M ONTRÉAL .

AI ACADEMY: A RTIFICIAL I NTELLIGENCE 101

F IRST W ORLD -C LASS OVERVIEW OF AI FOR A LL
VIP AI 101 C HEAT S HEET

A P REPRINT

Vincent Boucher∗
MONTRÉAL.AI
Montreal, Quebec, Canada
[email protected]

November 5, 2020

A BSTRACT
For the purpose of entrusting all sentient beings with powerful AI tools to learn, deploy and scale AI
in order to enhance their prosperity, to settle planetary-scale problems and to inspire those who, with
AI, will shape the 21st Century, MONTRÉAL.AI introduces this VIP AI 101 CheatSheet for All.

*MONTRÉAL.AI is preparing a global network of education centers.

**ALL OF EDUCATION, FOR ALL. MONTRÉAL.AI is developing a teacher (Saraswati AI) and
an agent learning to orchestrate synergies amongst academic disciplines (Polymatheia AI).
Curated Open-Source Codes and Science: http://www.academy.montreal.ai/.

Keywords AI-First · Artificial Intelligence · Deep Learning · Reinforcement Learning · Symbolic AI

1 AI-First
TODAY’S ARTIFICIAL INTELLIGENCE IS POWERFUL AND ACCESSIBLE TO ALL. AI is capable of transform-
ing industries and opens up a world of new possibilities. What’s important is what you do with AI and how you
embrace it. To pioneer AI-First innovations advantages: start by exploring how to apply AI in ways never thought of.
The Emerging Rules of the AI-First Era: Search and Learning.
"Search and learning are general purpose methods that continue to scale with increased computation, even as the
available computation becomes very great." — Richard Sutton in The Bitter Lesson
The Best Way Forward For AI2 .
"... so far as I’m concerned, system 1 certainly knows language, understands language... system 2... it does involve
certain manipulation of symbols... Gary Marcus ... Gary proposes something that seems very natural... a hybrid
architecture... I’m influenced by him... if you look introspectively at the way the mind works... you’d get to that
distinction between implicit and explicit... explicit looks like symbols." — Nobel Laureate Danny Kahneman at
AAAI-20 Fireside Chat with Daniel Kahneman https://vimeo.com/390814190
In The Next Decade in AI 3 , Gary Marcus proposes a hybrid, knowledge-driven, reasoning-based approach, centered
around cognitive models, that could provide the substrate for a richer, more robust AI than is currently possible.
∗
Founding Chairman at MONTRÉAL.AI http://www.montreal.ai and QUÉBEC.AI http://www.quebec.ai.
2
https://montrealartificialintelligence.com/aidebate/
3
https://arxiv.org/abs/2002.06177v3
 A PREPRINT - N OVEMBER 5, 2020

2 Getting Started

Tinker with neural networks in the browser with TensorFlow Playground http://playground.tensorflow.org/.

• AI Paygrades https://aipaygrad.es/.
• CS231n Python Tutorial With Google Colab4 .
• Learn with Google AI https://ai.google/education/.
• Made With ML Topics https://madewithml.com/topics/.
• One Place for Everything AI https://aihub.cloud.google.com/.
• Deep Learning Drizzle https://deep-learning-drizzle.github.io.
• Google Dataset Search (Blog5 ) https://datasetsearch.research.google.com.
• AI Literacy for K-12 School Children https://aieducation.mit.edu/resources.
• Learning resources from DeepMind https://deepmind.com/learning-resources.
• Papers With Code (Learn Python 3 in Y minutes6 ) https://paperswithcode.com/state-of-the-art.

"Dataset Search has indexed almost 25 million of these datasets, giving you a single place to search for datasets and
find links to where the data is." — Natasha Noy

The Measure of Intelligence (Abstraction and Reasoning Corpus7 ) https://arxiv.org/abs/1911.01547.

v Growing Neural Cellular Automata, Mordvintsev et al. https://distill.pub/2020/growing-ca/.

2.1 In the Cloud

Colab 8 . Practice Immediately 9 . Labs10 : Introduction to Deep Learning (MIT 6.S191)

• Free GPU compute via Colab https://colab.research.google.com/notebooks/welcome.ipynb.

• Colab can open notebooks directly from GitHub by simply replacing "http://github.com" with
"http://colab.research.google.com/github/ " in the notebook URL.
• Colab Pro https://colab.research.google.com/signup.

2.2 On a Local Machine

JupyterLab is an interactive development environment for working with notebooks, code and data 11 .

• Install Anaconda https://www.anaconda.com/download/ and launch ‘Anaconda Navigator’

• Update Jupyterlab and launch the application. Under Notebook, click on ‘Python 3’

IDE: Visual Studio Code https://code.visualstudio.com/.

"If we truly reach AI, it will let us know." — Garry Kasparov

4
https://colab.research.google.com/github/cs231n/cs231n.github.io/blob/master/python-colab.ipynb
5
https://blog.google/products/search/discovering-millions-datasets-web/
6
https://learnxinyminutes.com/docs/python3/
7
https://github.com/fchollet/ARC
8
https://medium.com/tensorflow/colab-an-easy-way-to-learn-and-use-tensorflow-d74d1686e309
9
https://colab.research.google.com/github/madewithml/practicalAI/
10
https://colab.research.google.com/github/aamini/introtodeeplearning_labs
11
https://blog.jupyter.org/jupyterlab-is-ready-for-users-5a6f039b8906

2
 A PREPRINT - N OVEMBER 5, 2020

3 Deep Learning
Learning according to Mitchell (1997):

"A computer program is said to learn from experience E with respect to some class of tasks T and performance
measure P, if its performance at tasks in T, as measured by P, improves with experience E." — Tom Mitchell

After the Historical AI Debate12 : "Yoshua Bengio and Gary Marcus on the Best Way Forward for AI" https://
montrealartificialintelligence.com/aidebate/, there have been clarifications on the term "deep learning"13 .

"Deep learning is inspired by neural networks of the brain to build learning machines which discover rich and useful
internal representations, computed as a composition of learned features and functions." — Yoshua Bengio

"DL is constructing networks of parameterized functional modules and training them from examples using
gradient-based optimization." — Yann LeCun

"... replace symbols by vectors and logic by continuous (or differentiable) functions." — Yann LeCun

Deep learning allows computational models that are composed of multiple processing layers to learn REPRESEN-
TATIONS of (raw) data with multiple levels of abstraction[2]. At a high-level, neural networks are either encoders,
decoders, or a combination of both14 . Introductory course http://introtodeeplearning.com. See also Table 1.

Table 1: Types of Learning, by Alex Graves at NeurIPS 2018

Name With Teacher Without Teacher
Active Reinforcement Learning / Active Learning Intrinsic Motivation / Exploration
Passive Supervised Learning Unsupervised Learning

Figure 1: Multilayer perceptron (MLP).

Deep learning assumes that the data was generated by the composition of factors potentially at multiple levels in a
hierarchy15 . Deep learning (distributed representations + composition) is a general-purpose learning procedure.

"When you first study a field, it seems like you have to memorize a zillion things. You don’t. What you need is to identify
the 3-5 core principles that govern the field. The million things you thought you had to memorize are various
combinations of the core principles." — J. Reed
12
https://www.zdnet.com/article/devils-in-the-details-in-bengio-marcus-ai-debate/
13
https://www.zdnet.com/article/whats-in-a-name-the-deep-learning-debate/
14
https://github.com/lexfridman/mit-deep-learning
15
https://www.deeplearningbook.org

3
 A PREPRINT - N OVEMBER 5, 2020

"1. Multiply things together

2. Add them up
3. Replaces negatives with zeros
4. Return to step 1, a hundred times."
— Jeremy Howard

v Linear Algebra. Prof. Gilbert Strang16 .

v Dive into Deep Learning http://d2l.ai.
v Minicourse in Deep Learning with PyTorch17 .
v How to do Research At the MIT AI Lab (1988)18 .
v Introduction to Artificial Intelligence, Gilles Louppe19 .
v Energy-Based Models for Continual Learning, Anonymous20 .
v Fast and Easy Infinitely Wide Networks with Neural Tangents21 .
v Deep Learning. The full deck of (600+) slides, Gilles Louppe22 .
v These Lyrics Do Not Exist https://theselyricsdonotexist.com.
v AI and Wargaming, Goodman et al. https://arxiv.org/abs/2009.08922.
v Backward Feature Correction: How Deep Learning Performs Deep Learning23 .
v A Selective Overview of Deep Learning https://arxiv.org/abs/1904.05526.
v The Missing Semester of Your CS Education https://missing.csail.mit.edu.
v fastai: A Layered API for Deep Learning https://arxiv.org/abs/2002.04688.
v Thinking Fast and Slow in AI, Booch et al. https://arxiv.org/abs/2010.06002.
v Anatomy of Matplotlib https://github.com/matplotlib/AnatomyOfMatplotlib.
v Data project checklist https://www.fast.ai/2020/01/07/data-questionnaire/.
v Using Nucleus and TensorFlow for DNA Sequencing Error Correction, Colab Notebook24 .
v Machine Learning for Physicists https://machine-learning-for-physicists.org.
v Flow-edge Guided Video Completion, Gao et al. https://arxiv.org/abs/2009.01835.
v The world as a neural network, Vitaly Vanchurin https://arxiv.org/abs/2008.01540.
v Generalized Energy Based Models, Michael Arbel, Liang Zhou and Arthur Gretton, 202025 .
v Representing Scenes as Neural Radiance Fields for View Synthesis. Mildenhall et al., 202026 .
v PoseNet Sketchbook https://googlecreativelab.github.io/posenet-sketchbook/.
v The Neural Network, A Visual Introduction. Vivek Verma : https://youtu.be/UOvPeC8WOt8.
v Synthetic Data for Deep Learning, Sergey I. Nikolenko https://arxiv.org/abs/1909.11512.
v Removing people from complex backgrounds in real time using TensorFlow.js in the web browser27 .
v A Recipe for Training Neural Networks https://karpathy.github.io/2019/04/25/recipe/.
v TensorFlow Datasets: load a variety of public datasets into TensorFlow programs (Blog28 | Colab29 ).
v Denoising Diffusion Probabilistic Models, Ho et al., 2020 https://arxiv.org/abs/2006.11239.
v The Markov-Chain Monte Carlo Interactive Gallery https://chi-feng.github.io/mcmc-demo/.
v NeurIPS 2019 Implementations https://paperswithcode.com/conference/neurips-2019-12.
v Involutive MCMC: a Unifying Framework, Neklyudov et al. https://arxiv.org/abs/2006.16653.
v Algebra, Topology, Differential Calculus, and Optimization Theory For Computer Science and Machine Learning30 .
v How to Choose Your First AI Project https://hbr.org/2019/02/how-to-choose-your-first-ai-project.
v Blog | MIT 6.S191 https://medium.com/tensorflow/mit-introduction-to-deep-learning-4a6f8dde1f0c.
v A Fortran-Keras Deep Learning Bridge for Scientific Computing, Ott et al. https://arxiv.org/abs/2004.10652.
16
https://ocw.mit.edu/courses/mathematics/18-06-linear-algebra-spring-2010/video-lectures/
17
https://github.com/Atcold/pytorch-Deep-Learning-Minicourse
18
http://dspace.mit.edu/bitstream/handle/1721.1/41487/AI_WP_316.pdf
19
https://glouppe.github.io/info8006-introduction-to-ai/pdf/lec-all.pdf
20
https://openreview.net/forum?id=j5d9qacxdZa
21
https://ai.googleblog.com/2020/03/fast-and-easy-infinitely-wide-networks.html
22
https://github.com/glouppe/info8010-deep-learning/raw/v2-info8010-2019/pdf/lec-all.pdf
23
https://arxiv.org/abs/2001.04413
24
https://colab.research.google.com/github/google/nucleus/blob/master/nucleus/examples/dna_
sequencing_error_correction.ipynb
25
https://arxiv.org/abs/2003.05033
26
http://www.matthewtancik.com/nerf
27
https://github.com/jasonmayes/Real-Time-Person-Removal
28
https://medium.com/tensorflow/introducing-tensorflow-datasets-c7f01f7e19f3
29
https://colab.research.google.com/github/tensorflow/datasets/blob/master/docs/overview.ipynb
30
https://drive.google.com/file/d/1sJvLQwxMyu89t2z4Zf9tD7O7efnbIUyB/view

