In this paper, we build from the preliminary work in [1] on using deep learning architectures to control inferential glut in a logical reasoning agent. Inferential glut is the well-known problem that symbolic reasoners encounter combinatorial explosions in the numbers of formulae they have to process. In [1] we proposed a neural architecture that employed a simple deep-reinforcement-learning network to prioritize potential inferences and control glut. This architecture was able to successfully control pre-inferential glut (inferential glut arising from inefficiencies in the reasoner’s algorithm) and a limited form of inferential glut (that didn’t use state). Work on this architecture continues; in this work we explore the use of Decision Transformers [2] as a replacement for the original deep reinforcement learning model. The decision transformer models reinforcement learning as a sequence modelling problem – specifically it employs transformers to generate the action likely to lead to a given return given the trajectory of (states, actions, returns) seen so far. In [2], the authors showed that using a pre-trained and fine-tuned transformer (such a GPT2) can lead to near state-of-the-art on offline reinforcement learning tasks. Since glut control as envisioned is largely a matter of selectively attending to the most promising avenues of reasoning, the attention-based mechanisms which underwrite transformers seem a natural fit for our application. In this paper we will explore the effectiveness of using a fine-tuned GPT3 model [3] to predict salient actions as a proof-of-concept.

In this paper, we address the problem of inferential glut – specifically the problem that symbolic reasoners tend to produce vastly more deductive information than they can tractably process. As a concrete example, let us imagine an agent which is searching for a specific book. Entering a study, it encounters, among other things, a bookshelf, a lamp, and a table. If the agent has a knowledge base with a large number of facts about lamps and tables, it may spend an inordinate amount of resources reasoning about lamps and tables, and such irrelevant reasoning can prevent it from drawing conclusions that are helpful for finding the book. In fact, such swamping can also arise immediately and devastatingly even without distractor entities, as discussed below. Of course, the agent could also have special information guiding the reasoning in ways appropriate to the specific task (e.g., books tend to be on shelves). But how is this special information obtained? Here we address this as a learning issue and report on preliminary results from our approach. As an example, consider a round library with one continuous circular shelf along its circumference, where each book B has exactly one "left" neighbor left(B) and another “right” neighbor right(B), where these function symbols simply refer to those two neighbors of B. This might be encoded by the following rules K for some book a:

Estimating the correspondences between pixels in sequences of images is a critical first step for a myriad of tasks including vision-aided navigation (e.g., visual odometry (VO), visual-inertial odometry (VIO), and visual simultaneous localization and mapping (VSLAM)) and anomaly detection. We introduce a new unsupervised deep neural network architecture called the Visual Inertial Flow (VIFlow) network and demonstrate image correspondence and optical flow estimation by an unsupervised multi-hypothesis deep neural network receiving grayscale imagery and extra-visual inertial measurements. VIFlow learns to combine heterogeneous sensor streams and sample from an unknown, un-parametrized noise distribution to generate several (4 or 8 in this work) probable hypotheses on the pixel-level correspondence mappings between a source image and a target image . We quantitatively benchmark VIFlow against several leading vision-only dense correspondence and flow methods and show a substantial decrease in runtime and increase in efficiency compared to all methods with similar performance to state-of-the-art (SOA) dense correspondence matching approaches. We also present qualitative results showing how VIFlow can be used for detecting anomalous independent motion.

Aimed at improving size, weight, and power (SWaP)-constrained robotic vision-aided state estimation, we describe our unsupervised, deep convolutional-deconvolutional sensor fusion network, Multi-Hypothesis DeepEfference (MHDE). MHDE learns to intelligently combine noisy heterogeneous sensor data to predict several probable hypotheses for the dense, pixel-level correspondence between a source image and an unseen target image. We show how our multi-hypothesis formulation provides increased robustness against dynamic, heteroscedastic sensor and motion noise by computing hypothesis image mappings and predictions at 76–357 Hz depending on the number of hypotheses being generated. MHDE fuses noisy, heterogeneous sensory inputs using two parallel, inter-connected architectural pathways and n (1–20 in this work) multi-hypothesis generating sub-pathways to produce n global correspondence estimates between a source and a target image. We evaluated MHDE on the KITTI Odometry dataset and benchmarked it against the vision-only DeepMatching and Deformable Spatial Pyramids algorithms and were able to demonstrate a significant runtime decrease and a performance increase compared to the next-best performing method.

