1.
Sebastian JORDAN MONTANO 1
sebastian.jordan@inria.fr
pharo-ai
Oleksandr ZAITSEV 2,1
oleksandr.zaitsev@arolla.fr
1 Inria, Univ. Lille, CNRS, Centrale Lille, UMR 9189 - CRIStAL
2 Arolla, Paris
August 2022

2.
a modular library for shallow machine
learning in Pharo
github.com/pharo-ai
We introduce pharo-ai v0.8

3.
Artificial Intelligence
Machine Learning
Deep
Learning
Techniques which enable
machines to mimic human
behaviour
Subset of AI techniques
which use statistical methods
to enable machines to
improve with experience
Subset of ML which uses
deep neural networks
(A*, knowledge based, …)
(linear regression, k-means,
naive bayes, …)
(convolutional networks,
GAN, Seq2seq, …)
3

4.
Machine Learning
Supervised Unsupervised
- No labeled data
- Extract patterns
from the data
- Labeled data
- First learn from
examples, then
apply to new data
(classification,
regression,…)
(clustering, anomaly
detection,…)
4

5.
Why do we need a ML library in Pharo
- We would like to contribute to the work that is currently being
developed by different people (Univ. Chile, Object Profile-Chile,
PolyMathOrg, Semantics-Bolivia, CIRAD-France).
- We want to provide tools for the Pharo community people
interested in doing ML and AI.
5

6.
How do we position ourselves
Python R Pharo
Data Analysis &
Manipulation
pandas data.frame, dplyr DataFrame
Algebra & Statistics numpy, scipy MASS, SparseM PolyMath
Shallow Learning scikit-learn caret, ml3 pharo-ai
Deep Learning TensorFlow, Keras TensorFlow, Keras TensorFlow, Keras
Visualisation matplotlib ggplot Roassal
6

7.
Algorithms and libraries available in
pharo-ai ecosystem

8.
8
Ready to be used
Work in progress
DataFrame
Datasets
Normalisation
Random Data
Partitioner
Data Generator
Data manipulation
Gaussian
Mixture Models
K-Nearest
Neighbours
Hierarchical
Clustering
K-Means
Clustering
Linear
Regression
Logistic
Regression
Support Vector
Machines
Regression
Principal
Component
Analysis
Metrics
PolyMath
Fast Linear
Algebra
Scienti
fi
c Computation
Edit Distances

9.
9
Ready to be used
Work in progress
Random Forest
Decision Trees
Naive Bayes
Classi
fi
er
Classi
fi
cation
OpenAI
OpenAI Binding
Graph Algorithms
Graph Algorithms
A-Priori
Data Mining
Natural Language Processing
Stop Words
N-Gram
Language Model
Spelling
Correction
TF-IDF

10.
10
Ready to be used
Work in progress
DataFrame
Datasets
Normalisati
on
Random Data
Data
Generator
Data
Gaussian
Mixture Models
K-
Nearest
Hierarchical
Clustering
K-Means
Clustering
Linear
Regression
Logistic
Regression
Support
Vector Machines
Regression
Principal
Component
Metrics
PolyMath
Fast
Linear Algebra
Scienti
fi
c Computation
Edit
Distances
Random
Forest
Decision
Trees
Naive
Bayes
Classi
fi
ca
OpenAI
OpenAI
Graph
Algorithms
Graph
A-Priori
Data Mining
Natural Language
Stop Words
N-
Gram
Spelling
Correction
TF-IDF

11.
Inspector Extensions for DataFrame
11

12.
Inspector Extensions for DataFrame
12

13.
Other Machine Learning Projects
13

14.
pharo-ai wiki
14

15.
Linear Regression benchmarks

16.
16
Dataset size
Name Columns Rows Size
Small 200,000 20 82 Mb
Medium 1,000,000 20 411 Mb
Large 5,000,000 20 2,06 Gb

17.
17
Pharo & LAPACK vs Scikit-learn
Time
in
seconds
(less
is
better)
0
4
8
12
16
Pharo & LAPACK scikit-learn
Small Dataset Medium Dataset Large Dataset
8x faster
5x faster
8x faster

18.
18
Pure Pharo vs Python
Time
in
seconds
(less
is
better)
0
10000
20000
30000
40000
Pure Pharo Pure Python
Small Dataset Medium Dataset
5x faster
15x faster

19.
19
Time in seconds (less is better)
Small
Medium
Pure Pharo vs Pharo & LAPACK
1820 times faster
2103 times faster
Pharo Pharo & LAPACK
19min40sec 1h44min
0.6sec 2.9sec
Help

20.
Visit Us !
Play, Use, and Contribute
20
https://github.com/pharo-ai/wiki
pharo-ai wiki:
Start here