You Can AI Like an Expert
You Can AI Like an Expert
The goal:
Automated AI workflow
The goal:
Automated AI workflow
Automated AI workflow
The wider goal:
Automated computation workflow
The wider goal:
Automated computation workflow
Automated computation workflow
Computation ⊃ {Statistics, modeling, visualization, machine learning, signal processing, geometry, image processing, maths, semantics, networks, queues, geodesy, random processes, audio, survival analysis
Computation ⊃ {Statistics, modeling, visualization, machine learning, signal processing, geometry, image processing, maths, semantics, networks, queues, geodesy, random processes, audio, survival analysis
The Answer:
Symbolic computation
The Answer:
Symbolic computation
Symbolic computation
Symbolic computation
Examples:
Symbolic computation
Examples:
Examples:
In[]:=
ReplaceAll+3,xu-1
1+x
2
x
3
x
Out[]=
u
2
(-1+u)
3
(-1+u)
In[]:=
FullForm+3
1+x
2
x
3
x
Out[]//FullForm=
Times[Power[Plus[1,x],Rational[1,2]],Power[Plus[Power[x,2],Times[3,Power[x,3]]],-1]]
Symbolic computation
Examples:
Symbolic computation
Examples:
Examples:
In[]:=
Out[]=
Symbolic computation:
All data is the same
Symbolic computation:
All data is the same
All data is the same
Symbolic computation:
Unified representation of structure, data & meaning
Symbolic computation:
Unified representation of structure, data & meaning
Unified representation of structure, data & meaning
Automated Machine Learning
The model workflow
Automated Machine Learning
The model workflow
The model workflow
Building machine learning models
Building machine learning models
Example: Rock Paper Scissors
Example: Rock Paper Scissors
What about real machine learning?
What about real machine learning?
High dimensional output (richer loss function control)
High dimensional output (richer loss function control)
Task specific model architectures
Task specific model architectures
Richer encoder/decoder control
Richer encoder/decoder control
Automating neural networks
Automating neural networks
Basic example
Basic example
Encoding & decoding
Encoding & decoding
Available components
Available components
Neural network...
...Like an expert
Neural network...
...Like an expert
...Like an expert
Neural network...
...Like an expert
Neural network...
...Like an expert
...Like an expert
Use existing networks
Use existing networks
Retrain existing networks
Retrain existing networks
Adapt existing networks
Adapt existing networks
The goal:
Automated AI workflow
The goal:
Automated AI workflow
Automated AI workflow
Automated AI workflow:
Import, clean, model, validate, deploy
Automated AI workflow:
Import, clean, model, validate, deploy
Import, clean, model, validate, deploy
Automated AI workflow:
Cleaning
Automated AI workflow:
Cleaning
Cleaning
Example: Missing values
Example: Missing values
Example: Anomaly detection
Example: Anomaly detection
Automated AI workflow:
Validation
Automated AI workflow:
Validation
Validation
Standard validation
Standard validation
Model inspection
Model inspection
Automated AI workflow:
Deployment
Automated AI workflow:
Deployment
Deployment
The real goal:
Automated computation workflow
The real goal:
Automated computation workflow
Automated computation workflow
Use machine learning when:
Data rich and understanding poor
Use machine learning when:
Data rich and understanding poor
Data rich and understanding poor
Multiparadigm computation:
Injecting human knowledge
Multiparadigm computation:
Injecting human knowledge
Injecting human knowledge
Multiparadigm computation:
Apply machine learning
Multiparadigm computation:
Apply machine learning
Apply machine learning
Summary
Summary
More automation is always important
More automation is always important
Richest data structures and broad tools enable specific tasks
Richest data structures and broad tools enable specific tasks
“Doing AI” is becoming routine
“Doing AI” is becoming routine
The “value” is the whole workflow
The “value” is the whole workflow