4
 A PREPRINT - N OVEMBER 5, 2020

GitHub31 .
v A Wholistic View of Continual Learning with Deep Neural Networks: Forgotten Lessons and the Bridge to Active
and Open World Learning. Mundt et al., 202032 .

3.1 Universal Approximation Theorem

The universal approximation theorem states that a feed-forward network with a single hidden layer containing a finite
number of neurons can solve any given problem to arbitrarily close accuracy as long as you add enough parameters.
Neural Networks + Gradient Descent + GPU33 :
• Infinitely flexible function: Neural Network (multiple hidden layers: Deep Learning)34 .
• All-purpose parameter fitting: Backpropagation3536 . Backpropagation is the key algorithm that makes training
deep models computationally tractable and highly efficient37 . The backpropagation procedure is nothing more
than a practical application of the chain rule for derivatives.

Figure 2: All-purpose parameter fitting: Backpropagation.

• Fast and scalable: GPU.

"You have relatively simple processing elements that are very loosely models of neurons. They have connections coming
in, each connection has a weight on it, and that weight can be changed through learning." — Geoffrey Hinton

Deep learning : connect a dataset, a model, a cost function and an optimization procedure.

"Deep learning has fully solved the curse of dimensionality. It vanished like an RNN gradient!" — Ilya Sutskever

When a choice must be made, just feed the (raw) data to a deep neural network (Universal function approximators).

3.2 Convolution Neural Networks (Useful for Images | Space)

Richer innate priors : innateness that enables learning.

A significant percentage of Deep Learning breakthroughs comes from reusable constructs and parameters sharing. The
deep convolutional network is a construct that reuses weights in multiple locations (parameters sharing in space)38 .
31
https://github.com/scientific-computing/FKB
32
https://arxiv.org/abs/2009.01797
33
http://wiki.fast.ai/index.php/Lesson_1_Notes
34
http://neuralnetworksanddeeplearning.com/chap4.html
35
https://github.com/DebPanigrahi/Machine-Learning/blob/master/back_prop.ipynb
36
https://www.jeremyjordan.me/neural-networks-training/
37
https://colah.github.io/posts/2015-08-Backprop/
38
https://twitter.com/iamtrask/status/949439556499230720

5
 A PREPRINT - N OVEMBER 5, 2020

"Virtually all modern observers would concede that genes and experience work together; it is “nature and nurture”,
not “nature versus nurture”. No nativist, for instance, would doubt that we are also born with specific biological
machinery that allows us to learn. Chomsky’s Language Acquisition Device should be viewed precisely as an innate
learning mechanism, and nativists such as Pinker, Peter Marler (Marler, 2004) and myself (Marcus, 2004) have
frequently argued for a view in which a significant part of a creature’s innate armamentarium consists not of specific
knowledge but of learning mechanisms, a form of innateness that enables learning." — Gary Marcus, Innateness,
AlphaZero, and Artificial Intelligence39

The deep convolutional network, inspired by Hubel and Wiesel’s seminal work on early visual cortex, uses hierarchical
layers of tiled convolutional filters to mimic the effects of receptive fields, thereby exploiting the local spatial correlations
present in images[1]. See Figure 4. Demo https://ml4a.github.io/demos/convolution/.

0 1 1 1×1 0×0 0×1 0

0 0 1 1×0 1×1 0×0 0 1 4 3 4 1
0 0 0 1×1 1×0 1×1 0 1 0 1 1 2 4 3 3
0 0 0 1 1 0 0 ∗ 0 1 0 = 1 2 3 4 1
0 0 1 1 0 0 0 1 0 1 1 3 3 1 1
0 1 1 0 0 0 0 3 3 1 1 0
1 1 0 0 0 0 0
I K I∗K

Figure 3: 2D Convolution. Source: Cambridge Coding Academy

A ConvNet is made up of Layers. Every Layer has a simple API: It transforms an input 3D volume to an output 3D
volume with some differentiable function that may or may not have parameters40 . Reading41 .
In images, local combinations of edges form motifs, motifs assemble into parts, and parts form objects4243 .
Representation learning : the language of neural networks. The visual vocabulary of a convolutional neural network
seems to emerge from low level features such as edges and orientations, and builds up textures, patterns and composites,
. . . and builds up even further into complete objects. This relates to Wittgenstein’s "language-game" in Philosophical
Investigations44 , where a functional language emerge from simple tasks before defining a vocabulary45 .

Figure 4: Architecture of LeNet-5, a Convolutional Neural Network. LeCun et al., 1998

"DL is essentially a new style of programming – "differentiable programming" – and the field is trying to work out the
reusable constructs in this style. We have some: convolution, pooling, LSTM, GAN, VAE, memory units, routing units,
etc." — Thomas G. Dietterich
39
https://arxiv.org/abs/1801.05667
40
http://cs231n.github.io/convolutional-networks/
41
https://ml4a.github.io/ml4a/convnets/
42
http://yosinski.com/deepvis
43
https://distill.pub/2017/feature-visualization/
44
https://en.wikipedia.org/wiki/Philosophical_Investigations
45
https://media.neurips.cc/Conferences/NIPS2018/Slides/Deep_Unsupervised_Learning.pdf

6
 A PREPRINT - N OVEMBER 5, 2020

v Image Classification from Scratch46 .

v CS231N : Convolutional Neural Networks for Visual Recognition47 .
v Introduction to Graph Convolutional Network (GCN). Alfredo Canziani48 .
v Deep Plastic Surgery: Robust and Controllable Image Editing with Human-Drawn Sketches. Yang et al.49 .
v CNN Explainer: Learning Convolutional Neural Networks with Interactive Visualization. Wang et al. 50 51 .
v An Overview of Early Vision in InceptionV1 https://distill.pub/2020/circuits/early-vision/.
v Neural Voice Puppetry: Audio-driven Facial Reenactment. Thies et al. https://arxiv.org/abs/1912.05566.
v TensorSpace (https://tensorspace.org) offers interactive 3D visualizations of LeNet, AlexNet and Inceptionv3.

3.3 Recurrent Neural Networks (Useful for Sequences | Time)

Recurrent neural networks are networks with loops in them, allowing information to persist52 . RNNs process an input
sequence one element at a time, maintaining in their hidden units a ‘state vector’ that implicitly contains information
about the history of all the past elements of the sequence[2]. For sequential inputs. See Figure 5.

ht h0 h1 h2 h3 ht

A = A A A A ... A

xt x0 x1 x2 x3 xt

Figure 5: RNN Layers Reuse Weights for Multiple Timesteps.

Figure 6: Google Smart Reply System is built on a pair of recurrent neural networks. Diagram by Chris Olah

"I feel like a significant percentage of Deep Learning breakthroughs ask the question “how can I reuse weights in
multiple places?” – Recurrent (LSTM) layers reuse for multiple timesteps – Convolutional layers reuse in multiple
locations. – Capsules reuse across orientation." — Andrew Trask

46
https://colab.research.google.com/drive/1umJnCp8tZ7UDTYSQsuWdKRhqbHts38AC
47
https://www.youtube.com/playlist?list=PLzUTmXVwsnXod6WNdg57Yc3zFx_f-RYsq
48
https://atcold.github.io/pytorch-Deep-Learning/en/week13/13-3/
49
https://arxiv.org/abs/2001.02890
50
https://arxiv.org/abs/2004.15004
51
http://poloclub.github.io/cnn-explainer/
52
http://colah.github.io/posts/2015-08-Understanding-LSTMs/

7
 A PREPRINT - N OVEMBER 5, 2020

v CS224N : Natural Language Processing with Deep Learning53 .

v Long Short-Term-Memory (LSTM), Sepp Hochreiter and Jürgen Schmidhuber54 .
v The Unreasonable Effectiveness of Recurrent Neural Networks, blog (2015) by Andrej Karpathy55 .
v Understanding LSTM Networks http://colah.github.io/posts/2015-08-Understanding-LSTMs/.
v Can Neural Networks Remember? Slides by Vishal Gupta: http://vishalgupta.me/deck/char_lstms/.