Due both to the speed and quality of their sensors and restrictive on-board computational capabilities, current state-of-the-art (SOA) size, weight, and power (SWaP) constrained autonomous robotic systems are limited in their abilities to sample, fuse, and analyze sensory data for state estimation. Aimed at improving SWaP-constrained robotic state estimation, we present Multi-Hypothesis DeepEfference (MHDE) - an unsupervised, deep convolutional-deconvolutional sensor fusion network that learns to intelligently combine noisy heterogeneous sensor data to predict several probable hypotheses for the dense, pixel-level correspondence between a source image and an unseen target image. This new multi-hypothesis formulation of our previous architecture, DeepEfference [1], has been augmented to handle dynamic heteroscedastic sensor and motion noise and computes hypothesis image mappings and predictions at 150-400 Hz depending on the number of hypotheses being generated. MHDE fuses noisy, heterogeneous sensory inputs using two parallel architectural pathways and n (1, 2, 4, or 8 in this work) multi-hypothesis generation subpathways to generate n pixel-level predictions and correspondences between source and target images. We evaluated MHDE on the KITTI Odometry dataset [2] and benchmarked it against DeepEfference [1] and DeepMatching [3] by mean pixel error and runtime. MHDE with 8 hypotheses outperformed DeepEfference in root mean squared (RMSE) pixel error by 103% in the maximum heteroscedastic noise condition and by 18% in the noise-free condition. MHDE with 8 hypotheses was over 5, 000% faster than DeepMatching with only a 3% increase in RMSE.

As the human eyeball saccades across the visual scene, humans maintain egocentric visual positional constancy despite retinal motion identical to an egocentric shift of the scene. Characterizing the underlying biological computations enabling visual constancy can inform methods of robotic localization by serving as a model for intelligently integrating complimentary, heterogeneous information. Here we present DeepEfference, a bio-inspired, unsupervised, deep sensorimotor network that learns to predict the sensory consequences of self-generated actions. DeepEfference computes dense image correspondences [1] at over 500 Hz and uses only a single monocular grayscale image and a low-dimensional extra-modal motion estimate as data inputs. Designed for robotic applications, DeepEfference employs multi-level fusion via two parallel pathways to learn dense, pixel-level predictions and correspondences between source and target images. We present quantitative and qualitative results from the SceneNet RGBD [2] and KITTI Odometry [3] datasets and demonstrate an approximate runtime decrease of over 20,000% with only a 12% increase in mean pixel matching error compared to DeepMatching [4] on KITTI Odometry.

