Russian Federation

The ability to explain decisions made by AI systems is highly sought after, especially in domains where human lives are at stake such as medicine or autonomous vehicles. While it is often possible to approximate the input-output relations of deep neural networks with a few human-understandable rules, the discovery of the double descent phenomena suggests that such approximations do not accurately capture the mechanism by which deep neural networks work. Double descent indicates that deep neural networks typically operate by smoothly interpolating between data points rather than by extracting a few high level rules. As a result, neural networks trained on complex real world data are inherently hard to interpret and prone to failure if asked to extrapolate. To show how we might be able to trust AI despite these problems we introduce the concept of self-explaining AI. Self-explaining AIs are capable of providing a human-understandable explanation of each decision along with confidence levels for both the decision and explanation. Some difficulties with this approach along with possible solutions are sketched. Finally, we argue it is important that deep learning based systems include a &quot;warning light&quot; based on techniques from applicability domain analysis to warn the user if a model is asked to extrapolate outside its training distribution.

AGI-20

Self-explaining AI as an alternative to interpretable AI

self-explaining ai

trustworthiness

double descent

explainability

interpretability

explainable ai

robustness

The purpose of the AAAI conference is to promote research in artificial intelligence (AI) and scientific exchange among AI researchers, practitioners, scientists, and engineers in affiliated disciplines. 

AAAI 2021

The ability to explain decisions made by AI systems is highly sought after, especially in domains where human lives are at stake such as medicine or autonomous vehicles. While it is often possible to approximate the input-output relations of deep neural networks with a few human-understandable rules, the discovery of the double descent phenomena suggests that such approximations do not accurately capture the mechanism by which deep neural networks work. Double descent indicates that deep neural networks typically operate by smoothly interpolating between data points rather than by extracting a few high level rules. As a result, neural networks trained on complex real world data are inherently hard to interpret and prone to failure if asked to extrapolate. To show how we might be able to trust AI despite these problems we introduce the concept of self-explaining AI. Self-explaining AIs are capable of providing a human-understandable explanation of each decision along with confidence levels for both the decision and explanation. Some difficulties with this approach along with possible solutions are sketched. Finally, we argue it is important that deep learning based systems include a "warning light" based on techniques from applicability domain analysis to warn the user if a model is asked to extrapolate outside its training distribution.

technical paper

The AGI conference series is the only major conference series devoted wholly and specifically to the creation of AI systems possessing general intelligence at the human level and ultimately beyond. By gathering together active researchers in the field, for presentation of results and discussion of ideas, we accelerate our progress toward our common goal.

Underline hosted the two main days of AGI 2020 on Wednesday June 24th and Thursday June 25th. Both days were **live-streamed** and you can read the full program bellow. All presentations are available for a **asynchronous viewing**. All presenters are grouped into panels. You can also hear guest talks by **Ben Goertzel** (SingularityNET), **Valery Tarasov** (Semenov Institute of Chemical Physics), **Tatiana Shavrina** (Sberbank) and **Joscha Bach** (MIT Media Lab). 

The AGI conference series is the only major conference series devoted wholly and specifically to the creation of AI systems possessing general intelligence at the human level and ultimately beyond. By gathering together active researchers in the field, for presentation of results and discussion of ideas, we accelerate our progress toward our common goal.

The paper is dedicated to the use of distributed hyperdimensional vec-tors to represent sensory information in the sign-based cognitive architecture, in which the image component of a sign is encoded by a causal matrix. The hyper-dimensional representation allows us to fix the precedent dimension of the causal matrix and accumulate information in it during the interaction of the system with the environment. Due to the high dimensionality of vectors, it is possible to re-duce the representation and reasoning on the entities related to them to simple operations on vectors. In this work we show how hyperdimensional representa-tions are embedded in an existing sign formalism and provide examples of scene encoding.

Hyperdimensional Representations in Semiotic Approach to AGI

We consider a task-oriented approach to AGI, when any cognitive problem, perhaps superior to human ability, has sense given a criterion of its solution. In the frame of this approach, we consider the task of purposeful behavior in a complex probabilistic environment, where behavior is organized through self-learning. For that purpose, we suggest cognitive architecture that relies on the theory of functional systems. The architecture is based on the main notions of this theory: goal, result, anticipation of the result. The logical structure of this theory was analyzed and used for the control system of purposeful behavior development. This control system contains a hierarchy of functional systems that organizes purposeful behavior. The control system was used for modeling agents to solve the foraging task.

