AGI 18

Talk Session 1, 17:00-18:00, Wednesday 22 August 2018: Algorithms

André Cyr, Frédéric Thériault, Matthew Ross and Sylvain Chartier, “Associative memory: an embodied spiking neural network robotic implementation.”
This article proposes a minimalist bio-inspired associative memory (AM) mechanism that can be implemented in a spiking neural network acting as a controller for virtual and real robots. As such, several main features of a general AM concept were reproduced. Using the strength of temporal coding at the single spike resolution level, this study approaches the AM phenomenon with basic examples in the visual modality. Specifically, these embedded neural architectures include varying time delays in synaptic links and asymmetry in the spike-timing dependent plasticity learning rules to solve visual tasks of pattern-matching, pattern-completion and noise-tolerance for autoassociative and heteroassociative memories. This preliminary work could serve as a step toward future comparative analysis with traditional artificial neural network paradigm.
Slides

Nil Geisweiller. “Partial Operator Induction with Beta Distributions,“
A specialization of Solomonoff Operator Induction considering partial operators described by second order probability distributions, and more specifically Beta distributions, is introduced. An estimate to predict the second order probability of new data, obtained by averaging the second order distributions of partial operators, is derived. The problem of managing the partiality of the operators is presented. A simplistic solution based on estimating the Kolomogorov complexity of perfect completions of partial operators is given.
Slides

Talk Session 2, 11:45-13:00, Thursday 23 August 2018: Cognitive Architectures

Khin Hua Ng, Zhiyuan Du and Gee Wah Ng.”DSO Cognitive Architecture: Implementation and Validation of the Global Workspace Enhancement”
An enhanced DSO Cognitive Architecture design was recently introduced to augment its cognitive functions by incorporating the Global Workspace Theory. A computational implementation of this new design is described in detail in this paper. The implementation is built as a distributed system with parallel pipelines of specialised processes, executing asynchronously. Competition initiated by these processes, and facilitated by the attention mechanism and global broadcast mechanism, leads to pipelines being dynamically created and allows disconnected pipelines to influence the processing of others. To validate the implementation, it was applied to a traffic control problem and experimental results showed increase in performance gain using the enhanced cognitive architecture.

Joscha Bach and Katherine Gallagher. “Request Confirmation Networks in MicroPsi 2”]
To combine neural learning with the sequential detection of hierarchies of sensory features, and to facilitate planning and script execution, we propose Request Confirmation Networks (ReCoNs). ReCoNs are spreading activation networks with units that contain an activation and a state, and are connected by typed directed links that indicate partonomic relations and spatial or temporal succession. By passing activation along the links, ReCoNs can perform both neural computations and controlled script execution. We demonstrate the application of ReCoNs in the context of performing simple arithmetic, based on camera images of mathematical expressions.

Xiang Li, Patrick Hammer, Pei Wang and Hongling Xie. “Functionalist Emotion Model in NARS”
Emotions play a crucial role in different cognitive functions, such as action selection and decision-making processes. This paper describes a new appraisal model for the emotion mechanism of NARS, an AGI system. Different from the previous appraisal model where emotions are triggered by the specific context, the new appraisal evaluates the relations between the system and its goals, based on a new set of criteria, including desirability, belief, and anticipation. Our work focuses on the functions of emotions and how emotional reactions could help NARS to improve capacities of different cognitive functions.
Slides

Talk Session 3, 17:00-18:30, Thursday 23 August 2018: Modeling and Evaluation

Eray Ozkural. “Zeta Distribution and Transfer Learning Problem”
We explore the relations between the zeta distribution and algorithmic information theory via a new model of the transfer learning problem. The program distribution is approximated by a zeta distribution with parameter near $1$. We model the training sequence as a stochastic process. We analyze the upper temporal bound for learning a training sequence and its entropy rates, assuming an oracle for the transfer learning problem. We argue from empirical evidence that power-law models are suitable for natural processes. Four sequence models are proposed. Random typing model is like no-free lunch where transfer learning does not work. Zeta process independently samples programs from the zeta distribution. A model of common sub-programs inspired by genetics uses a database of sub-programs. An evolutionary zeta process samples mutations from Zeta distribution. The
analysis of stochastic processes inspired by evolution suggest that AI may be feasible in nature, countering
no-free lunch sort of arguments.
Slides

