Reproducing Rise of AI over 8 decades plot

Python

DataViz

Our World in Data

#2024plotnineContest

Reproduction of the Our World in Data Plot on the Rise of AI over 8 decades

Author

Affiliation

Dr. Greg Chism

School of Information, University of Arizona

Introduction

This project uses {plotnine} to reproduce my {ggplot2} version of The rise of artificial intelligence over the last 8 decades (Roser 2022) plot by Our World in Data. The plot has some slight differences based on the data used in my version vs the original.

The original plot is licensed under CC-BY by the authors Charlie Giattino, Edouard Mathieu, and Max Roser. This project is licensed under the same.

The data is taken from (Sevilla et al. 2022) – Compute Trends Across Three eras of Machine Learning. Published in arXiv on March 9, 2022. The data are freely available to the public. Note that the available data to the public continues to be updated.

The original plot has fewer data points, likely due to additions by the data authors after the original article’s release.

Challenges

tl;dr

I had not used {plotnine} before, and my Python familiarity is adequate at best. Beyond this inherent difficulty, here are a few observations I made about {plotnine}:

There are some clearly missing components from {ggplot2}, due to {plotnine} being built using {matplotlib} as an underlying framework.
- coord_polar is currently not available (GitHub Issue) - I didn’t need this, but I had several other plot ideas that weren’t possible.
- geom_curve is currently not available (GitHub Issue)
- It seems impossible to include {ggtext} capabilities (GitHub Issue)
- unit is not possible given the current backend (GitHub Issue)
- geom_segment doesn’t seem to work the same as in {ggplot2}: if the size value was too small, it looks like a "dotted" line type reverted to solid. Maybe this is just due to size differences between the parameters in {plotnine} vs {ggplot2}
- Note that function names are different - e.g., {ggplot2}: theme(axis.text.y = …) vs. {plotnine}: theme(axis_text_y = …). This was cumbersome at first…
The documentation is not clear in several instances
- Adding ha = "right" for text alignment wasn’t clear to me until I found it randomly while searching for something else. (e.g., plotnine.annotate)
- There are great examples in geoms_* (e.g., plotnine.geom_segment), but none in several other geoms_* (e.g., plotnine.geom_errorbar) - I realize that this is a primary goal of the {plotnine} competition.
There are several differences between {plotnine} and {ggplot2}, which is understandable, but some are a little confusing - e.g., using ha = "right" in plotnine.annotate but not in plotnine.theme elements.
- I could not see how to make plot_margin work like it does in {ggplot2}. It seems like maybe through plot_margin_* options, but I wasn’t patient enough to try this…
- I also could not figure out how to make major and minor grid line changes like through panel_grid_minor_x, so I ended up giving up and drawing them as individual plotnine.geom_segments. This was not ideal, but it worked.
- I lastly could not get axis_ticks to work at all, probably because of theme_void but I’m not certain. I had to create tick marks through geom_segment instead.
I don’t like how Python code is typically styled, so I added spacing around operators and styled code as if it were in R. In particular, I line broke the plot code a lot, so the code is longer, but more readable (e.g., shorter lines of code)

Load required packages

# Data Manipulation and Analysis
import pandas as pd  # Data manipulation and analysis
import numpy as np   # Numerical operations

# Data Cleaning
from skimpy import clean_columns  # Cleaning column names

# Date and Time Parsing
from dateutil import parser    # Parsing dates
from datetime import datetime  # Working with date and time

# Data Visualization
from plotnine import (
    ggplot,            # Creating plots
    aes,               # Aesthetic mappings
    theme,             # Customizing plot themes
    geom_point,        # Scatter plots
    scale_y_log10,     # Logarithmic y-axis scaling
    element_text,      # Customizing text elements in themes
    element_blank,     # Creating blank elements
    scale_color_manual,# Manual color scales
    scale_x_datetime,  # Datetime x-axis scaling
    labs,              # Plot labels
    theme_void,        # Void theme
    geom_segment,      # Drawing segments
    arrow,             # Adding arrows to segments
    annotate           # Adding annotations
)

# Axis Breaks and Formatting
from mizani.breaks import breaks_log  # Logarithmic axis breaks

# Font Management for Plots
import matplotlib.font_manager as font_manager  # Managing fonts in matplotlib (under the hood of plotnine)

# Paths to the fonts within the working directory
playfair_path = 'fonts/Playfair_Display/PlayfairDisplay-VariableFont_wght.ttf'
lato_path = 'fonts/Lato/Lato-Regular.ttf'

# Add Playfair Display and Lato fonts to matplotlib's font manager
font_manager.fontManager.addfont(playfair_path)
font_manager.fontManager.addfont(lato_path)

Dataset Setup

# Load data
data = pd.read_csv("data/trendsInLLMs.csv")

# Clean column names
data = clean_columns(data)

# Select desired columns
data = data[['system', 'publication_date', 'training_compute_flop', 'domain']]

# Drop rows with missing values
data.dropna(inplace=True)

# Convert 'training_compute_flop' to numeric 
data['training_compute_flop'] = pd.to_numeric(data['training_compute_flop'], errors='coerce')

# Convert 'publication_date' to datetime to handle dates
data['publication_date'] = pd.to_datetime(data['publication_date'], errors='coerce')
data['publication_date'] = data['publication_date'].apply(lambda x: x if x <= datetime.today() else x.replace(year = x.year - 100))

# Handle the 'domain' column mapping and reordering
domain_mapping = {
    'Vision': 'Vision', 'Games': 'Games', 'Drawing': 'Drawing', 'Language': 'Language',
    'Speech': 'Language'
}
data['domain'] = data['domain'].map(lambda x: domain_mapping.get(x, 'Other'))
data['domain'] = pd.Categorical(data['domain'], categories = ["Vision", "Games", "Drawing", "Language", "Other"], ordered = True)

Metadata

After the wrangling from above, here is the metadata:

column	dtype	description
`system`	`object`	The AI system in question
`publication_date`	`datetime64[ns]`	Year-Month-Day for when the AI system was published
`training_compute_flop`	`float64`	Total amount of floating point operations used to train the model
`domain`	`category`	Vision, speech, language, games, etc

Examine data

data.head()

	system	publication_date	training_compute_flop	domain
1	Falcon 180B	2023-09-06	3.780000e+24	Language
2	Swift	2023-08-30	5.340000e+16	Other
3	Jais	2023-08-29	3.080000e+22	Language
6	Llama 2	2023-07-18	9.000000e+23	Language
7	Claude 2	2023-07-11	3.300000e+25	Language

data.tail()

	system	publication_date	training_compute_flop	domain
545	ADALINE	1960-06-30	9900.0	Vision
548	Samuel Neural Checkers	1959-07-01	428000000.0	Games
549	Pandemonium (morse)	1959-02-01	600000000.0	Other
550	Perceptron Mark I	1957-01-01	695000.0	Vision
556	Theseus	1950-07-02	40.0	Other

data.info()

