Prosek Point - Building A

Prosecká 851/64,

190 00

Prague, Czech Republic

  • Black Facebook Icon
  • Black Twitter Icon
  • Black YouTube Icon

Check out our sister company Keen Software House

Get in touch at

GoodAI Bibliography


The present bibliography lists sources of interest from the viewpoint of GoodAI’s general AI research.

Our effort is informed by the following working assumptions:

  • Agents need the ability to grow their cognitive resources as they encounter progressively harder problems;

  • The growth of agents can be aided by human-designed curricula.


Most of the entries are recent papers of specific interest to this research direction. In addition, we include

  • Books and a few older papers with valuable background information;

  • Presentation slides and videos for some material not covered in papers.


The material is organized by research area:

  1. AI school: curricula and learning environments

  2. Machine learning topics: Growing architectures and more

  3. Cognitive architectures

  4. AI safety

AI school

This section contains material on curricula and learning environments.


Curriculum learning

Yoshua Bengio, Jérôme Louradour, Ronan Collobert, Jason Weston


This paper introduces curriculum learning: learning accelerated by presenting easy examples first and progressively increasing the difficulty. The paper was a product of the BabyAI project. See also Bengio’s presentation on Deep Architectures for Baby AI.

Type: Very relevant


A Roadmap towards Machine Intelligence

Tomas Mikolov, Armand Joulin, Marco Baroni


An interesting proposal for a language-based environment for raising intelligent agents.

Type: Very relevant


Artificial Tasks for Artificial Intelligence (video, slides)

Antoine Bordes, Jason Weston, Sumit Chopra, Tomas Mikolov, Armand Joulin, Léon Bottou


On the shortcomings of present AI systems and how to learn for AGI.

Type: Very relevant


Cognitive development (book)

John H. Flavell, Patricia H. Miller, Scott A. Miller

This textbook describes the stages of human cognitive development.


Type: Very relevant


Why Artificial Intelligence Needs a Task Theory

Kristinn R. Thórisson, Jordi Bieger, Thröstur Thorarensen, Jóna S. Sigurðardóttir, Bas R. Steunebrink


An AI task theory, the authors argue, can give us more rigorous ways of comparing and evaluating intelligent behavior.

Type: Somewhat relevant

Machine learning


Growing architectures

Growing architectures learn new things while retaining existing knowledge.


Powerplay: Training an increasingly general problem solver by continually searching for the simplest still unsolvable problem

Jürgen Schmidhuber


The author presents a framework for automatically discovering problems inspired by playful behavior in animals and humans.

Type: Very relevant


Progressive Neural Networks

Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, Raia Hadsell


Progressive neural networks adapt to new tasks by growing new columns while retaining previously acquired knowledge.

Type: Very relevant


The Cascade-Correlation Learning Architecture

Scott Fahlman Christian, Christian Lebiere


The Cascade-Correlation architecture accelerates learning by adding one hidden neuron at the time, keeping the preceding hidden weights frozen.

Type: Very relevant


Learning without Forgetting

Zhizhong Li, Derek Hoiem


The authors present a method for adding capabilities to convolutional neural networks (CNNs) while retaining existing capabilities.

Type: Relevant


Growing Recursive Self-Improvers

Bas R. Steunebrink, Kristinn R. Thórisson, Jürgen Schmidhuber


Introduces experience-based AI, a class of systems capable of continuous self-improvement.

Type: Somewhat relevant


Bounded Recursive Self-Improvement

Eric Nivel, Kristinn R. Thórisson, Bas R. Steunebrink, Haris Dindo, Giovanni Pezzulo, M. Rodriguez, C. Hernandez, Dimitri Ognibene, Jürgen Schmidhuber, Ricardo Sanz, Helgi Páll Helgason, Antonio Chella, Gudberg K. Jonsson


The authors prototype “a machine that becomes increasingly better at behaving in underspecified circumstances, in a goal-directed way, on the job, by modeling itself and its environment as experience accumulates.”

Type: Somewhat relevant

Self-optimizing architectures

Determining the topology and hyperparameters for a model is labor intensive. The architectures described here learn their structure from the training data.


Online incremental feature learning with denoising autoencoders

Guanyu Zhou, Kihyuk Sohn, Honglak Lee


This paper introduces an algorithm for learning features incrementally and determining the optimal model complexity from the data.

Type: Relevant


Convolutional Neural Fabrics

Shreyas Saxena, Jakob Verbeek


Selecting the optimal convolutional network architecture for a given task remains an open problem. The authors propose a “neural fabric” that learns its structure from the data.

Type: Somewhat relevant

Deep architectures


Deep learning

Yann LeCun, Yoshua Bengio, Geoffrey Hinton


A brief and readable overview of recent progress in deep learning.

Type: Very relevant


Deep learning in neural networks: An overview

Jürgen Schmidhuber


A more comprehensive overview of deep learning.

Type: Very relevant

Extracting and Composing Robust Features with Denoising Autoencoders Pascal Vincent, Hugo Larochelle, Yoshua Bengio, Pierre-Antoine Manzagol


Noise robust learning of Autoencoders.

Type: Relevant

Compositional learning

Compositional learning allows a learner to extends its abilities by recombining existing skills.


Learning to compose neural networks for question answering

Jacob Andreas, Marcus Rohrbach, Trevor Darrell, Dan Klein


Learning how to compose (modules) neural networks to solve more complex tasks. One part of the algorithm learns how to do the composition of the modules and, in parallel, the specific modules are learned.