3.4 Transformers

Figure 7: Attention Is All You Need. Vaswani et al., 2017 : https://arxiv.org/abs/1706.03762.

Transformers are generic, simples and exciting machine learning architectures designed to process a connected set of
units (tokens in a sequence, pixels in an image, etc.) where the only interaction between units is through self-attention.
Transformers’ performance limit seems purely in the hardware (how big a model can be fitted in GPU memory)56 .
The fundamental operation of transformers is self-attention (a sequence-to-sequence operation, Figure 8): an attention
mechanism relating different positions of a single sequence in order to compute a representation of the same sequence57 .
Let’s call the input vectors (of dimension k) :
x1 , x2 , ..., xt (1)
Let’s call the corresponding output vectors (of dimension k) :
y1 , y2 , ..., yt (2)

The self attention operation takes a weighted average over all the input vectors :
X
yi = wij xj (3)
j

The weight wij is derived from a function over xi and xj . The simplest option is the dot product (with softmax) :

T
exi xj
wij = P xT x (4)
ei j
j

53
https://www.youtube.com/playlist?list=PLU40WL8Ol94IJzQtileLTqGZuXtGlLMP_
54
https://www.bioinf.jku.at/publications/older/2604.pdf
55
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
56
http://www.peterbloem.nl/blog/transformers
57
https://lilianweng.github.io/lil-log/2018/06/24/attention-attention.html

8
 A PREPRINT - N OVEMBER 5, 2020

Figure 8: Self-attention. By Peter Bloem : http://www.peterbloem.nl/blog/transformers.

Transformers are Graph Neural Networks58 .

v The Transformer Family. By Lilian Weng59 .
v Transformers Notebooks. By Hugging Face60 .
v Text classification with Transformer. Colab61 .
v Making Transformer networks simpler and more efficient62 .
v AttentioNN: All about attention in neural networks described as colab notebooks63 .
v Attention Is All You Need, Vaswani et al. https://arxiv.org/abs/1706.03762.
v Efficient Transformers: A Survey, Tay et al. https://arxiv.org/abs/2009.06732.
v How to train a new language model from scratch using Transformers and Tokenizers64 .
v Write With Transformer. By Hugging Face: https://transformer.huggingface.co.
v The Illustrated Transformer http://jalammar.github.io/illustrated-transformer/.
v How to generate text: using different decoding methods for language generation with Transformers65 .
v The annotated transformer (code) http://nlp.seas.harvard.edu/2018/04/03/attention.html.
v Attention and Augmented Recurrent Neural Networks https://distill.pub/2016/augmented-rnns/.
v Transformer model for language understanding. Tutorial showing how to write Transformer in TensorFlow 2.066 .
v End-to-End Object Detection with Transformers, Carion et al. https://arxiv.org/abs/2005.12872. Colab67 .
v Transformer in TensorFlow 2.0 (code) https://www.tensorflow.org/beta/tutorials/text/transformer.

3.4.1 Natural Language Processing (NLP) | BERT: A New Era in NLP

BERT (Bidirectional Encoder Representations from Transformers)[6] is a deeply bidirectional, unsupervised language
representation, pre-trained using only a plain text corpus (in this case, Wikipedia)68 .

• Reading: Unsupervised pre-training of an LSTM followed by supervised fine-tuning[7].

• TensorFlow code and pre-trained models for BERT https://github.com/google-research/bert.
• Better Language Models and Their Implications69 .
58
https://graphdeeplearning.github.io/post/transformers-are-gnns/
59
https://lilianweng.github.io/lil-log/2020/04/07/the-transformer-family.html
60
https://github.com/huggingface/transformers/tree/master/notebooks
61
https://colab.research.google.com/github/keras-team/keras-io/blob/master/examples/nlp/ipynb/
text_classification_with_transformer.ipynb
62
https://ai.facebook.com/blog/making-transformer-networks-simpler-and-more-efficient/
63
https://github.com/zaidalyafeai/AttentioNN
64
https://huggingface.co/blog/how-to-train
65
https://huggingface.co/blog/how-to-generate
66
https://www.tensorflow.org/tutorials/text/transformer
67
https://colab.research.google.com/drive/1rPm0-UrWHpJJRX9PsNb5SpzZiUlMh7wm
68
https://ai.googleblog.com/2018/11/open-sourcing-bert-state-of-art-pre.html
69
https://blog.openai.com/better-language-models/

9
 A PREPRINT - N OVEMBER 5, 2020

Figure 9: The two steps of how BERT is developed. Source https://jalammar.github.io/illustrated-bert/.

"I think transfer learning is the key to general intelligence. And I think the key to doing transfer learning will be the
acquisition of conceptual knowledge that is abstracted away from perceptual details of where you learned it from." —
Demis Hassabis

v Towards a Conversational Agent that Can Chat About. . . Anything70 .

v How to Build OpenAI’s GPT-2: "The AI That’s Too Dangerous to Release"71 .
v A Primer in BERTology: What we know about how BERT works, Rogers et al., 202072 .
v Play with BERT with your own data using TensorFlow Hub https://colab.research.google.com/github/
google-research/bert/blob/master/predicting_movie_reviews_with_bert_on_tf_hub.ipynb.

3.5 Unsupervised Learning

True intelligence will require independent learning strategies.

"Give a robot a label and you feed it for a second; teach a robot to label and you feed it for a lifetime." — Pierre
Sermanet

Unsupervised learning is a paradigm for creating AI that learns without a particular task in mind: learning for the
sake of learning73 . It captures some characteristics of the joint distribution of the observed random variables (learn the
underlying structure). The variety of tasks include density estimation, dimensionality reduction, and clustering.[4]74 .

"The unsupervised revolution is taking off!" — Alfredo Canziani

Self-supervised learning is derived form unsupervised learning where the data provides the supervision. E.g.
Word2vec75 , a technique for learning vector representations of words, or word embeddings. An embedding is a
mapping from discrete objects, such as words, to vectors of real numbers76 .

"The next revolution of AI won’t be supervised." — Yann LeCun

70
https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.html
71
https://blog.floydhub.com/gpt2/
72
https://arxiv.org/abs/2002.12327
73
https://deepmind.com/blog/unsupervised-learning/
74
https://media.neurips.cc/Conferences/NIPS2018/Slides/Deep_Unsupervised_Learning.pdf
75
https://jalammar.github.io/illustrated-word2vec/
76
http://projector.tensorflow.org

10
 A PREPRINT - N OVEMBER 5, 2020

Figure 10: A Simple Framework for Contrastive Learning of Visual Representations, Chen et al., 2020

"Self-supervised learning is a method for attacking unsupervised learning problems by using the mechanisms of
supervised learning." — Thomas G. Dietterich

v Self-Supervised Image Classification, Papers With Code77 .

v Self-supervised learning and computer vision, Jeremy Howard78 .
v Understanding Self-supervised Learning with Dual Deep Networks, Tian et al.79
v Momentum Contrast for Unsupervised Visual Representation Learning, He et al.80
v Data-Efficient Image Recognition with Contrastive Predictive Coding, Hénaff et al.81
v A Simple Framework for Contrastive Learning of Visual Representations, Chen et al.82
v FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence, Sohn et al.83
v Viewmaker Networks: Learning Views for Unsupervised Representation Learning, Tamkin et al.84
v Self-classifying MNIST Digits, Randazzo et al.: https://distill.pub/2020/selforg/mnist/.
v Self-Supervised Learning of Pretext-Invariant Representations, Ishan Misra, Laurens van der Maaten85 .

3.5.1 Generative Adversarial Networks

Simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model
D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is
to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game[3].

min max[IEx∼pdata (x) [logDθd (x)] + IEz∼pz (z) [log(1 − Dθd (Gθg (z)))]] (5)
θg θd

"What I cannot create, I do not understand." — Richard Feynman

Goodfellow et al. used an interesting analogy where the generative model can be thought of as analogous to a team of
counterfeiters, trying to produce fake currency and use it without detection, while the discriminative model is analogous
to the police, trying to detect the counterfeit currency. Competition in this game drives both teams to improve their
methods until the counterfeits are indistiguishable from the genuine articles. See Figure 11.
77
https://paperswithcode.com/task/self-supervised-image-classification
78
https://www.fast.ai/2020/01/13/self_supervised/
79
https://arxiv.org/abs/2010.00578
80
https://arxiv.org/abs/1911.05722
81
https://arxiv.org/abs/1905.09272
82
https://arxiv.org/abs/2002.05709
83
https://arxiv.org/abs/2001.07685
84
https://arxiv.org/abs/2010.07432
85
https://arxiv.org/abs/1912.01991

11
 A PREPRINT - N OVEMBER 5, 2020

Figure 11: GAN: Neural Networks Architecture Pioneered by Ian Goodfellow at University of Montreal (2014).

StyleGAN: A Style-Based Generator Architecture for Generative Adversarial Networks

• Paper http://stylegan.xyz/paper | Code https://github.com/NVlabs/stylegan.