<class 'pandas.core.frame.DataFrame'>
Index: 186 entries, 1 to 556
Data columns (total 4 columns):
 #   Column                 Non-Null Count  Dtype         
---  ------                 --------------  -----         
 0   system                 186 non-null    object        
 1   publication_date       186 non-null    datetime64[ns]
 2   training_compute_flop  186 non-null    float64       
 3   domain                 186 non-null    category      
dtypes: category(1), datetime64[ns](1), float64(1), object(1)
memory usage: 6.2+ KB

data.describe()

	publication_date	training_compute_flop
count	186	1.860000e+02
mean	2015-08-08 10:19:21.290322688	4.342449e+23
min	1950-07-02 00:00:00	4.000000e+01
25%	2015-07-06 00:00:00	1.395000e+18
50%	2019-10-23 00:00:00	8.460000e+20
75%	2021-07-27 00:00:00	3.245000e+22
max	2023-09-06 00:00:00	3.300000e+25
std	NaN	2.924188e+24

Insights from the data

Dates range from 1950-07-02 to 2023-09-06
Minimum value for training_compute_flop is 40, maximum value is 3.3e+25
186 observations (systems) with 0 missing values

Reproduced plot & code

Show the code

# Calculate y-axis breaks (log scale)
limits = (10, 1.e+27)
breaks_vals = breaks_log(14)(limits)

# Define y-axis labels
labels_vals = ["10 FLOP", "1,000 FLOP", "100,000 FLOP", "10 million FLOP",
               "1 billion FLOP", "100 billion FLOP", "10 trillion FLOP",
               "1 petaFLOP = 1 quadrillion FLOP", "100 petaFLOP", 
               "10,000 petaFLOP", "1 million petaFLOP", 
               "100 million petaFLOP", "10 billion petaFLOP", 
               "Training computation that was used\nto train the AI systems."]