Type: Relevant

Transfer learning

This section provides some background on knowledge transfer between different learners.


A survey on transfer learning

Sinno Jialin Pan, Qiang Yang


This paper provides a systematic overview of approaches to transfer learning.

TYPE: Very relevant

Sequence learning

This section discusses ways of learning temporal and other sequential patterns.


How the brain might work: A hierarchical and temporal model for learning and recognition (book)

Dileep George


Presents a theory about bottom-up and top-down processing of information, the importance of both of those directions, the role of spatial and temporal patterns and how is it connected to neocortex.

Type: Very relevant


Temporal Pattern Processing

DeLiang Wang


This paper provides an overview of approaches to the representation and processing of temporal patterns.

Type: Somewhat relevant

Program learning

By learning programs or algorithms an agent can represent and reuse procedural knowledge.


Neural Programmer-Interpreters

Scott Reed, Nando de Freitas


Winner of best paper award at ICLR 2016. This paper uses strong supervision to train a large controller which is able to compose hard coded instructions into programs. The methodology is very general as multiple instruction sets are shown on a diverse array of tasks. And the approach generalises to larger task variants quite well.

Type: Very relevant


Neural Programmer: Inducing Latent Programs with Gradient Descent

Arving Neelakantan, Quoc V. Lee, Ilya Sutskever


This paper uses weak supervision and gradient descent to train an agent to perform table lookups on data using a number of intrinsic operators. The agent is trained only with what the answer should be, so it tries to converge onto a sequence of instructions which satisfy the answer. If the NPI paper is for the architecture, this is for the method.

Type: Very relevant


Neural GPUs Learn Algorithms

Łukasz Kaiser, Ilya Sutskever


Instead of creating deep models, this model is designed to be wide. It is based on the convolution-gated recurrent units and is able to do e.g. addition or multiplication. It can be trained on small examples and is able to generalize to long instances.

Type: Somewhat relevant


Programming with a Differentiable Forth Interpreter

Sebastian Riedel, Matko Bošnjak, Tim Rocktäschel


The proposed system incorporates prior procedural knowledge as program sketches with slots that can be filled with learnable behaviour.

Type: Somewhat relevant

Spike-timing-dependent plasticity (STDP)

The papers presented here apply STDP to machine learning problems.


Learning dynamic Boltzmann machines with spike-timing dependent plasticity

Takayuki Osogami, Makoto Otsuka


Unsupervised learning rule and temporal pattern representations for Boltzmann machines.           

Type: Very relevant           

Solving the distal reward problem through linkage of STDP and dopamine signaling                   

E. M. Izhikevich


On the problem of delayed reward in biological and artificial neural networks.                        

Type: Somewhat relevant

Reinforcement learning

This section provides some papers describing recent progress in reinforcement learning.


Data-Efficient Learning of Feedback Policies from Image Pixels using Deep Dynamical Models

John-Alexander M. Assael, Niklas Wahlstrom, Thomas B. Schon, Marc Peter Deisenroth


Principle of network composed of multiple parts, one for encoding the current state of the environment into more compressed representation. The other for predicting next state in this compressed state.

Type: Relevant


Tunable and Generic Problem Instance Generation for Multi-objective Reinforcement Learning

Garrett, Deon, Jordi Bieger, Kristinn R. Thórisson


The paper introduces a tool to help researchers identify causes of failure and success of reinforce learning algorithms. The focus is primarily on multi-objective reinforcement learning.

Type: Somewhat relevant


Control of Memory, Active Perception, and Action in Minecraft

Junhyuk Oh, Valliappa Chockalingam, Satinder Singh, Honglak Lee


How to use memory with DQN. It can remember where the stuff is placed in the environment, how to navigate there, and the designer can even visualized what it does.

Type: Somewhat relevant



Representing Word Meaning and Order Information in a Composite Holographic Lexicon

Michael N. Jones, Douglas J. K. Mewhort


Unsupervised extraction of meaning of words from text, which is based on holographic representations.

Type: Somewhat relevant


Distributed Representations of Sentences and Documents

Quoc V. Le, Tomas Mikolov


The paper introduces the Paragraph Vector, an unsupervised framework that learns continuous distributed vector representations for variable-length texts.

Type: Somewhat relevant

Cognitive architectures


Thinking, Fast and Slow (book)

Daniel Kahneman


High level point of view to the mind. What it does and why? How we solve some problems and what we decide to remember? The book provides summary (high-level) summary of the Mr. Kahneman work. The key point is that our mind consist of two major parts: lazy system that instantly decides what to do, and more accurate decision making part that carefully thinks about the next action. One system is is faster, another one slower. The book discuss how, and why, we choose between these two and how it affects us.

Type: Relevant


How to Build a Brain: A Neural Architecture for Biological Cognition (book)

Chris Eliasmith


Description of heterogeneous modular engineered architecture, which combines holographic representations and spiking neural networks to produce biologically plausible information processing during solving selected cognitive tasks.

Type: Somewhat relevant


Cognitive Architectures and Autonomy: A Comparative Review

Kristinn R. Thórisson, Helgi Páll Helgasson


Good review and comparison of some selected cognitive architecture designs.

Type: Relevant


How to Create a Mind (book)

Ray Kurzweil


Overall discussion on the probable artificial mind architecture.

Type: Somewhat relevant

AI safety


Superintelligence: Paths, dangers, strategies (book)

Nick Bostrom


An influential discussion of AI risks.

Type: Very relevant