• StyleGAN for art. Colab https://colab.research.google.com/github/ak9250/stylegan-art.
• This Person Does Not Exist https://thispersondoesnotexist.com.
• Which Person Is Real? http://www.whichfaceisreal.com.
• This Resume Does Not Exist https://thisresumedoesnotexist.com.
• This Waifu Does Not Exist https://www.thiswaifudoesnotexist.net.
• Encoder for Official TensorFlow Implementation https://github.com/Puzer/stylegan-encoder.
• How to recognize fake AI-generated images. By Kyle McDonald86 .

v GAN in Keras. Colab87 .

v 100,000 Faces Imagined by a GAN https://generated.photos.
v Introducing TF-GAN: A lightweight GAN library for TensorFlow 2.088 .
v Generative Adversarial Networks (GANs) in 50 lines of code (PyTorch)89 .
v Few-Shot Adversarial Learning of Realistic Neural Talking Head Models90 .
v Wasserstein GAN http://www.depthfirstlearning.com/2019/WassersteinGAN.
v GANpaint Paint with GAN units http://gandissect.res.ibm.com/ganpaint.html.
v StyleGAN2 Distillation for Feed-forward Image Manipulation. Viazovetskyi et al.91 Code92 .
v A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications. Gui et al.93 .
v CariGANs: Unpaired Photo-to-Caricature Translation. Cao et al.: https://cari-gan.github.io.
v Infinite-resolution (CPPNs, GANs and TensorFlow.js) https://thispicturedoesnotexist.com.
v PyTorch pretrained BigGAN https://github.com/huggingface/pytorch-pretrained-BigGAN.
v GANSynth: Generate high-fidelity audio with GANs! Colab http://goo.gl/magenta/gansynth-demo.
v SC-FEGAN: Face Editing Generative Adversarial Network https://github.com/JoYoungjoo/SC-FEGAN.
v Demo of BigGAN in an official Colaboratory notebook (backed by a GPU) https://colab.research.google.
com/github/tensorflow/hub/blob/master/examples/colab/biggan_generation_with_tf_hub.ipynb

86
https://medium.com/@kcimc/how-to-recognize-fake-ai-generated-images-4d1f6f9a2842
87
https://colab.research.google.com/drive/1CQ2XTMoUB7b9i9USUh4kp8BoCag1z-en
88
https://medium.com/tensorflow/introducing-tf-gan-a-lightweight-gan-library-for-tensorflow-2-0-36d767e1abae
89
https://medium.com/@devnag/generative-adversarial-networks-gans-in-50-lines-of-code-pytorch-e81b79659e3f
90
https://arxiv.org/abs/1905.08233
91
https://arxiv.org/abs/2003.03581
92
https://github.com/EvgenyKashin/stylegan2-distillation
93
https://arxiv.org/abs/2001.06937

12
 A PREPRINT - N OVEMBER 5, 2020

3.5.2 Variational AutoEncoder

Variational Auto-Encoders94 (VAEs) are powerful models for learning low-dimensional representations See Figure 12.
Disentangled representations are defined as ones where a change in a single unit of the representation corresponds to a
change in single factor of variation of the data while being invariant to others (Bengio et al. (2013).

Figure 12: Variational Autoencoders (VAEs): Powerful Generative Models.

v Colab95 : "Debiasing Facial Detection Systems." AIEthics

v NVAE: A Deep Hierarchical Variational Autoencoder, Arash Vahdat and Jan Kautz 96 .
v Reading: Disentangled VAE’s (DeepMind 2016) https://arxiv.org/abs/1606.05579.
v Slides: A Few Unusual Autoencoders https://colinraffel.com/talks/vector2018few.pdf.
v MusicVAE: Learning latent spaces for musical scores https://magenta.tensorflow.org/music-vae.
v Generative models in Tensorflow 2 https://github.com/timsainb/tensorflow2-generative-models/.
v SpaceSheet: Interactive Latent Space Exploration with a Spreadsheet https://vusd.github.io/spacesheet/.

3.5.3 Capsule
Stacked Capsule Autoencoders. The inductive biases in this unsupervised version of capsule networks give rise to
object-centric latent representations, which are learned in a self-supervised way—simply by reconstructing input images.
Clustering learned representations is enough to achieve unsupervised state-of-the-art classification performance on
MNIST (98.5%). Reference: blog by Adam Kosiorek.97 Code98 .
Capsules learn equivariant object representations (applying any transformation to the input of the function has the same
effect as applying that transformation to the output of the function).

Figure 13: Stacked Capsule Autoencoders. Image source: Blog by Adam Kosiorek.

4 Autonomous Agents
We are on the dawn of The Age of Artificial Intelligence.

"In a moment of technological disruption, leadership matters." — Andrew Ng

94
https://arxiv.org/abs/1906.02691v2
95
https://colab.research.google.com/github/aamini/introtodeeplearning_labs/blob/master/lab2/Part2_
debiasing_solution.ipynb
96
https://arxiv.org/abs/2007.03898
97
http://akosiorek.github.io/ml/2019/06/23/stacked_capsule_autoencoders.html
98
https://github.com/google-research/google-research/tree/master/stacked_capsule_autoencoders

13
 A PREPRINT - N OVEMBER 5, 2020

An autonomous agent is any device that perceives its environment and takes actions that maximize its chance of
success at some goal. At the bleeding edge of AI, autonomous agents can learn from experience, simulate worlds and
orchestrate meta-solutions. Here’s an informal definition99 of the universal intelligence of agent π 100 :

X
Υ(π) := 2−K(µ) Vµπ (6)
µ∈E

"Intelligence measures an agent’s ability to achieve goals in a wide range of environments." — Legg and Hutter, 2007

4.1 Deep Reinforcement Learning

Figure 14: An Agent Interacts with an Environment.

Reinforcement learning (RL) studies how an agent can learn how to achieve goals in a complex, uncertain environment
(Figure 14) [5]. Recent superhuman results in many difficult environments combine deep learning with RL (Deep
Reinforcement Learning). See Figure 15 for a taxonomy of RL algorithms.
ú Spinning Up in Deep RL - Proximal Policy Optimization (PPO), Colab Notebook101 .
v RL Tutorial, Behbahani et al.102 .
v An Opinionated Guide to ML Research103 .
v CS 188 : Introduction to Artificial Intelligence104 .
v Introduction to Reinforcement Learning by DeepMind105 .
v Isaac Gym https://developer.nvidia.com/isaac-gym.
v Discovering Reinforcement Learning Algorithms, Oh et al.106 .
v The NetHack Learning Environment, Küttler et al.107 GitHub108 .
v "My Top 10 Deep RL Papers of 2019" by Robert Tjarko Lange109 .
v Behavior Priors for Efficient Reinforcement Learning, Tirumala et al.110 .
v Deep tic-tac-toe https://zackakil.github.io/deep-tic-tac-toe/.
99
https://arxiv.org/abs/0712.3329
100
Where µ is an environment, K is the Kolmogorov complexity function, E is the space of all computable reward summable
environmental measures with respect to the reference machine U and the value function Vµπ is the agent’s “ability to achieve”.
101
https://colab.research.google.com/drive/1piaU0x7nawRpSLKOTaCEdUG0KAR2OXku
102
https://github.com/eemlcommunity/PracticalSessions2020/blob/master/rl/EEML2020_RL_Tutorial.ipynb
103
http://joschu.net/blog/opinionated-guide-ml-research.html
104
https://inst.eecs.berkeley.edu/~cs188/fa18/
105
https://www.youtube.com/watch?v=2pWv7GOvuf0&list=PLqYmG7hTraZDM-OYHWgPebj2MfCFzFObQ
106
https://arxiv.org/abs/2007.08794
107
https://arxiv.org/abs/2006.13760
108
https://github.com/facebookresearch/nle
109
https://roberttlange.github.io/posts/2019/12/blog-post-9/
110
https://arxiv.org/pdf/2010.14274.pdf

14
 A PREPRINT - N OVEMBER 5, 2020

v A Framework for Reinforcement Learning and Planning, Moerland et al. 111 .

v Automatic Curriculum Learning For Deep RL: A Short Survey, Portelas et al.112 .
v ALLSTEPS: Curriculum-driven Learning of Stepping Stone Skills, Xie et al. 113 .
v Decoupling Representation Learning from Reinforcement Learning, Stooke et al.114 Code115
v Chip Placement with Deep Reinforcement Learning https://arxiv.org/abs/2004.10746.
v RL Unplugged: Benchmarks for Offline Reinforcement Learning, Gulcehre et al.116 GitHub117 .
v CS 287: Advanced Robotics118 . https://people.eecs.berkeley.edu/~pabbeel/cs287-fa19/.
v Combining Deep Reinforcement Learning and Search for Imperfect-Information Games, Brown et al.119 .
v MDP Homomorphic Networks: Group Symmetries in Reinforcement Learning, Elise van der Pol et al.120 .
v Too many cooks: Bayesian inference for coordinating multi-agent collaboration, Wang et al.121 GitHub122 .
v One Policy to Control Them All: Shared Modular Policies for Agent-Agnostic Control, Huang et al.123 . Code124 .
v Decentralized Reinforcement Learning: Global Decision-Making via Local Economic Transactions, Chang et al. 125 .

Figure 15: A Taxonomy of RL Algorithms. Source: Spinning Up in Deep RL by Achiam et al. | OpenAI

4.1.1 Model-Free RL | Value-Based

The goal in RL is to train the agent to maximize the discounted sum of all future rewards Rt , called the return:

Rt = rt + γrt+1 + γ 2 rt+2 + . . . (7)

The Q-function captures the expected total future reward an agent in state s can receive by executing a certain action a:

Q(s, a) = E[Rt ] (8)

The optimal policy should choose the action a that maximizes Q(s,a):

π ∗ (s) = argmaxa Q(s, a) (9)

111
https://arxiv.org/abs/2006.15009
112
https://arxiv.org/abs/2003.04664
113
https://www.cs.ubc.ca/~van/papers/2020-allsteps/
114
https://arxiv.org/abs/2009.08319
115
https://github.com/astooke/rlpyt/tree/master/rlpyt/ul
116
https://arxiv.org/abs/2006.13888
117
https://github.com/deepmind/deepmind-research/tree/master/rl_unplugged
118
https://people.eecs.berkeley.edu/~pabbeel/cs287-fa19/exam/cs287-fa19-exam-study-handout.pdf
119
https://arxiv.org/abs/2007.13544
120
https://arxiv.org/abs/2006.16908
121
https://arxiv.org/abs/2003.11778
122
https://github.com/rosewang2008/gym-cooking
123
https://arxiv.org/abs/2007.04976
124
https://huangwl18.github.io/modular-rl
125
https://arxiv.org/abs/2007.02382

15
 A PREPRINT - N OVEMBER 5, 2020

Figure 16: Open-Source RL Algorithms https://docs.google.com/spreadsheets/d/1EeFPd-XIQ3mq_

9snTlAZSsFY7Hbnmd7P5bbT8LPuMn0/

Figure 17: DQN Training Algorithm. Volodymyr Mnih, Deep RL Bootcamp

• Q-Learning: Playing Atari with Deep Reinforcement Learning (DQN). Mnih et al, 2013[10]. See Figure 17.