# Define plot
plot = (
  ggplot(
  data, # Target data
  aes(
    x = 'publication_date',     # x-axis
    y = 'training_compute_flop' # y-axis
  )
  ) +
        # Begins segments to build x-axis grid lines
        geom_segment(
          aes( 
            x = datetime(1940, 1, 1),
            xend = datetime(1940, 1, 1),
            y = 0.1,
            yend = 1e+25
          ),
          size = 0.1,
          linetype = "dotted",
          color = "#eaeaea"
        ) +        
        geom_segment(
          aes(
            x = datetime(1950, 1, 1), 
            xend = datetime(1950, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) + 
        geom_segment(
          aes(
            x = datetime(1960, 1, 1), 
            xend = datetime(1960, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1970, 1, 1), 
            xend = datetime(1970, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1980, 1, 1), 
            xend = datetime(1980, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1990, 1, 1), 
            xend = datetime(1990, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(2000, 1, 1), 
            xend = datetime(2000, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(2010, 1, 1), 
            xend = datetime(2010, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(2020, 1, 1), 
            xend = datetime(2020, 1, 1), 
            y = 0.1, 
            yend = 1e+25
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        # Begins segments to build y-axis grid lines
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6, 
            yend = 6
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) + 
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+2, 
            yend = 6e+2
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+4, 
            yend = 6e+4
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+6, 
            yend = 6e+6
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+8, 
            yend = 6e+8
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+10, 
            yend = 6e+10
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+12, 
            yend = 6e+12
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+14, 
            yend = 6e+14
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+16, 
            yend = 6e+16
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+18, 
            yend = 6e+18
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+20, 
            yend = 6e+20
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+22, 
            yend = 6e+22
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        geom_segment(
          aes(
            x = datetime(1940, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 6e+24, 
            yend = 6e+24
          ), 
          size = 0.1, 
          linetype = "dotted", 
          color = "#eaeaea"
        ) +
        # Line segment for timeline annotations
        geom_segment(
          aes(
            x = datetime(1953, 1, 1), 
            xend = datetime(2025, 1, 1), 
            y = 3.5, 
            yend = 3.5
          ), 
          size = 0.1, 
          linetype = "solid", 
          color = "#666666"
        ) +
        # Scatterplot
        geom_point(
          aes(
            color = 'domain'
          ), 
          size = 2.5,
          alpha = 0.75
        ) + 
        # Begin arrow line segments
        geom_segment(
          aes(
            x = datetime(1980, 4, 1), 
            xend = datetime(1980, 4, 1), 
            y = 1.75e+07, 
            yend = 1.5e+08
            ), 
            arrow = arrow(
              angle = 30,
              length = 0.025,
              type = "closed"
            ), 
            color = "#B4477A", 
            size = 0.25,
            lineend = "butt"
        ) + 
        geom_segment(
          aes(
            x = datetime(2003, 7, 15), 
            xend = datetime(2012, 2, 27), 
            y = 5e+16, 
            yend = 4.5e+17
            ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#B4477A", 
          size = 0.25, 
          lineend = "butt"
        ) +
        geom_segment(
          aes(
            x = datetime(2006, 6, 15), 
            xend = datetime(2019, 6, 19), 
            y = 1.2e+18, 
            yend = 4e+19
            ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#4B946C", 
          size = 0.25, 
          lineend = "butt"
        ) +
        geom_segment(
          aes(
            x = datetime(2009, 6, 15), 
            xend = datetime(2019, 7, 15), 
            y = 1.5e+19, 
            yend = 9e+19
          ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#9674B0", 
          size = 0.25, 
          lineend = "butt"
        ) +
        geom_segment(
          aes(
            x = datetime(2009, 6, 15), 
            xend = datetime(2015, 6, 27), 
            y = 4.5e+20, 
            yend = 1.33e+21
          ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#4B946C", 
          size = 0.25, 
          lineend = "butt"
        ) +
        geom_segment(
          aes(
            x = datetime(2010, 2, 15), 
            xend = datetime(2021, 3, 21), 
            y = 1.5e+22, 
            yend = 7.90e+21
          ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#476589", 
          size = 0.25, 
          lineend = "butt"
        ) +
        geom_segment(
          aes(
            x = datetime(2011, 3, 15), 
            xend = datetime(2021, 1, 5), 
            y = 1.5e+23, 
            yend = 4.7e+22
          ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#D8847C", 
          size = 0.25, 
          lineend = "butt"
        ) +
        geom_segment(
          aes(
            x = datetime(2012, 6, 15), 
            xend = datetime(2020, 1, 28), 
            y = 2.5e+24, 
            yend = 4e+23
          ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#476589", 
          size = 0.25, 
          lineend = "butt"
          ) +
        geom_segment(
          aes(
            x = datetime(2014, 12, 15), 
            xend = datetime(2021, 6, 4), 
            y = 3e+25, 
            yend = 3e+24
          ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#476589", 
          size = 0.25, 
          lineend = "butt"
        ) +
        geom_segment(
          aes(
            x = datetime(2016, 12, 15), 
            xend = datetime(2021, 12, 4), 
            y = 3e+26, 
            yend = 5e+24
          ), 
          arrow = arrow(
            angle = 30, 
            length = 0.025, 
            type = "closed"
          ), 
          color = "#476589", 
          size = 0.25, 
          lineend = "butt"
        ) +
        # Begin annotations for desired points
        annotate(
          geom = "label",
          x = datetime(1951, 6, 5),
          y = 40,
          label = "Theseus: built in 1950 and trained on around 40 floating point operations (FLOP)\nTheseus was a small robotic mouse, developed by Claude Shannon,\nthat could navigate a simple maze and remember its course.",
          size = 8,
          color = "#9674B0",
          label_size = 0,
          ha = "left"
        ) + 
        annotate(
          geom = "label", 
          x = datetime(1961, 6, 1), 
          y = 9900, 
          label = "ADALINE: built in 1960 and trained on aroiund 9,900 FLOP\nAn early single-layer artificial neural network.", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0, 
          ha = "left"
        ) +
        annotate(
          geom = "label", 
          x = datetime(1957, 12, 1), 
          y = 125000, 
          label = "Perceptron Mark I: built in 1957/58, 695,000 FLOP\nRegarded as the first artificial neural network, it could visually distinguish cards marked on the left side\nfrom those marked on the right, but it could not learn to recognize many other patterns.", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0, 
          ha = "left"
        ) +
        annotate(
          geom = "label", 
          x = datetime(1996, 9, 1), 
          y = 1400000000, 
          label = "Fuzzy NN", 
          size = 8, 
          color = "#476589", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1980, 4, 1), 
          y = 1e+07, 
          label = "Neocognitron: built in 1980, 228 million FLOP\nA precursor of modern vision systems. It could recognize\nhandwritten Japanese characters and a few other patterns.", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1959, 2, 1), 
          y = 1.6e+09, 
          label = "Pandemonium (morse)", 
          size = 8, 
          color = "#9674B0", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1959, 10, 1), 
          y = 1.33e+08, 
          label = "Samuel Neural Checkers",
          size = 8, 
          color = "#4B946C", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1986, 6, 6), 
          y = 3.25e+10, 
          label = "NetTalk: 1987; 81 billion FLOP\nNetTalk was able to learn to pronounce some English text by given\ntext as input and matching it to phonetic transcriptions. Among its many\nlimitations, it did not perform the visual recognition of the text itself.", 