Logical probabilistic biologically inspired cognitive architecture

This presentation describes ideas on how to develop engineering tools based on knowledge graphs to manage a systems complexity.

Approach for development of engineering tools based on knowledge graphs and context separation

"This position paper suggests the existence of a plurality of ""general-purpose''
AGI paradigms, each specific to a domain of experience. These paradigms are studied to answer the question of which AGI will be developed first. Finally, in order to make the case for AGI based on symbolic experience, preliminary results from Semiotic AI are discussed."

Experience-specific AGI Paradigms

"A possible practical engineering approach to creation of the general artificial
intelligence is considered. The choice of approach is based on modular hierarchical representation of knowledge, where each module uses its own methods of representation and knowledge processing. Work with knowledge is done by a hierarchical multi-agent system. The description of system’s individual elements and information about the current development state are given."

The use of engineering approach in creation of artificial general intelligence

Creating Strong AI means to develop artificial intelligence to the point where the machine's intellectual capability is in a way equal to a human's. Science is definitely one of the summits of human intelligence, the other being the art. Scientific research consists in creating hypotheses that are limited applicability models (methods) implying lossy information compression. In this article, we show that this paradigm is not unique to the science and is common to the most developed areas of human activities, like business and engineering. Thus, we argue, a Strong AI should possess a capability to build such models. Still, the known tests to confirm the human-level AI do not address this consideration. Based on the above we suggest a series of six tests of rising complexity to check if AI have achieved the human-level intelligence (Explanation, Problem-setting, Refutation, New phenomenon prediction, Business creation, Theory creation), five of which are new to the AGI literature.

How do you test the strength of AI?

"Autonomous knowledge transfer from a known task to a new
one requires discovering task similarities and knowledge generalization
without the help of a designer or teacher. How transfer mechanisms in
such learning may work is still an open question. Transfer of knowledge
makes most sense for learners for whom novelty is regular (other things
being equal), as in the physical world. When new information must be
unied with existing knowledge over time, a cumulative learning mech-
anism is required, increasing the breadth, depth, and accuracy of an
agent's knowledge over time, as experience accumulates. Here we address
the requirements for what we refer to as autonomous cumulative
transfer learning (ACTL) in novel task-environments, including implementation
and evaluation criteria, and how it relies on the process of
similarity and ampliative reasoning. While the analysis here is theoretical,
the fundamental principles of the cumulative learning mechanism in
our theory have been implemented and evaluated in a running system
described priorly. We present arguments for the theory from an empirical
as well as analytical viewpoint."

Autonomous Cumulative Transfer Learning

In recent works, the DSO Cognitive Architecture’s design is enhanced by incorporating the concept of the Global Workspace Theory(GWT). The theory proposes that consciousness is realised through the competition of massive, specialised, parallelised processes and thus parallelised, unsynchronised cognitive processes become sequential through such bottleneck. Due to the concurrent nature of DSO Cognitive Architecture, coordination of the different parallel processes through this competition mechanism can be difficult and if not handled properly, will create inconsistent results. In this work, we propose a preliminary framework to coordinate the different processes by process composition which borrows concepts from automated planning. Processes, its argument signature and its output are abstracted into higher level type abstractions which can be used to compose with other processes based on matching the output types to argument types. This is known as process composition and it represents a sketch of how different process can coordinate with one another. We combined this with the current design of the DSO Cognitive Architecture and illustrate an example in crowd anomaly detection

Towards Dynamic Process Composition in the DSO Cognitive Architecture

This paper discuses attentional control mechanism of several systems in context of Artificial General Intelligence. Attentional control mechanism of OpenNARS, an implementation of Non-Axiomatic Reasoning System for research purposes is being introduced with description of the related functions and demonstration examples. Paper also implicitly compares OpenNARS attentional mechanism with the one found in other Artificial General Intelligence systems.

An Attentional Control Mechanism for Reasoning and Learning

A pragmatic design for a general purpose reasoner incorporating the Non-Axiomatic Logic (NAL) and Non-Axiomatic Reasoning System (NARS) theory. The architecture and attentional control differ in many respects to the OpenNARS implementation. Key changes include; an event driven control process, separation of sensorimotor from semantic inference and a different handling of resource constraints.

Downloads

Next from AGI-20

Hyperdimensional Representations in Semiotic Approach to AGI

Similar lecture

Does Explainable Artificial Intelligence Improve Human Decision-Making?

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AGI-20

Hyperdimensional Representations in Semiotic Approach to AGI

Similar lecture

Does Explainable Artificial Intelligence Improve Human Decision-Making?

Downloads