"There’s no limit to intelligence." — David Silver

v Q-Learning in enormous action spaces via amortized approximate maximization, de Wiele et al.126 .
v TF-Agents (DQN Tutorial) | Colab https://colab.research.google.com/github/tensorflow/agents.

4.1.2 Model-Free RL | Policy-Based

An RL agent learns the stochastic policy function that maps state to action and act by sampling policy.

126
https://arxiv.org/abs/2001.08116

16
 A PREPRINT - N OVEMBER 5, 2020

Figure 18: Policy Gradient Directly Optimizes the Policy.

Run a policy for a while (code: https://gist.github.com/karpathy/a4166c7fe253700972fcbc77e4ea32c5):

τ = (s0 , a0 , r0 , s1 , a1 , r1 , . . . , sT −1 , aT −1 , rT −1 , sT ) (10)

Increase probability of actions that lead to high rewards and decrease probability of actions that lead to low rewards:
"T −1 #
X
∇θ Eτ [R(τ )] = Eτ ∇θ log π(at |st , θ)R(τ ) (11)
t=0

πθ (s, α1 )
πθ (s, α2 )
πθ (s, α3 )
πθ (s, α4 )
πθ (s, α5 )
s

Vψ (s)

Figure 19: Asynchronous Advantage Actor-Critic (A3C). Source: Petar Velickovic

• Policy Optimization: Asynchronous Methods for Deep Reinforcement Learning (A3C). Mnih et al, 2016[8].
• Policy Optimization: Proximal Policy Optimization Algorithms (PPO). Schulman et al, 2017[9].

v Phasic Policy Gradient. Cobbe et al.127 Code128 .

v Deep Reinforcement Learning for Playing 2.5D Fighting Games. Li et al.129 .
v rlpyt: A Research Code Base for Deep Reinforcement Learning in PyTorch. Adam Stooke, Pieter Abbee130 .

4.1.3 Model-Based RL
In Model-Based RL, the agent generates predictions about the next state and reward before choosing each action.
127
https://arxiv.org/abs/2009.04416
128
https://github.com/openai/phasic-policy-gradient
129
https://arxiv.org/abs/1805.02070
130
https://arxiv.org/abs/1909.01500

17
 A PREPRINT - N OVEMBER 5, 2020

Figure 20: World Model’s Agent consists of: Vision (V), Memory (M), and Controller (C). | Ha et al, 2018[11]

• Learn the Model: Recurrent World Models Facilitate Policy Evolution (World Models131 ). The world model
agent can be trained in an unsupervised manner to learn a compressed spatial and temporal representation of
the environment. Then, a compact policy can be trained. See Figure 20. Ha et al, 2018[11].
• Learn the Model: Learning Latent Dynamics for Planning from Pixels https://planetrl.github.io/.
• Given the Model: Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
(AlphaZero). Silver et al, 2017[14]. AlphaGo Zero Explained In One Diagram132 .

v Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model. Schrittwieser et al.133 . Pseudocode134 .

4.1.4 Toward a General AI-Agent Architecture: SuperDyna (General Dyna-style RL Agent)

"Intelligence is the computational part of the ability to predict and control a stream of experience." — Rich Sutton

SuperDyna.135 The ambition: a general AI agent for Artificial Biological Reinforcement Learning.

1. Interact with the world: sense, update state and take an action
2. Learn from what just happened: see what happened and learn from it
3. Plan: (while there is time remaining in this time step) imagine hypothetical states and actions you might take
4. Discover : curate options and features and measure how well they’re doing

The first complete and scalable general AI-agent architecture that has all the most important capabilities and desiderata:

• Acting, learning, planning, model-learning, subproblems, and options.

• Function approximation, partial observability, non-stationarity and stochasticity.
• Discovery of state features, and thereby of subproblems, options and models.
• All feeding back to motivate new, more-abstract features in a virtuous cycle of discovery.

Presentation by Richard Sutton (starts at 15 min.)136 .

"In practice, I work primarily in reinforcement learning as an approach to artificial intelligence. I am exploring ways to
represent a broad range of human knowledge in an empirical form–that is, in a form directly in terms of experience–and
in ways of reducing the dependence on manual encoding of world state and knowledge." — Richard S. Sutton
131
https://worldmodels.github.io
132
https://applied-data.science/static/main/res/alpha_go_zero_cheat_sheet.png
133
https://arxiv.org/abs/1911.08265
134
https://arxiv.org/src/1911.08265v2/anc/pseudocode.py
135
https://insidehpc.com/2020/02/video-toward-a-general-ai-agent-architecture/
136
https://slideslive.com/38921889/biological-and-artificial-reinforcement-learning-4

18
 A PREPRINT - N OVEMBER 5, 2020

Figure 21: Inner Loop of a General Dyna-Style RL Agent (SuperDyna).

Figure 22: SuperDyna: Virtuous cycle of discovery.

4.1.5 Improving Agent Design

Via Reinforcement Learning: Blog137 . arXiv138 . ASTool https://github.com/hardmaru/astool/.
Via Evolution: Video139 . Evolved Creatures http://www.karlsims.com/evolved-virtual-creatures.html.

"The future of high-level APIs for AI is... a problem-specification API. Currently we only search over network weights,
thus "problem specification" involves specifying a model architecture. In the future, it will just be: "tell me what data
you have and what you are optimizing"." — François Chollet

v Teacher algorithms for curriculum learning of Deep RL in continuously parameterized environments140 .

4.1.6 OpenAI Baselines

High-quality implementations of reinforcement learning algorithms https://github.com/openai/baselines.
137
https://designrl.github.io
138
https://arxiv.org/abs/1810.03779
139
https://youtu.be/JBgG_VSP7f8
140
https://arxiv.org/abs/1910.07224

19
 A PREPRINT - N OVEMBER 5, 2020

Figure 23: A comparison of the original LSTM cell vs. two new good generated. Top left: LSTM cell. [19]

ú Colab Notebook https://colab.research.google.com/drive/1amdIQaHWyc8Av_DoM5yFYHyYvyqD5BZX.

4.1.7 Google Dopamine and A Zoo of Agents

Dopamine is a research framework for fast prototyping of reinforcement learning algorithms.141 .
A Zoo of Atari-Playing Agents: Code142 , Blog143 and Colaboratory notebook https://colab.research.google.
com/github/uber-research/atari-model-zoo/blob/master/colab/AtariZooColabDemo.ipynb.

4.1.8 TRFL : TensorFlow Reinforcement Learning

TRFL ("truffle"): a library of reinforcement learning building blocks https://github.com/deepmind/trfl.

4.1.9 bsuite : Behaviour Suite for Reinforcement Learning

A collection of experiments that investigate core capabilities of RL agents http://github.com/deepmind/bsuite.

4.2 Evolution Strategies (ES)

In her Nobel Prize in Chemistry 2018 Lecture "Innovation by Evolution: Bringing New Chemistry to Life" (Nobel
Lecture)†144 , Prof. Frances H. Arnold said :

"Nature ... invented life that has flourished for billions of years. (...) Equally awe-inspiring is the process by which
Nature created these enzyme catalysts and in fact everything else in the biological world. The process is evolution, the
grand diversity-generating machine that created all life on earth, starting more than three billion years ago. (...)
evolution executes a simple algorithm of diversification and natural selection, an algorithm that works at all levels
of complexity from single protein molecules to whole ecosystems." — Prof. Frances H. Arnold

ú Demo: ES on LunarLanderContinuous-v2. Colab Notebook145 . Python Code146

141
https://github.com/google/dopamine
142
https://github.com/uber-research/atari-model-zoo
143
https://eng.uber.com/atari-zoo-deep-reinforcement-learning/
144
https://onlinelibrary.wiley.com/doi/epdf/10.1002/anie.201907729
145
https://colab.research.google.com/drive/1PpYYaihoJWiszZh1vhKXvmN2X9KnLA7i
146
https://drive.google.com/file/d/1YlKNorK21GMffz-29omEB7q_iLYRlmXD/view?usp=sharing

20
 A PREPRINT - N OVEMBER 5, 2020

Evolution and neural networks proved a potent combination in nature.

"Evolution is a slow learning algorithm that with the sufficient amount of compute produces a human brain." —
Wojciech Zaremba

Natural evolutionary strategy directly evolves the weights of a DNN and performs competitively with the best deep
reinforcement learning algorithms, including deep Q-networks (DQN) and policy gradient methods (A3C)[21].

Figure 24: https://colab.research.google.com/github/karpathy/randomfun/blob/master/es.ipynb.

Neuroevolution, which harnesses evolutionary algorithms to optimize neural networks, enables capabilities that are
typically unavailable to gradient-based approaches, including learning neural network building blocks, architectures
and even the algorithms for learning[12].

". . . evolution — whether biological or computational — is inherently creative, and should routinely be expected to
surprise, delight, and even outwit us." — The Surprising Creativity of Digital Evolution, Lehman et al.[22]

The ES algorithm is a “guess and check” process, where we start with some random parameters and then repeatedly:

1. Tweak the guess a bit randomly, and

2. Move our guess slightly towards whatever tweaks worked better.

Neural architecture search has advanced to the point where it can outperform human-designed models[13].

"Caterpillar brains LIQUIFY during metamorphosis, but the butterfly retains the caterpillar’s memories!" — M. Levin

"Open-ended" algorithms are algorithms that endlessly create. Brains and bodies evolve together in nature.

"We’re machines," says Hinton. ""We’re just produced biologically (...)" — Katrina Onstad, Toronto Life

v Evolution Strategies147 .
v VAE+CPPN+GAN148 .
v Demo: ES on CartPole-v1149 .
v AutoML-Zero: Evolving Machine Learning Algorithms From Scratch, Real et al.150 Code151 .
v Spiders Can Fly Hundreds of Miles Riding the Earth’s Magnetic Fields152 .
v A Visual Guide to ES http://blog.otoro.net/2017/10/29/visual-evolution-strategies/.
v Xenobots A scalable pipeline for designing reconfigurable organisms, Kriegman et al.153 . Learn154 . Evolve155 .