          size = 8, 
          color = "#476589", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(1990, 1, 1), 
          y = 2.15e+11, 
          label = "ALVINN", 
          size = 8, 
          color = "#9674B0", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1993, 8, 1), 
          y = 4.34e+10, 
          label = "Zip CNN", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(2000, 8, 18), 
          y = 1.29e+10, 
          label = "System 11", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1986, 8, 18), 
          y = 3.24e+08, 
          label = "Back-propagation", 
          size = 8, 
          color = "#9674B0", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1993, 6, 1), 
          y = 1.2e+08, 
          label = "Innervator", 
          size = 8, 
          color = "#9674B0", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(2004, 6, 1), 
          y = 2.27e+11, 
          label = "RNN for speech", 
          size = 8, 
          color = "#476589", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1991, 1, 1), 
          y = 1.82e+13, 
          label = "TD-Gammon: 1992; 18 trillion FLOP\nTD-Gammon learned to play backgammon at a high\nlevel, just below the top human players of the time.", 
          size = 8, 
          color = "#4B946C", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2002, 6, 1), 
          y = 2.81e+12, 
          label = "LeNet-5", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(2000, 11, 1), 
          y = 2.10e+13, 
          label = "LSTM", 
          size = 8, 
          color = "#476589", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(2001, 12, 8), 
          y = 1.75e+14, 
          label = "Decision tree", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2002, 1, 1), 
          y = 1.3e+15, 
          label = "NPLM", 
          size = 8, 
          color = "#476589", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2003, 7, 15), 
          y = 2e+16, 
          label = "AlexNet: 2012; 470 petaFLOP\nA pivotal early deep learning system, or neural network with many layers, that\ncould recognize images of objects such as dogs and cars at near-human level.", 
          size = 8, 
          color = "#B4477A", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2006, 1, 11), 
          y = 9e+17, 
          label = "MuZero: 2019; 48,000 petaFLOP\nMuZero is a single system that achieved superhuman performance at Go\nchess, and shogi (Japanese chess) - all without ever being told the rules.", 
          size = 8, 
          color = "#4B946C", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2009, 6, 15), 
          y = 2e+19, 
          label = "AlphaFold: 2020; 100,000 petaFLOP\nAlphaFold was a major advance toward solving the protein-folding problem in biology.", 
          size = 8, 
          color = "#9674B0", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2009, 6, 15), 
          y = 3e+20, 
          label = "AlphaGo: 2016; 1.9 million petaFLOP\nAlphaGo defeated 18-time champion Lee Sedol at the ancient and highly\ncomplex board game Go. The best Go players are no longer human.", 
          size = 8, 
          color = "#4B946C", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2010, 6, 15), 
          y = 9e+21, 
          label = "NEO: 2021; 1.1 million petaFLOP\nRecommendation systems like Facebook's NEO determine what you see on\nyour social media feed, online shopping, streaming services, and more.", 
          size = 8, 
          color = "#476589", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2011, 6, 15), 
          y = 2e+23, 
          label = "DALL-E: 2021; 47 million petaFLOP\nDALL-E can generate high-quality images from writteen descriptions.", 
          size = 8, 
          color = "#D8847C", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2012, 6, 15), 
          y = 3e+24, 
          label = "GPT-3: 2021; 314 million petaFLOP\nGPT-3 can produce high-quality text that is\noften indistinguishable from human writing.", 
          size = 8, 
          color = "#476589", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2014, 12, 15), 
          y = 5e+25, 
          label = "PaLM: 2022; built in 2022 and trained on 2.5 petaFLOP\nPaLM can generate high-quality text, explain some jokes, cause & effect, and more.", 
          size = 8, 
          color = "#476589", 
          label_size = 0, 
          ha = "right"
        ) +
        annotate(
          geom = "label", 
          x = datetime(2016, 12, 15), 
          y = 5e+26, 
          label = "Minerva: 2022; built in 2022 and trained on 2.7 petaFLOP\nMinerva can solve complex mathematical problems at the college level.", 
          size = 8, 
          color = "#476589", 
          label_size = 0, 
          ha = "right"
        ) +
        # Begins timeline annotations above the x-axis
        annotate(
          geom = "label", 
          x = datetime(1940, 1, 1), 
          y = 3.5, 
          label = "Information Age Begins", 
          size = 9.5, 
          color = "#666666", 
          label_size = 0, 
          ha = "left"
        ) + 
        annotate(
          geom = "label", 
          x = datetime(1940, 1, 1), 
          y = 0.75, 
          label = "The first electronic computers\nwere developed in the 1940s", 
          size = 8, 
          color = "#666666", 
          label_size = 0, 
          ha = "left"
        ) +
        annotate(
          geom = "label", 
          x = datetime(1980, 1, 1), 
          y = 3.5, 
          label = "Pre Deep Learning Era", 
          size = 9.5, 
          color = "#666666", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(1980, 1, 1), 
          y = 0.75, 
          label = "Training computation grew in line with\nMoore's law, doubling every 20 months", 
          size = 8, 
          color = "#666666", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(2018, 1, 1),
          y = 3.5, 
          label = "Deep Learning Era", 
          size = 9.5, 
          color = "#666666", 
          label_size = 0
        ) +
        annotate(
          geom = "label", 
          x = datetime(2025, 1, 1), 
          y = 0.75, 
          label = "Increases in training computation accelerated,\ndoubling roughly every 6 months", 
          size = 8, 
          color = "#666666", 
          label_size = 0, 
          ha = "right"
        ) +
        # Scale the y-axis as logarithmic
        scale_y_log10(
          breaks = breaks_vals,
          labels = labels_vals
        ) + 
        # Manually assign colors to Domain categories
        scale_color_manual(
          values = ["#B4477A", "#4B946C", "#D8847C", "#476589", "#9674B0"],
          labels = ["Vision", "Games", "Drawing", "Language", "Other"]
        ) +
        # Scale the x-axis as datatime
        scale_x_datetime(
          date_breaks = "10 years",
          date_labels = "%Y"
        ) + 
        # Set titles within the plot
        labs(
          title = "The rise of artificial intelligence over the last 8 decades: As training\ncomputation has increased, AI systems have become more powerful",
          subtitle = "Computation is measured in floating point operators (FLOP). One FLOP is equivalent to one addition,\nsubtraction, multiplication, or division of two decimals. The data is shown on a logarithmic scale:\neach grid line is a 100-fold increase in computation.",
          caption = "Data: Sevilla et al. (2022) | Visual: Greg Chism | Original: OurWorldInData.org - Charlie Giattino, Edouard Mathieu, and Max Roser\nLicensed under CC-BY | #plotnine",
          x = None,
          y = None,
          color = "The color indicates the domain of the AI system:"
        ) + 
        # Set base theme void
        theme_void(
          base_family = "Lato"
        ) + 
        # Manually set individual theme elements
        theme(
          # Plot title
          plot_title = element_text(
            size = 17.5,
            family = "Playfair Display",
            hjust = 0
          ),
          # Plot subtitle
          plot_subtitle = element_text(
            color = "#666666",
            hjust = 0,
            linespacing = 1.25
          ),
          # Plot caption
          plot_caption = element_text(
            color = "#666666",
            size = 8,
            hjust = 0,
            linespacing = 1.25
          ),
          # Legend position
          legend_position = "top",
          # Legend title
          legend_title = element_text(
            size = 10,
            color = "#666666",
            hjust = 1
          ), 
          # Legend text
          legend_text = element_text(
            size = 10,
            color = "#666666"
            ), 
        # x-axis text
        axis_text_x = element_text(
          size = 12,
          color = "#666666",
          vjust = 0
        ),
        # y-axis text
        axis_text_y = element_text(
          size = 12,
          color = "#666666",
          linespacing = 1.25
        ),
        # Figure size
        figure_size = (10, 11)
        ) 
        )