We discuss Perry’s notion of the essential indexical and the requirement that robots interacting with humans (and other robots) be able to reason about themselves in a grounded way. We describe an approach based on grounding symbols via an analogue of the neural mechanism of efference copy and approaching symbolic reasoning via active logic – a situated framework for logical and temporal reasoning. The Essential Indexical Robots that interact with humans should be able to make sense of indexical terms (such as “I” and “now”). For example, a robotic helper should ideally be able to interpret human utterances such as: I am getting hungry. I don’t want to eat now, but would like you to get me a snack from the kitchen in 10 minutes. Please make sure to check that you didn’t spill anything after you’re done, and make sure to clean yourself up too if needed. Understanding such a sentence will require, along with much else, an ability to use indexical notions of “I”, “you”, “now” and “in 10 minutes.” A brief digression into philosophy will demonstrate that this is a more difficult problem than it first appears to be. See (Anderson and Perlis 2005) for a fuller account with explicit ties to robotics. The meat of our digression comes from a seminal 1979 essay by Stanford philosopher John Perry. In it, he describes the peculiar experience of following a trail of sugar around a grocery store. Perry spent some time pushing his cart up the aisle on one side of a tall counter, then down the other, noting all the while that the trail grew thicker and thicker. Eventually the truth dawned on him: as he put it, "I was the shopper I was trying to catch." Perry spent the rest of his seminal paper trying precisely to characterize the changes in his beliefs that precipitated the change in his Copyright © 2016, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. behavior -which was, of course, that he stopped looking for the shopper with the torn sugar sack. One should like to say that Perry came to believe that “I am the shopper making the mess.” Things turn out not to be so simple. What Perry finds is that in a traditional Fregean framework of de dicto belief -where belief is characterized as a particular relation between a subject and a proposition -there is no room for indexicals like “I” or “now.” For if we accept Perry's construal of the traditional "doctrine of propositions," then we agree that for a proposition S to equal proposition S' is precisely for S and S' to have not only the same truth-value, nor merely the same truth-conditions -that is, merely share the same reference -but that they have also the same intension, or sense. As Perry explains, “Atlanta is the capital of Georgia” and “Atlanta is the capital of the largest state East of the Mississippi” are not the same propositions -though they are certainly true or false under the same conditions -for I can clearly believe one but not the other without inconsistency, depending on what I believe about Georgia1. But then it cannot be that Perry came to believe “I am the shopper with the torn sack,” because that sentence does not even identify a proposition. It is not true or false absolutely. It is ambiguous, depending on who says it. Nor does it help, as Perry points out, to argue that indexicals like “I” or “now” or “this” are communicative shortcuts standing in for some propositions α, β, or γ. For if we claim that all Perry came to believe was, in fact, "α is making a mess," where α unambiguously picks out Perry2, we have not yet explained his subsequent change in behavior -for we also believe this proposition, and yet haven't stopped reading to check our shopping carts. No; Perry must also have believed "I am α." Perry goes on to describe in detail the way in which indexicals like “I” and “now” trouble the waters of tradition1 Hilary Putnam argued against fixing reference in general based solely on truth-conditional criteria. See (Lakoff 1990) for a good exposition. 2 Perhaps, Perry suggests, α might be “the only bearded philosopher in a Safeway store West of the Mississippi.” The 2016 AAAI Fall Symposium Series: Artificial Intelligence for Human-Robot Interaction Technical Report FS-16-01

In a previous paper, we discussed the desirability of a formal model of a self that emerged when a process became immediately aware of its own processing. In this paper we discuss some of the features that go into a realization of this pro-cessual self in active logic. A description is given of some specific extensions to active logic that would move toward allowing for a computational processual self.

A fully autonomous agent recognizes new problems, explains what causes such problems, and generates its own goals to solve these problems. Our approach to this goal-driven model of autonomy uses a methodology called the Note-Assess-Guide procedure. It instantiates a monitoring process in which an agent notes an anomaly in the world, assesses the nature and cause of that anomaly, and guides appropriate modifications to behavior. This report describes a novel approach to the note phase of that procedure. Adistance, a sliding-window statistical distance metric, is applied to numerical vector representations of intermediate states from plans generated for two symbolic domains. Using these representations, the metric is able to detect anomalous world states caused by restricting the actions available to the planner.

Cognitive Science at present is highly fragmented across multiple disciplines and even within those disciplines. It not only lacks an underlying explanatory framework but even a unifying theme or concept other than the not-very-helpful ones of mind and behavior. In this paper, we survey some ways in which a robust notion of self may be "waiting" to be fleshed out across the allied cognitive sciences in what we think may become a powerfully unifying and explanatory role.

The paper discusses the benefits of metacognitive guidance for a natural language dialog agent. These capabilities may be included directly in the agent or through a general purpose external module. We report on the specific case of handling pause time in dialog, using a metacognitive loop within the agent, and discuss future experiments implementing guidance for this example also using the general module. Keywords-metacognition; dialog management

The MIDCA cognitive architecture has at its core a process sequence consisting of three phases. They are to note an anomaly, to assess what causes the anomaly, and then to guide a response to the anomaly. Here we present a novel approach to the first phase and discuss implications for the second. Our method detects a shift in streams of symbolic data that signal an anomaly and trigger in-depth understanding of the input. The approach uses a metric called the A-distance, normally used to detect shifts in distributions underlying numeric data. Instead, using a novel plan representation, we apply this metric to streams of changing predicates in various environments. Empirical results show that over a range of circumstances we are able to detect changes in the underlying domain. We suggest that these results apply to metacognitive as well as to cognitive processing.