Mark Wernsdorfer. “How Failure Facilitates Success”
For robotic systems that interact with real-world environments, it is plain impossible to cover all the underlying patterns that govern the environment’s behavior. Real environments may as well be considered erratic. One way to deal with this fundamental uncertainty is to describe this behavior at least with consistent probabilities. This paper proposes another another way: to incorporate failed expectations into a deterministic model that can provide useful information in case the same failure happens again. A new type of model is developed that embraces fallibility. It uses wrong predictions from the past to provide a more appropriate description of the future. This type of model is based on a new conception of partially observable systems. Like similar conceptions, it assumes a hidden state in these systems. Unlike previous conceptions, it does not suggest that transitions between hidden states can be predicted prior to the fact. Instead, their effects have to be observed first, as behavior that would not have been possible according to the model’s current conception of the system. This discrepancy enables the state of the model to self-correct in a continual coupling with the system that it describes.
Slides

Giacomo Spigler. “The Temporal Singularity: time-accelerated simulated civilizations and their implications”
Provided significant future progress in artificial intelligence and computing, it may ultimately be possible to create multiple Artificial General Intelligences (AGIs), and possibly entire societies living within simulated environments. In that case, it should be possible to improve the problem solving capabilities of the system by increasing the speed of the simulation. If a minimal simulation with sufficient capabilities is created, it might manage to increase its own speed by accelerating progress
in science and technology, in a way similar to the Technological Singularity. This may ultimately lead to large simulated civilizations unfolding at extreme temporal speedups, achieving what from the outside would look like a Temporal Singularity. Here we discuss the feasibility of the minimal simulation and the potential advantages, dangers, and connection to the Fermi paradox of the Temporal Singularity. The medium-term importance of the topic derives from the amount of computational power required to start the process, which could be available within the next decades, making the Temporal Singularity theoretically possible before the end of the century.
Slides

Ondřej Vadinský. “Towards General Evaluation of Intelligent Systems: Using Semantic Analysis to Improve Environments in the AIQ Test”
This paper conducted a semantic analysis of environment programs that are used in the Algorithmic Intelligence Quotient test to evaluate the intelligence of agents. The analysis identified several classes of programs that are non-discriminative or contain pointless code adversely affecting the testing process. Extensions of the test were implemented and verified that reduce the proportion of problematic programs thus increasing the suitability of the Algorithmic Intelligence Quotient test as a general artificial intelligence evaluation method.
Slides

Talk Session 4, 9:00-10:30, Friday 24 August 2018: Philosophy, Social Sciences, and Knowledge Acquistion

Jeremy Turner and Steve Dipaola. “Transforming Kantian Aesthetic Principles into Qualitative Hermeneutics for Contemplative AGI Agents”
This paper introduces an interdisciplinary qualitative hermeneutic approach to the engineering and computer science methodological paradigm(s) for assessing a contemplative-agent’s cognitive capabilities at a level corresponding to Artificial General Intelligence (AGI). This paper has utilized cognitive intensity levels from Kantian aesthetic philosophy to qualitatively re-address Russell and Norvig’s canonical agent-categories as they scale upward towards AGI cognition levels. These Kantian levels allow the AGI-agent designer to consider the cognitive interplay between computational representations of the imagination and reason as they relate to motivationally-nuanced teleological notions of self-interest versus disinterestedness. While the AGI level is the thematic focus, lower intensity levels are also introduced in order to set the appropriate cognitive benchmarks for higher levels corresponding to truly contemplative AGI-agents. This paper first contextualizes and endorses Kant’s analytical framework before discussing the appropriately corresponding agent-categories. This paper concludes with a brief discussion of the particular methodological and hermeneutic contribution of Kant’s aesthetic philosophical framework to the AGI domain.
Slides

Jakub Mlynář, Hamed Alavi, Himanshu Verma and Lorenzo Cantoni. “Towards a Sociological Conception of Artificial Intelligence”
Social sciences have been always formed and influenced by the development of society, adjusting the conceptual, methodological, and theoretical frameworks to emerging social phenomena. In recent years, with the leap in the advancement of Artificial Intelligence (AI) and the proliferation of its everyday applications, “non-human intelligent actors” are increasingly becoming part of the society. This is manifested in the evolving realms of smart home systems, autonomous vehicles, chatbots, intelligent public displays, etc. In this paper, we present a prospective research project that takes one of the pioneering steps towards establishing a “distinctively sociological” conception of AI. Its first objective is to extract the existing conceptions of AI as perceived by its technological developers and (possibly differently) by its users. In the second part, capitalizing on a set of interviews with experts from social science domains, we will explore the new imaginable conceptions of AI that do not originate from its technological possibilities but rather from societal necessities. The current formal ways of defining AI are grounded in the technological possibilities, namely machine learning methods and neural network models. But what exactly is AI as a social phenomenon, which may act on its own, can be blamed responsible for ethically problematic behavior, or even endanger people’s employment? We argue that such conceptual investigation is a crucial step for further empirical studies of phenomena related to AI’s position in current societies, but also will open up ways for critiques of new technological advancements with social consequences in mind from the outset.
Slides