# Show the plot        
ggplot.show(plot)

1: Where the segment begins on the x-axis (datetime)
2: Where the segment ends on the x-axis (datetime)
3: Where the segment begins on the y-axis (log-scale)
4: Where the segment ends on the y-axis (log-scale)
5: Line thickness (size)
6: Line type (dotted)
7: Line color (#eaeaea)
8: Color points by the domain of the AI system
9: Make the point size 2.5
10: Make the alpha (transparency) 0.75 to better view overlap
11: Make the line segment end in an arrow
12: Angle of the arrow (30-degrees)
13: length of the arrow head (0.025)
14: Type of arrow (“closed”)
15: Type of line segment end (“butt”) - see here
16: Set geom type as “label”
17: Where the annotation is on the x-axis (datetime)
18: Where the annotation is on the y-axis (datetime)
19: The label to be printed
20: The (text) size of the annotation
21: The (text) color of the annotation
22: The border line size of the label (reduced to 0)
23: The horizontal justification of the label (left-aligned)
24: Log breaks defined above
25: Labels for each break defined above
26: Manually defined colors
27: Order of labels
28: Set axis breaks to every 10-years
29: Set the lable to years only
30: Set plot title
31: Set plot subtitle
32: Set plot caption
33: No x-axis title
34: No y-axis title
35: Set color legend title
36: Set the base font family to Lato (imported above)
37: Plot title font size
38: Plot title font family as Playfair Display (imported above)
39: Plot title horizontal justification
40: Plot subtitle text color
41: Plot subtitle horizontal justification
42: Plot subtitle line spacing
43: Plot caption text color
44: Plot caption font size
45: Plot caption horizontal justification
46: Plot caption line spacing
47: Legend position to “top”
48: Set legend title font size
49: Set legend title text color
50: Set legend title horizontal
51: Set legend text font size
52: Set legend element text color
53: Set x-axis text font size
54: Set x-axis text color
55: Set x-axis text vertical justification
56: Set y-axis text font size
57: Set y-axis text color
58: Set y-axis line spacing
59: Set to 10 x 11in

Find the code here

Find the original article here

Amendments

I had to make some changes from the original. Some changes were design choices to help reduce clutter. Others were to add additional clarity.

I could not reproduce the axes annotations without literally drawing the entire axis as an annotation (like I did through {ggplot2}. The information is helpful, but also clutters the axis a little too much.
The caption was reduced to save space. Having technically two captions wouldn’t have been possible anyway.
I added a subtitle to summarise the y-axis annotation information. I think this is helpful.
I chose to keep the standard axes text that you can create instead of adding each as an annotation. This better showcases what {plotnine} is capable of by default (e.g., plotnine.theme_void).
I removed the x-axis tick marks to simplify the plot a little. This could have only been done with geom_segment.
I added an annotation where Information Age Begins in 1940. This makes sense, and only really wouldn’t work if the plot was more cluttered (like in the original).

Conclusions

Overall, I found that {plotnine} is not a perfect facsimile of {ggplot2} due to it being built upon the {matplotlib} framework, but its pretty close:

Most of the functionality from my {ggplot2} version was translatable, with some exceptions (see above).
The “heart” of {ggplot2} seems to be in {plotnine}, making it a good way for R users to get into plotting through Python.
All-in-all, I think the plot I produced was pretty similar, so clearly the {plotnine} developers did a great job with the package.