147
https://lilianweng.github.io/lil-log/2019/09/05/evolution-strategies.html
148
https://colab.research.google.com/drive/1_OoZ3z_C5Jl5gnxDOE9VEMCTs-Fl8pvM
149
https://colab.research.google.com/drive/1bMZWHdhm-mT9NJENWoVewUks7cGV10go
150
https://arxiv.org/abs/2003.03384
151
https://github.com/google-research/google-research/tree/master/automl_zero
152
https://www.cell.com/current-biology/fulltext/S0960-9822(18)30693-6
153
https://www.pnas.org/content/early/2020/01/07/1910837117
154
https://cdorgs.github.io
155
https://github.com/skriegman/reconfigurable_organisms

21
 A PREPRINT - N OVEMBER 5, 2020

4.3 Self Play

Silver et al.[15] introduced an algorithm based solely on reinforcement learning, without human data, guidance or
domain knowledge. Starting tabula rasa (and being its own teacher!), AlphaGo Zero achieved superhuman performance.
AlphaGo Zero showed that algorithms matter much more than big data and massive amounts of computation.

"Self-Play is Automated Knowledge Creation." — Carlos E. Perez

Self-play mirrors similar insights from coevolution. Transfer learning is the key to go from self-play to the real world156 .

"Open-ended self play produces: Theory of mind, negotiation, social skills, empathy, real language understanding." —
Ilya Sutskever, Meta Learning and Self Play

v How To Build Your Own MuZero AI Using Python157 .

v AlphaGo - The Movie | Full Documentary https://youtu.be/WXuK6gekU1Y.
v TensorFlow.js Implementation of DeepMind’s AlphaZero Algorithm for Chess. Live Demo158 . | Code159 .
v An open-source implementation of the AlphaGoZero algorithm https://github.com/tensorflow/minigo.
v ELF OpenGo: An Open Reimplementation of AlphaZero, Tian et al.: https://arxiv.org/abs/1902.04522.

4.4 Multi-Agent Populations

"We design a Theory of Mind neural network – a ToMnet – which uses meta-learning to build models of the agents it
encounters, from observations of their behaviour alone." — Machine Theory of Mind, Rabinowitz et al.[25]

Cooperative Agents. Learning to Model Other Minds, by OpenAI[24], is an algorithm which accounts for the fact that
other agents are learning too, and discovers self-interested yet collaborative strategies. Also: OpenAI Five160 .

Figure 25: Facebook, Carnegie Mellon build first AI that beats pros in 6-player poker https://ai.facebook.
com/blog/pluribus-first-ai-to-beat-pros-in-6-player-poker

"Artificial Intelligence is about recognising patterns, Artificial Life is about creating patterns." — Mizuki Oka et al.

Active Learning Without Teacher. In Intrinsic Social Motivation via Causal Influence in Multi-Agent RL, Jaques et
al. (2018) https://arxiv.org/abs/1810.08647 propose an intrinsic reward function designed for multi-agent RL
(MARL), which awards agents for having a causal influence on other agents’ actions. Open-source implementation 161 .
156
http://metalearning-symposium.ml
157
https://medium.com/applied-data-science/how-to-build-your-own-muzero-in-python-f77d5718061a
158
https://frpays.github.io/lc0-js/engine.html
159
https://github.com/frpays/lc0-js/
160
https://blog.openai.com/openai-five/
161
https://github.com/eugenevinitsky/sequential_social_dilemma_games

22
 A PREPRINT - N OVEMBER 5, 2020

"Open-ended Learning in Symmetric Zero-sum Games," Balduzzi et al.: https://arxiv.org/abs/1901.08106

v Lenia and Expanded Universe, Bert Wang-Chak Chan https://arxiv.org/abs/2005.03742.
v Neural MMO v1.3: A Massively Multiagent Game Environment for Training and Evaluating Neural Networks,
Suarezet al.162 Project Page https://jsuarez5341.github.io, Video163 and Slides164 .
v Neural MMO: A massively multiagent env. for simulations with many long-lived agents. Code165 and 3D Client166 .

4.5 Deep Meta-Learning

Learning to Learn[16].

"The notion of a neural "architecture" is going to disappear thanks to meta learning." — Andrew Trask

v Stanford CS330: Multi-Task and Meta-Learning, Finn et al., 2019167 .

v Meta Learning Shared Hierarchies[18] (The Lead Author is in High School!).
v Causal Reasoning from Meta-reinforcement Learning https://arxiv.org/abs/1901.08162.
v Meta-Learning through Hebbian Plasticity in Random Networks, Elias Najarro and Sebastian Risi, 2020168 .
v Meta-Learning Symmetries by Reparameterization, Zhou et al., 2020 https://arxiv.org/abs/2007.02933.

4.5.1 MAML: Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

The goal of model-agnostic meta-learning for fast adaptation of deep networks is to train a model on a variety of
learning tasks, such that it can solve new learning tasks using only a small number of training samples[20].
X

θ ← θ − β∇θ LTi fθi0 (12)
Ti ∼p(T )

A meta-learning algorithm takes in a distribution of tasks, where each task is a learning problem, and it produces a
quick learner — a learner that can generalize from a small number of examples[17].

Figure 26: Diagram of Model-Agnostic Meta-Learning (MAML)

v How to Train MAML (Model-Agnostic Meta-Learning)169 .

v Meta-Learning with Implicit Gradients https://arxiv.org/abs/1909.04630.
162
https://arxiv.org/abs/2001.12004
163
https://youtube.com/watch?v=DkHopV1RSxw
164
https://docs.google.com/presentation/d/1tqm_Do9ph-duqqAlx3r9lI5Nbfb9yUfNEtXk1Qo4zSw/edit?usp=
sharing
165
https://github.com/openai/neural-mmo
166
https://github.com/jsuarez5341/neural-mmo-client
167
http://youtube.com/playlist?list=PLoROMvodv4rMC6zfYmnD7UG3LVvwaITY5
168
https://arxiv.org/abs/2007.02686
169
https://medium.com/towards-artificial-intelligence/how-to-train-maml-model-agnostic-meta-learning-90aa093f8e46

23
 A PREPRINT - N OVEMBER 5, 2020

v Colaboratory reimplementation of MAML (Model-Agnostic Meta-Learning) in TF 2.0170 .

v Torchmeta: A Meta-Learning library for PyTorch171 https://github.com/tristandeleu/pytorch-meta.

4.5.2 The Grand Challenge for AI Research | AI-GAs: AI-Generating Algorithms, an Alternate Paradigm for
Producing General Artificial Intelligence
In AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence172 , Jeff
Clune describes an exciting path that ultimately may be successful at producing general AI. The idea is to create an
AI-generating algorithm (AI-GA), which automatically learns how to produce general AI.
Three Pillars are essential for the approach: (1) Meta-learning architectures, (2) Meta-learning the learning algo-
rithms themselves, and (3) Generating effective learning environments.

• The First Pillar, meta-learning architectures, could potentially discover the building blocks : convolution, re-
current layers, gradient-friendly architectures, spatial tranformers, etc.
• The Second Pillar, meta-learning learning algorithms, could potentially learn the building blocks : intelligent
exploration, auxiliary tasks, efficient continual learning, causal reasoning, active learning, etc.
• The Third Pillar, generating effective and fully expressive learning environments, could learn things like :
co-evolution / self-play, curriculum learning, communication / language, multi-agent interaction, etc.

On Earth,

"( . . . ) a remarkably simple algorithm (Darwinian evolution) began producing solutions to relatively simple
environments. The ‘solutions’ to those environments were organisms that could survive in them. Those organism often
created new niches (i.e. environments, or opportunities) that could be exploited. Ultimately, that process produced all
of the engineering marvels on the planet, such as jaguars, hawks, and the human mind." — Jeff Clune

Turing Complete (universal computer) : an encoding that enables the creation any possible learning algorithm.
Darwin Complete : an environmental encoding that enables the creation of any possible learning environment.
v Learning to Continually Learn. Beaulieu et al. https://arxiv.org/abs/2002.09571. Code173 .
v Self-Organizing Intelligent Matter: A blueprint for an AI generating algorithm. Anonymous et al.174

"We propose an artificial life framework of interacting neural elements as a basis of an AI generating algorithm." —
Anonymous175

v Fully Differentiable Procedural Content Generation through Generative Playing Networks. Bontrageret et
al.176

5 Symbolic AI
v Neural Module Networks for Reasoning over Text. Gupta et al.177 Code.178
v Inductive Logic Programming: Theory and methods. Muggleton, S.; De Raedt, L.179
v (Original FOIL paper) Learning Logical Definitions from Relations. J.R. Quinlan.180
v Neural-Symbolic Learning and Reasoning: A Survey and Interpretation. Besold et al.181
v On neural-symbolic computing: suggested readings on foundations of the field. Luis Lamb182 .
170
https://colab.research.google.com/github/mari-linhares/tensorflow-maml/blob/master/maml.ipynb
171
https://medium.com/pytorch/torchmeta-a-meta-learning-library-for-pytorch-f76c2b07ca6d
172
https://arxiv.org/abs/1905.10985
173
https://github.com/uvm-neurobotics-lab/ANML
174
https://openreview.net/forum?id=160xFQdp7HR
175
https://openreview.net/forum?id=160xFQdp7HR
176
https://arxiv.org/abs/2002.05259
177
https://arxiv.org/abs/1912.04971
178
https://nitishgupta.github.io/nmn-drop
179
https://doi.org/10.1016/0743-1066(94)90035-3
180
https://link.springer.com/article/10.1023/A:1022699322624
181
https://arxiv.org/abs/1711.03902
182
https://twitter.com/luislamb/status/1218575842340634626

24
 A PREPRINT - N OVEMBER 5, 2020

v Neuro-symbolic A.I. is the future of artificial intelligence. Here’s how it works. Luke Dormehl183 .
v Dimensions of Neural-symbolic Integration - A Structured Survey. Sebastian Bader, Pascal Hitzler184 .
v Graph Neural Networks Meet Neural-Symbolic Computing: A Survey and Perspective. Lamb et al.185 .

Figure 27: Graph Neural Networks Meet Neural-Symbolic Computing: A Survey and Perspective. Lamb et al.