Jordi Bieger and Kristinn R. Thorisson. “Task Analysis for Teaching Cumulative Learners”
A generally intelligent machine (AGI) should be able to learn a wide range of tasks. Knowledge acquisition in complex and dynamic task-environments cannot happen all-at-once, and AGI-aspiring systems must thus be capable of *cumulative learning*: efficiently making use of existing knowledge during learning, supporting incremental increases in the scope of ability and knowledge — without catastrophic forgetting. Where relevant expertise is at hand the learning process can be aided by curriculum-based teaching, where a teacher divides a high-level task up into smaller and simpler pieces and presents them in an order that facilitates learning. Creating such a curriculum can benefit from expert knowledge of (a) the task domain, (b) the learning system itself, and (c) general teaching principles. Curriculum design for AI systems has so far been rather ad-hoc and limited to systems incapable of cumulative learning. We present a task analysis methodology that utilizes expert knowledge and can inform the construction of teaching curricula for cumulative learners. Inspired in part by methods from knowledge engineering and functional requirements analysis, our strategy decomposes high-level tasks in three ways based on involved actions, features and functionality. We show how this methodology can be used for a (simplified) *arrival control* task from the air traffic control domain, where extensive expert knowledge is available and teaching cumulative learners is required to facilitate the safe and trustworthy automation of complex workflows.
Slides

Kristinn R. Thórisson and Arthur Talbot. “Cumulative Learning With Causal-Relational Models”
In the quest for artificial general intelligence (AGI) questions remain unanswered about what kinds of representation are needed for the kind of flexibility and dynamism exhibited by the human mind. While some capabilities of a fully trained human mind have been replicated in a machine, a capacity for continued learning of many domains has yet to be realized, and proposals for how to achieve such continuous performance improvement and cumulative learning learning remain scarce. In particular, learning systems that can effectively carry skills learned priorly over to similar domains, or even very different ones – while seeming necessary feature of any AGI – have not been demonstrated to date. In this paper we describe a cumulative learning mechanism that produces causal-relational models to represent its environment, predict events and achieve goals in it. Such models, coupled with an appropriate modeling process, result in increasingly precise knowledge of a domain, generalize to other domains, and can run continuously throughout the lifetime of the agent that employs it. The methods have been implemented, with results in prior papers describing an autonomous agent that can learn complex human tasks and pragmatic natural language by observation. Here we focus on key principles of the modeling method and explain how it implements effective cumulative learning.
Slides

Talk Session 5, 14:45-16:45, Friday 24 August 2018: Computer Vision, Language Learning, Deep Learning

Alexey Potapov, Sergey Rodionov, Maxim Peterson, Oleg Scherbakov, Innokentii Zhdanov and Nikolai Skorobogatko. “Vision System for AGI: Problems and Directions”
What frameworks and architectures are necessary to create a vision system for AGI? In this paper, we propose a formal model that states the task of perception within AGI. We show the role of discriminative and generative models in achieving efficient and general solution of this task, thus specifying the task in more detail. We discuss some existing generative and discriminative models and demonstrate their insufficiency for our purposes. Finally, we discuss some architectural dilemmas and open questions.

Alexey Potapov, Innokentii Zhdanov, Oleg Scherbakov, Nikolai Skorobogatko, Hugo Latapie and Enzo Fenoglio. “Semantic Image Retrieval by Uniting Deep Neural Networks and Cognitive Architectures”
Image and video retrieval by their semantic content has been an important and challenging task for years, because it ultimately requires bridging the symbolic/subsymbolic gap. Recent successes in deep learning enabled detection of objects belonging to many classes greatly outperforming traditional computer vision techniques. However, deep learning solutions capable of executing retrieval queries are still not available. We propose a hybrid solution consisting of a deep neural network for object detection and a cognitive architecture for query execution. Specifically, we use YOLOv2 and OpenCog. Queries allowing the retrieval of video frames containing objects of specified classes and specified spatial arrangement are implemented.
Slides for the Two Papers Above

Eray Ozkural. “The Foundations of Deep Learning with a Path Towards General Intelligence”Like any formal field of science, AI may be approached axiomatically. We formulate requirements for a general-purpose, human-level AI system in terms of postulates. We review the methodology of deep learning, examining the explicit and tacit assumptions in deep learning research. Deep Learning methodology seeks to overcome limitations in traditional machine learning research as it combines facets of model richness, generality, and practical applicability. The methodology so far has produced outstanding results due to a productive synergy of function approximation, under plausible assumptions of irreducibility and the efficiency of back-propagation family of algorithms. We examine these winning traits of deep learning, and also observe the various known failure modes of deep learning. We conclude by giving recommendations on how to extend deep learning methodology to cover the postulates of general-purpose AI including modularity, and cognitive architecture. We also relate deep learning to advances in theoretical neuroscience research.
Slides