This all being said, I found that working with {plotnine} felt like I was going to break it with every new line. The package feels a little fragile due to this. Some of this however is from my lack of experience in Python.

References

Roser, Max. 2022. “The Brief History of Artificial Intelligence: The World Has Changed Fast — What Might Be Next?” Our World in Data.

Sevilla, Jaime, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn, and Pablo Villalobos. 2022. “Compute Trends Across Three Eras of Machine Learning.” https://doi.org/10.48550/ARXIV.2202.05924.

--- title: Reproducing Rise of AI over 8 decades plot author: - name: Dr. Greg Chism affiliations: - name: School of Information, University of Arizona description: Reproduction of the Our World in Data Plot on the Rise of AI over 8 decades format: html: theme: flatly code-tools: true code-line-numbers: true execute: warning: false message: false error: false toc: true code-annotations: hover categories: - Python - DataViz - AI - Our World in Data - "#2024plotnineContest" bibliography: references.bib --- ## Introduction This project uses {**plotnine**} to reproduce my {**ggplot2**} version of [**The rise of artificial intelligence over the last 8 decades**](https://ourworldindata.org/brief-history-of-ai) [@owid-brief-history-of-ai] plot by [Our World in Data](https://ourworldindata.org/). The plot has some slight differences based on the data used in my version vs the original. The original plot is licensed under [**CC-BY**](https://creativecommons.org/licenses/by/4.0/) by the authors Charlie Giattino, Edouard Mathieu, and Max Roser. This project is licensed under the same. The data is taken from [@sevilla2022] -- [Compute Trends Across Three eras of Machine Learning](https://arxiv.org/pdf/2202.05924.pdf). Published in arXiv on March 9, 2022. The data are [freely available](https://docs.google.com/spreadsheets/d/1AAIebjNsnJj_uKALHbXNfn3_YsT6sHXtCU0q7OIPuc4/edit#gid=0) to the public. Note that the available data to the public continues to be updated. The original plot has fewer data points, likely due to additions by the data authors after the original article's release. ### Challenges {#challenges} **tl;dr** I had not used {**plotnine**} before, and my Python familiarity is adequate at best. Beyond this inherent difficulty, here are a few observations I made about {**plotnine**}: - There are some clearly missing components from {**ggplot2**}, due to {**plotnine**} being built using {**matplotlib**} as an underlying framework. - `coord_polar` is currently not available ([GitHub Issue](https://github.com/has2k1/plotnine/issues/10)) - I didn't need this, but I had several other plot ideas that weren't possible. - `geom_curve` is currently not available ([GitHub Issue](https://github.com/has2k1/plotnine/issues/535)) - It seems impossible to include {**ggtext**} capabilities ([GitHub Issue](https://github.com/has2k1/plotnine/issues/612)) - `unit` is not possible given the current backend ([GitHub Issue](https://github.com/has2k1/plotnine/issues/612)) - `geom_segment` doesn't seem to work the same as in {**ggplot2**}: if the `size` value was too small, it looks like a `"dotted"` line type reverted to solid. Maybe this is just due to size differences between the parameters in {**plotnine**} vs {**ggplot2**} - [Note]{.underline} that function names are different - e.g., {**ggplot2**}: `theme(axis.text.y = …)` vs. {**plotnine**}: `theme(axis_text_y = …)`. This was cumbersome at first... - The documentation is not clear in several instances - Adding `ha = "right"` for text alignment wasn't clear to me until I found it randomly while searching for something else. (e.g., [`plotnine.annotate`](https://plotnine.org/reference/annotate.html)) - There are great examples in `geoms_*` (e.g., [`plotnine.geom_segment`](https://plotnine.org/reference/geom_segment.html#an-elaborate-range-plot)), but none in several other `geoms_*` (e.g., [`plotnine.geom_errorbar`](https://plotnine.org/reference/geom_errorbar.html#plotnine.geom_errorbar)) - I realize that this is a primary goal of the {**plotnine**} competition. - There are several differences between {**plotnine**} and {**ggplot2**}, which is understandable, but some are a little confusing - e.g., using `ha = "right"` in `plotnine.annotate` but not in `plotnine.theme` elements. - I could not see how to make [`plot_margin`](https://plotnine.org/reference/plot_margin.html) work like it does in {**ggplot2**}. It seems like maybe through `plot_margin_*` options, but I wasn't patient enough to try this... - I also could not figure out how to make major and minor grid line changes like through [`panel_grid_minor_x`](https://plotnine.org/reference/panel_grid_minor_x.html), so I ended up giving up and drawing them as individual [`plotnine.geom_segment`](https://plotnine.org/reference/geom_segment.html#an-elaborate-range-plot)s. This was not ideal, but it worked. - I lastly could not get [`axis_ticks`](https://plotnine.org/reference/axis_ticks.html) to work at all, probably because of `theme_void` but I'm not certain. I had to create tick marks through `geom_segment` instead. - I don't like how Python code is typically styled, so I added spacing around operators and styled code as if it were in R. In particular, I line broke the plot code a lot, so the code is longer, but more readable (e.g., shorter lines of code) ## Load required packages ```{python} #| label: load-pkgs #| message: false # Data Manipulation and Analysis import pandas as pd # Data manipulation and analysis import numpy as np # Numerical operations # Data Cleaning from skimpy import clean_columns # Cleaning column names # Date and Time Parsing from dateutil import parser # Parsing dates from datetime import datetime # Working with date and time # Data Visualization from plotnine import ( ggplot, # Creating plots aes, # Aesthetic mappings theme, # Customizing plot themes geom_point, # Scatter plots scale_y_log10, # Logarithmic y-axis scaling element_text, # Customizing text elements in themes element_blank, # Creating blank elements scale_color_manual,# Manual color scales scale_x_datetime, # Datetime x-axis scaling labs, # Plot labels theme_void, # Void theme geom_segment, # Drawing segments arrow, # Adding arrows to segments annotate # Adding annotations ) # Axis Breaks and Formatting from mizani.breaks import breaks_log # Logarithmic axis breaks # Font Management for Plots import matplotlib.font_manager as font_manager # Managing fonts in matplotlib (under the hood of plotnine) # Paths to the fonts within the working directory playfair_path = 'fonts/Playfair_Display/PlayfairDisplay-VariableFont_wght.ttf' lato_path = 'fonts/Lato/Lato-Regular.ttf' # Add Playfair Display and Lato fonts to matplotlib's font manager font_manager.fontManager.addfont(playfair_path) font_manager.fontManager.addfont(lato_path) ``` ## Dataset Setup ```{python} #| label: prepare-data #| message: FALSE # Load data data = pd.read_csv("data/trendsInLLMs.csv") # Clean column names data = clean_columns(data) # Select desired columns data = data[['system', 'publication_date', 'training_compute_flop', 'domain']] # Drop rows with missing values data.dropna(inplace=True) # Convert 'training_compute_flop' to numeric data['training_compute_flop'] = pd.to_numeric(data['training_compute_flop'], errors='coerce') # Convert 'publication_date' to datetime to handle dates data['publication_date'] = pd.to_datetime(data['publication_date'], errors='coerce') data['publication_date'] = data['publication_date'].apply(lambda x: x if x <= datetime.today() else x.replace(year = x.year - 100)) # Handle the 'domain' column mapping and reordering domain_mapping = { 'Vision': 'Vision', 'Games': 'Games', 'Drawing': 'Drawing', 'Language': 'Language', 'Speech': 'Language' } data['domain'] = data['domain'].map(lambda x: domain_mapping.get(x, 'Other')) data['domain'] = pd.Categorical(data['domain'], categories = ["Vision", "Games", "Drawing", "Language", "Other"], ordered = True) ``` ### Metadata After the wrangling from above, here is the metadata: | column | dtype | description | |-------------------------|------------------|-------------------------------------------------------------------| | `system` | `object` | The AI system in question | | `publication_date` | `datetime64[ns]` | Year-Month-Day for when the AI system was published | | `training_compute_flop` | `float64` | Total amount of floating point operations used to train the model | | `domain` | `category` | Vision, speech, language, games, etc | ### Examine data ::: panel-tabset ## Head + Tail ```{python} #| label: head data.head() ``` ```{python} #| label: tail data.tail() ``` ## Info ```{python} #| label: info data.info() ``` ## Describe ```{python} #| label: describe data.describe() ``` ::: ### Insights from the data 1. Dates range from 1950-07-02 to 2023-09-06 2. Minimum value for `training_compute_flop` is 40, maximum value is 3.3e+25 3. 