"The paper was inspired by the AIDebate, Gary Marcus writings, the AAAI2020 Firechat with Daniel Kahneman, and
surveys not only our work, but the work of many in these AI fields." — Luis Lamb

v DDSP: Differentiable Digital Signal Processing. Engel et al. Blog186 , Colab187 , Paper188 and Code189 .
v The compositionality of neural networks: integrating symbolism and connectionism. Hupkes et al.190
v Graph Neural Networks Meet Neural-Symbolic Computing: A Survey and Perspective. Lamb et al.191
v A computing procedure for quantification theory. M .Davis, H. Putnam. J. of ACM, Vol. 7, pp. 201-214, 1960
v Discovering Symbolic Models from Deep Learning with Inductive Biases, Cranmer et al.192 . Blog and code193 .
v Symbolic Pregression: Discovering Physical Laws from Distorted Video. Silviu-Marian Udrescu, Max Tegmark194
v (Workshop series on neurosymbolic AI) Neural-Symbolic Integration. Hitzler et al. http://neural-symbolic.org
v Graph Colouring Meets Deep Learning: Effective Graph Neural Network Models for Combinatorial Problems.
Lemos et al. https://arxiv.org/abs/1903.04598.
v Neural-Symbolic Relational Reasoning on Graph Models: Effective Link Inference and Computation from Knowledge
Bases. Lemos et al. https://arxiv.org/abs/2005.02525.
v Neural-Symbolic Computing: An Effective Methodology for Principled Integration of Machine Learning and
Reasoning. Garcez et al. https://arxiv.org/abs/1905.06088
v Differentiable Reasoning on Large Knowledge Bases and Natural Language. Minervini et al.195 Open-source
neuro-symbolic reasoning framework, in TensorFlow https://github.com/uclnlp/gntp.
v (Original ILP foundational work) Automatic Methods of Inductive Inference, Plotkin G.D. PhD thesis, University of
Edinburgh, 1970 https://era.ed.ac.uk/bitstream/handle/1842/6656/Plotkin1972.pdf;sequence=1.

6 Environments

Platforms for training autonomous agents.

"Run a physics sim long enough and you’ll get intelligence." — Elon Musk

183
https://www.digitaltrends.com/cool-tech/neuro-symbolic-ai-the-future/
184
https://arxiv.org/abs/cs/0511042
185
https://arxiv.org/abs/2003.00330
186
http://magenta.tensorflow.org/ddsp
187
http://g.co/magenta/ddsp-demo
188
http://g.co/magenta/ddsp-paper
189
http://github.com/magenta/ddsp
190
https://arxiv.org/abs/1908.08351
191
https://arxiv.org/abs/2003.00330
192
https://arxiv.org/abs/2006.11287
193
https://astroautomata.com/paper/symbolic-neural-nets/
194
https://arxiv.org/abs/2005.11212
195
https://arxiv.org/abs/1912.10824

25
 A PREPRINT - N OVEMBER 5, 2020

6.1 OpenAI Gym

"Situation awareness is the perception of the elements in the environment within a volume of time and space, and the
comprehension of their meaning, and the projection of their status in the near future." — Endsley (1987)

The OpenAI Gym https://gym.openai.com/ (Blog196 | GitHub197 ) is a toolkit for developing and comparing
reinforcement learning algorithms. What makes the gym so great is a common API around environments.

Figure 28: Robotics Environments https://blog.openai.com/ingredients-for-robotics-research/

"By framing the approach within the popular OpenAI Gym framework, design firms can create more realistic
environments – for instance, incorporate strength of materials, safety factors, malfunctioning of components under
stressed conditions, and plug existing algorithms into this framework to optimize also for design aspects such as energy
usage, easy-of-manufacturing, or durability." — David Ha198

ú Getting Started with the OpenAI Gym, Colab Notebook199

How to create new environments for Gym200 . Minimal example with code and agent (evolution strategies on foo-v0):

1. Download gym-foo https://drive.google.com/file/d/1r2A8J9CJjIQNwss246gATeD0LLMzpUT-/

view?usp=sharing
2. cd gym-foo
3. pip install -e .
4. python ES-foo.py

He’re another more difficult (for the agent!) new environment for Gym (evolution strategies on foo-v3):

1. Download gym-foo-v3201
2. cd gym-foo-v3
3. pip install -e .
4. python ES-foo-v3.py

ú Create a New Environment (foo) from Scratch, Colab Notebook202

v OpenAI Gym Environment for Trading203 .
v Fantasy Football AI Environment https://github.com/njustesen/ffai.
v Create custom gym environments from scratch — A stock market example204 .
196
https://blog.openai.com/openai-gym-beta/
197
https://github.com/openai/gym
198
https://designrl.github.io
199
https://colab.research.google.com/drive/1fBDH7xfpwH9SKj5J9TAH9XOTGJF61vJZ
200
https://github.com/openai/gym/blob/master/docs/creating-environments.md
201
https://drive.google.com/file/d/1cGncsXJ56UUKCO9MaRWJVTnxiQEnLuxS/view?usp=sharing
202
https://colab.research.google.com/drive/1hXW5hQn1MO4kjgc2W2wjyTwDcId5QGCD
203
https://github.com/hackthemarket/gym-trading
204
https://towardsdatascience.com/creating-a-custom-openai-gym-environment-for-stock-trading-be532be3910e

26
 A PREPRINT - N OVEMBER 5, 2020

v Spot Mini Mini OpenAI Gym Environment. Maurice Rahme, blog205 et code206 .
v IKEA Furniture Assembly Environment https://clvrai.github.io/furniture/.
v Minimalistic Gridworld Environment https://github.com/maximecb/gym-minigrid.
v DoorGym: A Scalable Door Opening Environment and Baseline Agent, Urakami et al., 2019207 .
v gym-gazebo2, a toolkit for reinforcement learning using ROS 2 and Gazebo, Lopez et al., 2019208 .
v OFFWORLD GYM Open-access physical robotics environment for real-world reinforcement learning209 .
v Safety Gym: environments to evaluate agents with safety constraints https://github.com/openai/safety-gym.
v TensorTrade: An open source reinforcement learning framework for training, evaluating, and deploying robust
trading agents https://github.com/tensortrade-org/tensortrade.

6.2 DeepMind Lab

DeepMind Lab: A customisable 3D platform for agent-based AI research https://github.com/deepmind/lab.

• DeepMind Control Suite https://github.com/deepmind/dm_control.

• Convert DeepMind Control Suite to OpenAI Gym Envs https://github.com/zuoxingdong/dm2gym.

6.3 Unity ML-Agents

Unity ML Agents allows to create environments where intelligent agents (Single Agent, Cooperative and Competitive
Multi-Agent and Ecosystem) can be trained using RL, neuroevolution, or other ML methods https://unity3d.ai.

• Announcing ML-Agents Unity Package v1.0! Mattar et al.210 .

• Getting Started with Marathon Environments for Unity ML-Agents211 https://github.com/

Unity-Technologies/marathon-envs.

• Arena: A General Evaluation Platform and Building Toolkit for Multi-Agent Intelligence212 .

6.4 AI Habitat

AI Habitat enables training of embodied AI agents (virtual robots) in a highly photorealistic and efficient 3D simulator,
before transferring the learned skills to reality. By Facebook AI Research https://aihabitat.org/.
Why the name Habitat? Because that’s where AI agents live!

6.5 POET: Paired Open-Ended Trailblazer

Diversity is the premier product of evolution. Endlessly generate increasingly complex and diverse learning environ-
ments213 . Open-endedness could generate learning algorithms reaching human-level intelligence[23].

• Implementation of the POET algorithm https://github.com/uber-research/poet.

• Enhanced POET: Open-Ended Reinforcement Learning through Unbounded Invention of Learning Challenges
and their Solutions. Wang et al., 2020 https://arxiv.org/abs/2003.08536. Code214 .

27
 A PREPRINT - N OVEMBER 5, 2020

Figure 29: Edge TPU - Dev Board https://coral.ai/products/dev-board/

Figure 30: The world’s largest chip : Cerebras Wafer Scale Engine https://www.cerebras.net

7 Deep-Learning Hardware

v Which GPU(s) to Get for Deep Learning, by Tim Dettmers215 .

v A Full Hardware Guide to Deep Learning, by Tim Dettmers216 .
v Jetson Nano. A small but mighty AI computer to create intelligent systems217 .
v Build AI that works offline with Coral Dev Board, Edge TPU, and TensorFlow Lite, by Daniel Situnayake218 .
205
https://moribots.github.io/project/spot-mini-mini
206
https://github.com/moribots/spot_mini_mini
207
https://arxiv.org/abs/1908.01887
208
https://arxiv.org/abs/1903.06278
209
https://gym.offworld.ai
210
https://blogs.unity3d.com/2020/05/12/announcing-ml-agents-unity-package-v1-0/
211
https://towardsdatascience.com/gettingstartedwithmarathonenvs-v0-5-0a-c1054a0b540c
212
https://arxiv.org/abs/1905.08085
213
https://eng.uber.com/poet-open-ended-deep-learning/
214
http://github.com/uber-research/poet
215
http://timdettmers.com/2019/04/03/which-gpu-for-deep-learning/
216
http://timdettmers.com/2018/12/16/deep-learning-hardware-guide/
217
https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano/
218
https://medium.com/tensorflow/build-ai-that-works-offline-with-coral-dev-board-edge-tpu-and-tensorflow-lite-70

28
 A PREPRINT - N OVEMBER 5, 2020

8 Deep-Learning Software
8.1 TensorFlow

TensorFlow Hub is a library for reusable ML modules https://www.tensorflow.org/hub. Tutorials219 .

TensorFlow.js allows machine learning to happen within the web browser https://www.tensorflow.org/js/.

• TF-Coder https://goo.gle/3gwTbB6.
• TensorFlow Lite for Microcontrollers220 .
• Intro to Keras for Researchers. Colab221 .
• Introduction to Keras for Engineers. Colab222 .
• TensorBoard in Jupyter Notebooks223 . Colab224 .
• TensorFlow 2.0 + Keras Crash Course. Colab225 .
• tf.keras (TensorFlow 2.0) for Researchers: Crash Course. Colab226 .
• TensorFlow Tutorials https://www.tensorflow.org/tutorials.
• Exploring helpful uses for BERT in your browser with TensorFlow.js227 .
• TensorFlow 2.0: basic ops, gradients, data preprocessing and augmentation, training and saving. Colab228 .