Mark Wernsdorfer. “A Time-critical Simulation of Language Comprehension”
Language comprehension is usually not understood as a time-critical task. Algorithms that simulate language comprehension are therefore conceded polynomial runtime and memory complexity as well as multiple passes over the same data. In most cases, practical applications for the computational comprehension of language are in fact not time-critical. Nonetheless, however, humans do perform the same task on-line with a single pass over the data (e.g. with speech or text) and without a polynomial increase in processing time. This calls for a more genuinely cognitive algorithmic approach to simulate the process of language comprehension. A formal conception of language is presented in this paper. A language model is developed from this conception and an algorithm is inferred that enables to generate such a model on-line and from a single pass over a linguistic sequence. The quality of the model is evaluated qualitatively, by comparing the resulting hierarchical segmentation to linguistic segmentations (e.g. syllables, words, sentences). The underlying algorithm is evaluated comparatively by contrasting the predictive performance of the generated model over time with the performance of an order-n Markov model that has been generated on-line as well.
Slides

Peter Sutor, Douglas Summers-Stay and Yiannis Aloimonos. “A Computational Theory for Life-Long Learning of Semantics”
Semantic vectors are learned from data to express semantic relationships between elements of information, for the purpose of solving and informing downstream tasks. Other models exist that learn to map and classify supervised data. However, the two worlds of learning rarely interact to inform one another dynamically, whether across types of data or levels of semantics, in order to form a unified model. We explore the research problem of learning these vectors and propose a framework for learning the semantics of knowledge incrementally and online, across multiple mediums of data, via binary vectors. We discuss the aspects of this framework to spur future research on this approach and problem.
Slides

Anton Kolonin, Andres Suarez, Alex Glushchenko, Ben Goertzel, Claudia Castillo, Linas Vepstas, Man Hin Leung and Oleg Baskov. “Unsupervised Language Learning in OpenCog”
We discuss technology capable to learn language with neither grammar, teachers, feedback, grounding, nor marked-up corpora. While the entire goal may be too ambitious and not achievable to full extent, we try to see how far we can advance learning grammar to substantial extent and basics of semantics at least. We present the current approach employed in the open source OpenCog Artificial Intelligence Platform, describe the cognitive pipeline being constructed and present some intermediate results.
Slides

Poster Session, 13:00-14:00, Wednesday 22 August 2018

Mark Wernsdorfer. A Phenomenologically Justifiable Simulation of Mental Modeling
Real-world agents need to learn how to react to their environment. To achieve this, it is crucial that they have a model of this environment that is adapted during interaction and although important aspects may be hidden. This paper presents a new type of model for partially observable environments that enables to represent hidden states but still can be generated and queried on-line. Agents can use such a model to predict the outcomes of their actions and to infer action policies. These policies turn out to be better than the optimal policy in a partially observable Markov decision process as it can be inferred, for example by Q- or SARSA-learning. The structure and generation of these models are both motivated by phenomenological considerations from semiotics and the philosophy of mind. They evolved from changes in the common reinforcement learning setting that simulate a more cognitively plausibility task. The performance of these models is compared to a baseline of Markov models for prediction and interaction in partially observable environments.

Anton Kolonin. Resource-constrained social evidence based cognitive model for empathy-driven artificial intelligence
Working model of social aspects of human and non-human intelligence is required for social embodiment of artificial general intelligence systems to explain, predict and manage behavioral patterns in multi-agent communities. For this purpose, we propose implementation of resource-constrained social evidence based model and discuss possible implications of its application.

Nadisha-Marie Aliman and Leon Kester. Hybrid Strategies Towards Safe “Self-Aware” Superintelligent Systems
Against the backdrop of increasing progresses in AI research paired with a rise of AI applications in decision-making processes, security-critical domains as well as in ethically relevant frames, a large-scale debate on possible safety measures encompassing corresponding long-term and short-term issues has emerged across different disciplines. One pertinent topic in this context which has been addressed by various AI Safety researchers is e.g. the AI alignment problem for which no final consensus has been achieved yet. In this paper, we present a multidisciplinary toolkit of AI Safety strategies combining considerations from AI and Systems Engineering as well as from Cognitive Science with a security mindset as often relevant in Cybersecurity. We elaborate on how AGI “Self-awareness” could complement different AI Safety measures in a framework extended by a jointly performed Human Enhancement procedure. Our analysis suggests that this hybrid framework could contribute to undertake the AI alignment problem from a new holistic perspective through security-building synergetic effects emerging thereof and could help to increase the odds of a possible safe future transition towards superintelligent systems.