186 observations (systems) with 0 missing values ## Reproduced plot & code ::: panel-tabset ## {plotnine} ```{python} #| label: plotnine-version #| code-fold: true #| code-summary: "Show the code" #| fig-height: 15 #| fig-width: 10 # Calculate y-axis breaks (log scale) limits = (10, 1.e+27) breaks_vals = breaks_log(14)(limits) # Define y-axis labels labels_vals = ["10 FLOP", "1,000 FLOP", "100,000 FLOP", "10 million FLOP", "1 billion FLOP", "100 billion FLOP", "10 trillion FLOP", "1 petaFLOP = 1 quadrillion FLOP", "100 petaFLOP", "10,000 petaFLOP", "1 million petaFLOP", "100 million petaFLOP", "10 billion petaFLOP", "Training computation that was used\nto train the AI systems."] # Define plot plot = ( ggplot( data, # Target data aes( x = 'publication_date', # x-axis y = 'training_compute_flop' # y-axis ) ) + # Begins segments to build x-axis grid lines geom_segment( aes( x = datetime(1940, 1, 1), #<1> xend = datetime(1940, 1, 1),#<2> y = 0.1, #<3> yend = 1e+25 #<4> ), size = 0.1, #<5> linetype = "dotted", #<6> color = "#eaeaea" #<7> ) + geom_segment( aes( x = datetime(1950, 1, 1), xend = datetime(1950, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1960, 1, 1), xend = datetime(1960, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1970, 1, 1), xend = datetime(1970, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1980, 1, 1), xend = datetime(1980, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1990, 1, 1), xend = datetime(1990, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(2000, 1, 1), xend = datetime(2000, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(2010, 1, 1), xend = datetime(2010, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(2020, 1, 1), xend = datetime(2020, 1, 1), y = 0.1, yend = 1e+25 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + # Begins segments to build y-axis grid lines geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6, yend = 6 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+2, yend = 6e+2 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+4, yend = 6e+4 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+6, yend = 6e+6 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+8, yend = 6e+8 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+10, yend = 6e+10 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+12, yend = 6e+12 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+14, yend = 6e+14 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+16, yend = 6e+16 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+18, yend = 6e+18 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+20, yend = 6e+20 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+22, yend = 6e+22 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + geom_segment( aes( x = datetime(1940, 1, 1), xend = datetime(2025, 1, 1), y = 6e+24, yend = 6e+24 ), size = 0.1, linetype = "dotted", color = "#eaeaea" ) + # Line segment for timeline annotations geom_segment( aes( x = datetime(1953, 1, 1), xend = datetime(2025, 1, 1), y = 3.5, yend = 3.5 ), size = 0.1, linetype = "solid", color = "#666666" ) + # Scatterplot geom_point( aes( color = 'domain' #<8> ), size = 2.5, #<9> alpha = 0.75 #<10> ) + # Begin arrow line segments geom_segment( aes( x = datetime(1980, 4, 1), xend = datetime(1980, 4, 1), y = 1.75e+07, yend = 1.5e+08 ), arrow = arrow( #<11> angle = 30, #<12> length = 0.025, #<13> type = "closed" #<14> ), color = "#B4477A", size = 0.25, lineend = "butt" #<15> ) + geom_segment( aes( x = datetime(2003, 7, 15), xend = datetime(2012, 2, 27), y = 5e+16, yend = 4.5e+17 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#B4477A", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2006, 6, 15), xend = datetime(2019, 6, 19), y = 1.2e+18, yend = 4e+19 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#4B946C", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2009, 6, 15), xend = datetime(2019, 7, 15), y = 1.5e+19, yend = 9e+19 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#9674B0", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2009, 6, 15), xend = datetime(2015, 6, 27), y = 4.5e+20, yend = 1.33e+21 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#4B946C", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2010, 2, 15), xend = datetime(2021, 3, 21), y = 1.5e+22, yend = 7.90e+21 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#476589", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2011, 3, 15), xend = datetime(2021, 1, 5), y = 1.5e+23, yend = 4.7e+22 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#D8847C", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2012, 6, 15), xend = datetime(2020, 1, 28), y = 2.5e+24, yend = 4e+23 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#476589", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2014, 12, 15), xend = datetime(2021, 6, 4), y = 3e+25, yend = 3e+24 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#476589", size = 0.25, lineend = "butt" ) + geom_segment( aes( x = datetime(2016, 12, 15), xend = datetime(2021, 12, 4), y = 3e+26, yend = 5e+24 ), arrow = arrow( angle = 30, length = 0.025, type = "closed" ), color = "#476589", size = 0.25, lineend = "butt" ) + # Begin annotations for desired points annotate( geom = "label", #<16> x = datetime(1951, 6, 5), #<17> y = 40, #<18> label = "Theseus: built in 1950 and trained on around 40 floating point operations (FLOP)\nTheseus was a small robotic mouse, developed by Claude Shannon,\nthat could navigate a simple maze and remember its course.", #<19> size = 8, #<20> color = "#9674B0", #<21> label_size = 0, #<22> ha = "left" #<23> ) + annotate( geom = "label", x = datetime(1961, 6, 1), y = 9900, label = "ADALINE: built in 1960 and trained on aroiund 9,900 FLOP\nAn early single-layer artificial neural network.", size = 8, color = "#B4477A", label_size = 0, ha = "left" ) + annotate( geom = "label", x = datetime(1957, 12, 1), y = 125000, label = "Perceptron Mark I: built in 1957/58, 695,000 FLOP\nRegarded as the first artificial neural network, it could visually distinguish cards marked on the left side\nfrom those marked on the right, but it could not learn to recognize many other patterns.", size = 8, color = "#B4477A", label_size = 0, ha = "left" ) + annotate( geom = "label", x = datetime(1996, 9, 1), y = 1400000000, label = "Fuzzy NN", size = 8, color = "#476589", label_size = 0 ) + annotate( geom = "label", x = datetime(1980, 4, 1), y = 1e+07, label = "Neocognitron: built in 1980, 228 million FLOP\nA precursor of modern vision systems. It could recognize\nhandwritten Japanese characters and a few other patterns.", size = 8, color = "#B4477A", label_size = 0 ) + annotate( geom = "label", x = datetime(1959, 2, 1), y = 1.6e+09, label = "Pandemonium (morse)", size = 8, color = "#9674B0", label_size = 0 ) + annotate( geom = "label", x = datetime(1959, 10, 1), y = 1.33e+08, label = "Samuel Neural Checkers", size = 8, color = "#4B946C", label_size = 0 ) + annotate( geom = "label", x = datetime(1986, 6, 6), y = 3.25e+10, label = "NetTalk: 1987; 81 billion FLOP\nNetTalk was able to learn to pronounce some English text by given\ntext as input and matching it to phonetic transcriptions. Among its many\nlimitations, it did not perform the visual recognition of the text itself.", size = 8, color = "#476589", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(1990, 1, 1), y = 2.15e+11, label = "ALVINN", size = 8, color = "#9674B0", label_size = 0 ) + annotate( geom = "label", x = datetime(1993, 8, 1), y = 4.34e+10, label = "Zip CNN", size = 8, color = "#B4477A", label_size = 0 ) + annotate( geom = "label", x = datetime(2000, 8, 18), y = 1.29e+10, label = "System 11", size = 8, color = "#B4477A", label_size = 0 ) + annotate( geom = "label", x = datetime(1986, 8, 18), y = 3.24e+08, label = "Back-propagation", size = 8, color = "#9674B0", label_size = 0 ) + annotate( geom = "label", x = datetime(1993, 6, 1), y = 1.2e+08, label = "Innervator", size = 8, color = "#9674B0", label_size = 0 ) + annotate( geom = "label", x = datetime(2004, 6, 1), y = 2.27e+11, label = "RNN for speech", size = 8, color = "#476589", label_size = 0 ) + annotate( geom = "label", x = datetime(1991, 1, 1), y = 1.82e+13, label = "TD-Gammon: 1992; 18 trillion FLOP\nTD-Gammon learned to play backgammon at a high\nlevel, just below the top human players of the time.", size = 8, color = "#4B946C", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2002, 6, 1), y = 2.81e+12, label = "LeNet-5", size = 8, color = "#B4477A", label_size = 0 ) + annotate( geom = "label", x = datetime(2000, 11, 1), y = 2.10e+13, label = "LSTM", size = 8, color = "#476589", label_size = 0 ) + annotate( geom = "label", x = datetime(2001, 12, 8), y = 1.75e+14, label = "Decision tree", size = 8, color = "#B4477A", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2002, 1, 1), y = 1.3e+15, label = "NPLM", size = 8, color = "#476589", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2003, 7, 15), y = 2e+16, label = "AlexNet: 2012; 470 petaFLOP\nA pivotal early deep learning system, or neural network with many layers, that\ncould recognize images of objects such as dogs and cars at near-human level.", size = 8, color = "#B4477A", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2006, 1, 11), y = 9e+17, label = "MuZero: 2019; 48,000 petaFLOP\nMuZero is a single system that achieved superhuman performance at Go\nchess, and shogi (Japanese chess) - all without ever being told the rules.", size = 8, color = "#4B946C", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2009, 6, 15), y = 2e+19, label = "AlphaFold: 2020; 100,000 petaFLOP\nAlphaFold was a major advance