8.2 PyTorch
• PyTorch primer. Colab229 .
• Get started with PyTorch, Cloud TPUs, and Colab230 .
• MiniTorch https://minitorch.github.io/index.html
• Effective PyTorch https://github.com/vahidk/EffectivePyTorch
• PyTorch internals http://blog.ezyang.com/2019/05/pytorch-internals/
• PyTorch Lightning Bolts https://github.com/PyTorchLightning/pytorch-lightning-bolts

9 AI Art | A New Day Has Come in Art Industry

The code (art-DCGAN) for the first artificial intelligence artwork ever sold at Christie’s auction house (Figure 31) is a
modified implementation of DCGAN focused on generative art: https://github.com/robbiebarrat/art-dcgan.

• The Creative AI Lab https://creative-ai.org.

• TensorFlow Magenta. An open source research project exploring the role of ML in the creative process.231 .
• Magenta Studio. A suite of free music-making tools using machine learning models!232 .
• Style Transfer Tutorial https://colab.research.google.com/github/tensorflow/docs/blob/
master/site/en/r2/tutorials/generative/style_transfer.ipynb
219
https://www.tensorflow.org/hub/tutorials
220
https://www.tensorflow.org/lite/microcontrollers
221
https://colab.research.google.com/drive/1qKPITTI879YHTxbTgYW_MAWMHFkbOBIk
222
https://colab.research.google.com/drive/1lWUGZarlbORaHYUZlF9muCgpPl8pEvve
223
https://www.tensorflow.org/tensorboard/tensorboard_in_notebooks
224
https://colab.research.google.com/github/tensorflow/tensorboard/blob/master/docs/tensorboard_in_
notebooks.ipynb
225
https://colab.research.google.com/drive/1UCJt8EYjlzCs1H1d1X0iDGYJsHKwu-NO
226
https://colab.research.google.com/drive/14CvUNTaX1OFHDfaKaaZzrBsvMfhCOHIR
227
https://blog.tensorflow.org/2020/03/exploring-helpful-uses-for-bert-in-your-browser-tensorflow-js.
html
228
https://colab.research.google.com/github/zaidalyafeai/Notebooks/blob/master/TF_2_0.ipynb
229
https://colab.research.google.com/drive/1DgkVmi6GksWOByhYVQpyUB4Rk3PUq0Cp
230
https://medium.com/pytorch/get-started-with-pytorch-cloud-tpus-and-colab-a24757b8f7fc
231
https://magenta.tensorflow.org
232
https://magenta.tensorflow.org/studio

29
 A PREPRINT - N OVEMBER 5, 2020

Figure 31: On October 25, 2018, the first AI artwork ever sold at Christie’s auction house fetched USD 432,500.

• AI x AR Paper Cubes https://experiments.withgoogle.com/paper-cubes.

• Photo Wake-Up https://grail.cs.washington.edu/projects/wakeup/.
• COLLECTION. AI Experiments https://experiments.withgoogle.com/ai.

"The Artists Creating with AI Won’t Follow Trends; THEY WILL SET THEM." — The House of Montréal.AI Fine Arts

v Tuning Recurrent Neural Networks with Reinforcement Learning233 .

v MuseNet. Generate Music Using Many Different Instruments and Styles!234 .
v Infinite stream of machine generated art. Valentin Vieriu https://art42.net.
v Deep Multispectral Painting Reproduction via Multi-Layer, Custom-Ink Printing. Shi et al.235 .
v Discovering Visual Patterns in Art Collections with Spatially-consistent Feature Learning. Shen et al.236 .
v Synthesizing Programs for Images using Reinforced Adversarial Learning, Ganin et al., 2018237 . Agents238 .

10 AI Macrostrategy: Aligning AGI with Human Interests

Montréal.AI Governance: Policies at the intersection of AI, Ethics and Governance.

Figure 32: A Map of Ethical and Right-Based Approaches https://ai-hr.cyber.harvard.edu/primp-viz.html

233
https://magenta.tensorflow.org/2016/11/09/tuning-recurrent-networks-with-reinforcement-learning
234
https://openai.com/blog/musenet/
235
http://people.csail.mit.edu/liangs/papers/ToG18.pdf
236
https://arxiv.org/pdf/1903.02678.pdf
237
http://proceedings.mlr.press/v80/ganin18a.html
238
https://github.com/deepmind/spiral

30
 A PREPRINT - N OVEMBER 5, 2020

"(AI) will rank among our greatest technological achievements, and everyone deserves to play a role in shaping it." —
Fei-Fei Li

Figure 33: The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies. Zheng et al.
https://arxiv.org/abs/2004.13332

v AI Index. http://aiindex.org.
v The State of AI Report. https://www.stateof.ai/.
v Malicious AI Report. https://arxiv.org/pdf/1802.07228.pdf.
v Artificial Intelligence and Human Rights. https://ai-hr.cyber.harvard.edu.
v The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies, Zheng et al.239 . Blog240 .
v Ethically Aligned Design, First Edition241 . From Principles to Practice https://ethicsinaction.ieee.org.
v ADDRESS PREPARED BY POPE FRANCIS FOR THE PLENARY ASSEMBLY OF THE PONTIFICAL
ACADEMY FOR LIFE242 .

"It’s springtime for AI, and we’re anticipating a long summer." — Bill Braun

References
[1] Mnih et al. Human-Level Control Through Deep Reinforcement Learning. In Nature 518, pages 529–533. 26
February 2015. https://storage.googleapis.com/deepmind-media/dqn/DQNNaturePaper.pdf
[2] Yann LeCun, Yoshua Bengio and Geoffrey Hinton. Deep Learning. In Nature 521, pages 436–444. 28 May 2015.
https://www.cs.toronto.edu/~hinton/absps/NatureDeepReview.pdf
[3] Goodfellow et al. Generative Adversarial Networks. arXiv preprint arXiv:1406.2661, 2014. https://arxiv.
org/abs/1406.2661
[4] Yoshua Bengio, Andrea Lodi, Antoine Prouvost. Machine Learning for Combinatorial Optimization: a Method-
ological Tour d’Horizon. arXiv preprint arXiv:1811.06128, 2018. https://arxiv.org/abs/1811.06128
[5] Brockman et al. OpenAI Gym. 2016. https://gym.openai.com
[6] Devlin et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint
arXiv:1810.04805, 2018. https://arxiv.org/abs/1810.04805
[7] Dai et al. Semi-supervised Sequence Learning. arXiv preprint arXiv:1511.01432, 2015. https://arxiv.org/
abs/1511.01432
[8] Mnih et al. Asynchronous Methods for Deep Reinforcement Learning. arXiv preprint arXiv:1602.01783, 2016.
https://arxiv.org/abs/1602.01783
239
https://arxiv.org/abs/2004.13332
240
https://blog.einstein.ai/the-ai-economist/
241
https://standards.ieee.org/content/dam/ieee-standards/standards/web/documents/other/ead1e.pdf
242
http://w2.vatican.va/content/francesco/en/speeches/2020/february/documents/papa-francesco_
20200228_accademia-perlavita.html

31
 A PREPRINT - N OVEMBER 5, 2020

[9] Schulman et al. Proximal Policy Optimization Algorithms. arXiv preprint arXiv:1707.06347, 2017. https:
//arxiv.org/abs/1707.06347
[10] Mnih et al. Playing Atari with Deep Reinforcement Learning. DeepMind Technologies, 2013. https://www.cs.
toronto.edu/~vmnih/docs/dqn.pdf
[11] Ha et al. Recurrent World Models Facilitate Policy Evolution. arXiv preprint arXiv:1809.01999, 2018. https:
//arxiv.org/abs/1809.01999
[12] Kenneth et al. Designing neural networks through neuroevolution. In Nature Machine Intelligence VOL 1, pages
24–35. January 2019. https://www.nature.com/articles/s42256-018-0006-z.pdf
[13] So et al. The Evolved Transformer. arXiv preprint arXiv:1901.11117, 2019. https://arxiv.org/abs/1901.
11117
[14] Silver et al. Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm. arXiv
preprint arXiv:1712.01815, 2017. https://arxiv.org/abs/1712.01815
[15] Silver et al. AlphaGo Zero: Learning from scratch. In DeepMind’s Blog, 2017. https://deepmind.com/blog/
alphago-zero-learning-scratch/
[16] Andrychowicz et al. Learning to learn by gradient descent by gradient descent. arXiv preprint arXiv:1606.04474,
2016. https://arxiv.org/abs/1606.04474
[17] Nichol et al. Reptile: A Scalable Meta-Learning Algorithm. 2018. https://blog.openai.com/reptile/
[18] Frans et al. Meta Learning Shared Hierarchies. arXiv preprint arXiv:1710.09767, 2017. https://arxiv.org/
abs/1710.09767
[19] Zoph and Le, 2017 Neural Architecture Search with Reinforcement Learning. arXiv preprint arXiv:1611.01578,
2017. https://arxiv.org/abs/1611.01578
[20] Finn et al., 2017 Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. arXiv preprint
arXiv:1703.03400, 2017. https://arxiv.org/abs/1703.03400
[21] Salimans et al. Evolution Strategies as a Scalable Alternative to Reinforcement Learning. 2017. https:
//blog.openai.com/evolution-strategies/
[22] Lehman et al. The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary
Computation and Artificial Life Research Communities. arXiv preprint arXiv:1803.03453, 2018. https://arxiv.
org/abs/1803.03453
[23] Wang et al. Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse
Learning Environments and Their Solutions. arXiv preprint arXiv:1901.01753, 2019. https://arxiv.org/abs/
1901.01753
[24] Foerster et al. Learning to Model Other Minds. 2018. https://blog.openai.com/
learning-to-model-other-minds/
[25] Rabinowitz et al. Machine Theory of Mind. arXiv preprint arXiv:1802.07740, 2018. https://arxiv.org/abs/
1802.07740

32