Robert Wünsche. Adaptive Compressed Search
Program-search as induction and abduction is one of the key pillars of any sufficiently advanced AGI.
In this paper, we present a mechanism to search for programs given a specific bias.
This bias is flexible to some degree.
Another novel attribute of the mechanism is the use of compression that selects simple programs over complex ones.
The complexity of the program is changing all the time over the lifetime of the agent.

Andreas Hein and Hélène Condat. Can Machines Design? An Artificial General Intelligence Approach
Can machines design? Can they come up with creative solutions to problems and build tools and artifacts across a wide range of domains? Recent advances in the field of computational creativity and formal Artificial General Intelligence (AGI) provide frameworks for machines with the general ability to design. In this paper we propose to integrate a formal computational creativity framework into the Gödel machine framework. We call this machine a design Gödel machine. Such a machine could solve a variety of design problems by generating novel concepts. In addition, it could change the way these concepts are generated by modifying itself. The design Gödel machine is able to improve its initial design program, once it has proven that a modification would increase its return on the utility function. Finally, we sketch out a specific version of the design Gödel machine which specifically aims at the design of complex software and hardware systems. Future work could be the development of a more formal version of the Design Gödel machine and a potential implementation.

Patrick Hammer and Tony Lofthouse. Goal-directed Procedure Learning
A novel method of Goal-directed Procedure Learning is presented that overcomes some of the drawbacks of the traditional approaches to planning and reinforcement learning. The necessary principles for acquiring goal-dependent behaviors, and the motivations behind this approach are explained. A concrete implementation exists in a Non-Axiomatic Reasoning System, OpenNARS, although we believe the findings may be generally applicable to other AGI systems.
Slides

Rafik Hadfi. Solving Tree Problems with Category Theory
Artificial Intelligence (AI) has long pursued models, theories, and techniques to imbue machines with human-like general intelligence. Yet even the currently predominant data-driven approaches in AI seem to be lacking humans’ unique ability to solve wide ranges of problems. This situation begs the question of the existence of principles that underlie general problem-solving capabilities. We approach this question through the mathematical formulation of analogies across different problems and solutions. We focus in particular on problems that could be represented as tree-like structures. Most importantly, we adopt a category-theoretic approach in formalising tree problems as categories, and in proving the existence of equivalences across apparently unrelated problem domains. We prove the existence of a functor between the category of tree problems and the category of solutions. We also provide a weaker version of the functor by quantifying equivalences of problem categories using a metric on tree problems.

Fredrik Mäkeläinen, Hampus Torén and Claes Strannegård. Efficient Concept Formation in Large State Spaces
General autonomous agents must be able to operate in previously unseen worlds with large state spaces. To operate successfully there, the agents must maintain their own models of the world, based on sets of internal states that are several orders of magnitude smaller.
For adaptive agents, this internal state space cannot be fixed, but must adapt continuously to new situations. This in turn requires mechanisms for forming and preserving those internal states that are critical to successful decision-making, while removing others.
In this paper we compare three general algorithms for learning and decision-making: (i) standard Q-learning, (ii) local Q-learning, and (iii) local Q-learning with improved formation rules, based on a novel mechanism for forming internal states. In an experiment with a state space of size $2^{32}$ and high levels of noise, it was found that local Q-learning with improved formation rules outperforms local Q-learning, which in turn outperforms standard Q-learning.

Andrej Dameski. A Comprehensive Ethical Framework for AI Entities
The participation of AI in society is expected to increase significantly, and with that the scope, intensity and significance of morally-burdened effects produced or otherwise related to AI, and the possible future advent of AGI. There is a lack of a comprehensive ethical framework for AI and AGI, which can help manage moral scenarios in which artificial entities are participants. Therefore, I propose such a framework in this text, and suggest that it can enable artificial entities to make morally sound decision in complex moral scenarios.

Pei Wang and Patrick Hammer. Perception from an AGI Perspective
This paper argues that according to the discovery of cognitive science, in AGI systems perception should be considered as subjective, active, and unified with other processes. This treatment of perception is fundamentally different from the traditional approaches in computer vision, pattern recognition, machine learning, and the other related fields, where perception is taken to be objective, passive, and modular. The conceptual design of perception in the AGI system NARS is introduced, where the three features are realized. Some preliminary testing cases are used to show the features of this novel approach.
Slides