toward solving the protein-folding problem in biology.", size = 8, color = "#9674B0", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2009, 6, 15), y = 3e+20, label = "AlphaGo: 2016; 1.9 million petaFLOP\nAlphaGo defeated 18-time champion Lee Sedol at the ancient and highly\ncomplex board game Go. The best Go players are no longer human.", size = 8, color = "#4B946C", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2010, 6, 15), y = 9e+21, label = "NEO: 2021; 1.1 million petaFLOP\nRecommendation systems like Facebook's NEO determine what you see on\nyour social media feed, online shopping, streaming services, and more.", size = 8, color = "#476589", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2011, 6, 15), y = 2e+23, label = "DALL-E: 2021; 47 million petaFLOP\nDALL-E can generate high-quality images from writteen descriptions.", size = 8, color = "#D8847C", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2012, 6, 15), y = 3e+24, label = "GPT-3: 2021; 314 million petaFLOP\nGPT-3 can produce high-quality text that is\noften indistinguishable from human writing.", size = 8, color = "#476589", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2014, 12, 15), y = 5e+25, label = "PaLM: 2022; built in 2022 and trained on 2.5 petaFLOP\nPaLM can generate high-quality text, explain some jokes, cause & effect, and more.", size = 8, color = "#476589", label_size = 0, ha = "right" ) + annotate( geom = "label", x = datetime(2016, 12, 15), y = 5e+26, label = "Minerva: 2022; built in 2022 and trained on 2.7 petaFLOP\nMinerva can solve complex mathematical problems at the college level.", size = 8, color = "#476589", label_size = 0, ha = "right" ) + # Begins timeline annotations above the x-axis annotate( geom = "label", x = datetime(1940, 1, 1), y = 3.5, label = "Information Age Begins", size = 9.5, color = "#666666", label_size = 0, ha = "left" ) + annotate( geom = "label", x = datetime(1940, 1, 1), y = 0.75, label = "The first electronic computers\nwere developed in the 1940s", size = 8, color = "#666666", label_size = 0, ha = "left" ) + annotate( geom = "label", x = datetime(1980, 1, 1), y = 3.5, label = "Pre Deep Learning Era", size = 9.5, color = "#666666", label_size = 0 ) + annotate( geom = "label", x = datetime(1980, 1, 1), y = 0.75, label = "Training computation grew in line with\nMoore's law, doubling every 20 months", size = 8, color = "#666666", label_size = 0 ) + annotate( geom = "label", x = datetime(2018, 1, 1), y = 3.5, label = "Deep Learning Era", size = 9.5, color = "#666666", label_size = 0 ) + annotate( geom = "label", x = datetime(2025, 1, 1), y = 0.75, label = "Increases in training computation accelerated,\ndoubling roughly every 6 months", size = 8, color = "#666666", label_size = 0, ha = "right" ) + # Scale the y-axis as logarithmic scale_y_log10( breaks = breaks_vals, #<24> labels = labels_vals #<25> ) + # Manually assign colors to Domain categories scale_color_manual( values = ["#B4477A", "#4B946C", "#D8847C", "#476589", "#9674B0"], #<26> labels = ["Vision", "Games", "Drawing", "Language", "Other"] #<27> ) + # Scale the x-axis as datatime scale_x_datetime( date_breaks = "10 years", #<28> date_labels = "%Y" #<29> ) + # Set titles within the plot labs( title = "The rise of artificial intelligence over the last 8 decades: As training\ncomputation has increased, AI systems have become more powerful", #<30> subtitle = "Computation is measured in floating point operators (FLOP). One FLOP is equivalent to one addition,\nsubtraction, multiplication, or division of two decimals. The data is shown on a logarithmic scale:\neach grid line is a 100-fold increase in computation.", #<31> caption = "Data: Sevilla et al. (2022) | Visual: Greg Chism | Original: OurWorldInData.org - Charlie Giattino, Edouard Mathieu, and Max Roser\nLicensed under CC-BY | #plotnine", #<32> x = None, #<33> y = None, #<34> color = "The color indicates the domain of the AI system:" #<35> ) + # Set base theme void theme_void( base_family = "Lato" #<36> ) + # Manually set individual theme elements theme( # Plot title plot_title = element_text( size = 17.5, #<37> family = "Playfair Display", #<38> hjust = 0 #<39> ), # Plot subtitle plot_subtitle = element_text( color = "#666666", #<40> hjust = 0, #<41> linespacing = 1.25 #<42> ), # Plot caption plot_caption = element_text( color = "#666666", #<43> size = 8, #<44> hjust = 0, #<45> linespacing = 1.25 #<46> ), # Legend position legend_position = "top", #<47> # Legend title legend_title = element_text( size = 10, #<48> color = "#666666", #<49> hjust = 1 #<50> ), # Legend text legend_text = element_text( size = 10, #<51> color = "#666666" #<52> ), # x-axis text axis_text_x = element_text( size = 12, #<53> color = "#666666", #<54> vjust = 0 #<55> ), # y-axis text axis_text_y = element_text( size = 12, #<56> color = "#666666", #<57> linespacing = 1.25 #<58> ), # Figure size figure_size = (10, 11) #<59> ) ) # Show the plot ggplot.show(plot) ``` 1. Where the segment begins on the x-axis (datetime) 2. Where the segment ends on the x-axis (datetime) 3. Where the segment begins on the y-axis (log-scale) 4. Where the segment ends on the y-axis (log-scale) 5. Line thickness (size) 6. Line type (dotted) 7. Line color (#eaeaea) 8. Color points by the domain of the AI system 9. Make the point size 2.5 10. Make the alpha (transparency) 0.75 to better view overlap 11. Make the line segment end in an arrow 12. Angle of the arrow (30-degrees) 13. length of the arrow head (0.025) 14. Type of arrow ("closed") 15. Type of line segment end ("butt") - see [here](https://ggplot2.tidyverse.org/reference/geom_segment.html#ref-examples) 16. Set geom type as "label" 17. Where the annotation is on the x-axis (datetime) 18. Where the annotation is on the y-axis (datetime) 19. The label to be printed 20. The (text) size of the annotation 21. The (text) color of the annotation 22. The border line size of the label (reduced to 0) 23. The horizontal justification of the label (left-aligned) 24. Log breaks defined above 25. Labels for each break defined above 26. Manually defined colors 27. Order of labels 28. Set axis breaks to every 10-years 29. Set the lable to years only 30. Set plot title 31. Set plot subtitle 32. Set plot caption 33. No x-axis title 34. No y-axis title 35. Set color legend title 36. Set the base font family to Lato (imported above) 37. Plot title font size 38. Plot title font family as Playfair Display (imported above) 39. Plot title horizontal justification 40. Plot subtitle text color 41. Plot subtitle horizontal justification 42. Plot subtitle line spacing 43. Plot caption text color 44. Plot caption font size 45. Plot caption horizontal justification 46. Plot caption line spacing 47. Legend position to "top" 48. Set legend title font size 49. Set legend title text color 50. Set legend title horizontal 51. Set legend text font size 52. Set legend element text color 53. Set x-axis text font size 54. Set x-axis text color 55. Set x-axis text vertical justification 56. Set y-axis text font size 57. Set y-axis text color 58. Set y-axis line spacing 59. Set to 10 x 11in ## {ggplot2} [**Find the code here**](https://gchism94.github.io/riseOfAIPlot/) ![](images/ggplot2-version.png){fig-align="center"} ## Original [**Find the original article here**](https://ourworldindata.org/brief-history-of-ai) ![](https://ourworldindata.org/uploads/2022/12/ai-training-computation-3.png){fig-align="center"} ::: ## Amendments I had to make some changes from the original. Some changes were design choices to help reduce clutter. Others were to add additional clarity. - I could not reproduce the axes annotations without literally drawing the entire axis as an annotation (like I did through {**ggplot2**}. The information is helpful, but also clutters the axis a little too much. - The caption was reduced to save space. Having technically two captions wouldn't have been possible anyway. - I added a subtitle to summarise the y-axis annotation information. I think this is helpful. - I chose to keep the standard axes text that you can create instead of adding each as an annotation. This better showcases what {**plotnine**} is capable of by default (e.g., `plotnine.theme_void`). - I removed the x-axis tick marks to simplify the plot a little. This could have only been done with `geom_segment`. - I added an annotation where Information Age Begins in 1940. This makes sense, and only really wouldn't work if the plot was more cluttered (like in the original). ## Conclusions Overall, I found that {**plotnine**} is not a perfect facsimile of {**ggplot2**} due to it being built upon the {**matplotlib**} framework, but its pretty close: - Most of the functionality from my {**ggplot2**} version was translatable, with some exceptions ([see above](#challenges)). - The "heart" of {**ggplot2**} seems to be in {**plotnine**}, making it a good way for R users to get into plotting through Python. - All-in-all, I think the plot I produced was pretty similar, so clearly the {**plotnine**} developers did a great job with the package. This all being said, I found that working with {**plotnine**} felt like I was going to break it with every new line. The package feels a little fragile due to this. Some of this however is from my lack of experience in Python.