary
And And
At At
1. A 8. Cat
Ate Ate
2. An 9. Cats
Bark Bark
Barked Barked
3. And 10. Dog
Cat Cat
4. At 11. Dogs
Cats Cats
5. Ate 12. Eat
Dog Dog
6. Bark
Dogs Dogs 7. Barked
Eat Eat
10
FROM WORDS TO NUMBERS
Inputs Outputs
0 0%
0 0% Dictionary
0 10%
0 5%
1. A 8. Cat
0 35%
2. An 9. Cats
0 0%
0 50%
3. And 10. Dog
0 0%
4. At 11. Dogs
0 0%
5. Ate 12. Eat
1 0%
6. Bark
0 0% 7. Barked
0 0%
11
FROM WORDS TO NUMBERS
Big
Giraffe
(.9, .9)
Llama
Bigger Dictionary
(-.9, .1)
1. A 32. Bark 63. Dogs
2. An 33. Barked 64. Eat
3. And 34. Cat 65. Eaten
4. At 35. Cats 66. A
5. Ate 36. Dog 67. An
6. Bark 37. Dogs 68. And
7. Barked 38. Eat 69. At
8. Cat 39. Eaten 70. Ate
Domestic Wild 9.
10.
Cats
Dog
40.
41.
A
An
71.
72.
Bark
Barked
11. Dogs 42. And 73. Cat
12. Eat 43. At 74. Cats
13. Eaten 44. Ate 75. Dog
14. A 45. Bark 76. Dogs
15. An 46. Barked 77. Eat
16. And 47. Cat 78. Eaten
17. At 48. Cats 79. …
Falcon 18.
19.
Ate
Bark
49.
50.
Dog
Dogs
80.
81.
…
…
(.85, -.65)
26. Eaten 57. Ate
12
FROM WORDS TO NUMBERS
13
RECURRENT NEURAL
NETWORKS
14
RECURRENT NEURAL NETWORKS
1. Cats
2. Dogs
3. Meow
4. Say
5. Woof
15
RECURRENT NEURAL NETWORKS
Inputs
16
RECURRENT NEURAL NETWORKS
0
“Cats say ___.”
0
0
0% “Dogs say ___.” Dictionary
0%
50%
1 1. Cats
50% 2. Dogs
0
RNN 0% 3. Meow
0 4. Say
0 Outputs 5. Woof
Embedding
0
Inputs
17
RECURRENT NEURAL NETWORKS
0
“Cats say ___.”
0
0
0% “Dogs say ___.” Dictionary
.1
0%
-.5
50%
1 1. Cats
.6
50% 2. Dogs
0
RNN 0% 3. Meow
0 4. Say
0 Outputs 5. Woof
Embedding
0
Inputs
18
RECURRENT NEURAL NETWORKS
0
“Cats say ___.”
0
0
“Dogs say ___.” Dictionary
.1
-.5
1 1. Cats
.6
2. Dogs
0
RNN 3. Meow
0 4. Say
0 Outputs 5. Woof
Embedding
0
Inputs
19
RECURRENT NEURAL NETWORKS
.1
“Cats say ___.”
-.5
.6
“Dogs say ___.” Dictionary
0 1. Cats
0
2. Dogs
RNN 3. Meow
0 4. Say
1 Outputs 5. Woof
Embedding
0
Inputs
20
RECURRENT NEURAL NETWORKS
.1
“Cats say ___.”
-.5
.6
0% “Dogs say ___.” Dictionary
-.3
0%
.2
100% 1. Cats
0
.5
0
0% 2. Dogs
RNN 0%
3. Meow
0 4. Say
1 Outputs 5. Woof
Embedding
0
Inputs
21
RECURRENT NEURAL NETWORKS
Input Input
RNN LSTM
Output Output
22
OTHER ARCHITECTURES
23
AUTOENCODERS
Inputs Outputs
24
AUTOENCODERS
Inputs Outputs
25
AUTOENCODERS
-.3 -.3
.6 .6
Encoder Decoder
26
GENERATIVE ADVERSARIAL NETWORKS (GANS)
Prediction
Real Discriminator
Images
Real
Generator
Fake
Noise Fake
Images
27
REINFORCEMENT LEARNING
Agent Environment
28
NEXT STEPS
29
ENABLING PORTABILITY WITH NGC CONTAINERS
Extensive
NGC Deep Learning Containers
- Diverse range of workloads and industry specific use cases
Optimized
- DL containers updated monthly
- Packed with latest features and superior performance
Secure & Reliable
- Scanned for vulnerabilities and crypto
- Tested on workstations, servers, & cloud instances
Scalable
- Supports multi-GPU & multi-node systems
Designed for Enterprise & HPC
- Supports Docker, Singularity & other runtimes
Run Anywhere
- Bare metal, VMs, Kubernetes
- x86, ARM, POWER
- Multi-cloud, on-prem, hybrid, edge
31
CLOSING THOUGHTS
32
COPYING ROCKET SCIENCE
33
LET’S GET STARTED!
34
35