To understand the language we use, we sometimes must turn language on itself, and we do this through an understanding of the use-mention distinction. In particular, we are able to recognize mentioned language: that is, tokens (e.g., words, phrases, sentences, letters, symbols, sounds) produced to draw attention to linguistic properties that they possess. Evidence suggests that humans frequently employ the use-mention distinction, and we would be severely handicapped without it; mentioned language frequently occurs for the introduction of new words, attribution of statements, explanation of meaning, and assignment of names. Moreover, just as we benefit from mutual recognition of the use-mention distinction, the potential exists for us to benefit from language technologies that recognize it as well. With a better understanding of the use-mention distinction, applications can be built to extract valuable information from mentioned language, leading to better language learning materials, precise dictionary building tools, and highly adaptive computer dialogue systems. This dissertation presents the first computational study of how the use-mention distinction occurs in natural language, with a focus on occurrences of mentioned language. Three specific contributions are made. The first is a framework for identifying and analyzing instances of mentioned language, in an effort to reconcile elements of previous theoretical work for practical use. Definitions for mentioned language, metalanguage, and quotation have been formulated, and a procedural rubric has been constructed for labeling instances of mentioned language. The second is a sequence of three labeled corpora of mentioned language, containing delineated instances of the phenomenon. The corpora illustrate the variety of mentioned language, and they enable analysis of how the phenomenon relates to sentence structure. Using these corpora, inter-annotator agreement studies have quantified the concurrence of human readers in labeling the phenomenon. The third contribution is a method for identifying common forms of mentioned language in text, using patterns in metalanguage and sentence structure. Although the full breadth of the phenomenon is likely to elude computational tools for the foreseeable future, some specific, common rules for detecting and delineating mentioned language have been shown to perform well.

Researchers have studied problems in metacognition both in computers and in humans. In response some have implemented models of cognition and metacognitive activity in various architectures to test and better define specific theories of metacognition. However, current theories and implementations suffer from numerous problems and lack of detail. Here we illustrate the problems with two different computational approaches. The Meta-Cognitive Loop and Meta-AQUA both examine the metacognitive reasoning involved in monitoring and reasoning about failures of expectations, and they both learn from such experiences. But neither system presents a full accounting of the variety of known metacognitive phenomena, and, as far as we know, no extant system does. The problem is that no existing cognitive architecture directly addresses metacognition. Instead, current architectures were initially developed to study more narrow cognitive functions and only later were they modified to include higher level attributes. We claim that the solution is to develop a metacognitive architecture outright, and we begin to outline the structure that such a foundation might have.

This dissertation studies the problem of facilitating semantic search across disparate ontologies that are developed by different organizations. There is tremendous potential in enabling users to search independent ontologies and discover knowledge in a serendipitous fashion, i.e., often completely unintended by the developers of the ontologies. The main difficulty with such search is that users generally do not have any control over the naming conventions and content of the ontologies. Thus terms must be appropriately mapped across ontologies based on their meaning. The meaning-based search of data is referred to as semantic search, and its facilitation (aka semantic interoperability) then requires mapping between ontologies.In relational databases, searching across organizational boundaries currently involves the difficult task of setting up a rigid information integration system. Linked Data representations more flexibly tackle the problem of searching across organizational boundaries on the Web. However, there exists no consensus on how ontology mapping should be performed for this scenario, and the problem is open. We lay out the foundations of semantic search on the Web of Data by comparing it to keyword search in the relational model and by providing effective mechanisms to facilitate data interoperability across organizational boundaries.We identify two sharply distinct goals for ontology mapping based on realworld use cases. These goals are: (i) ontology development, and (ii) facilitating interoperability. We systematically analyze these goals, side-by-side, and contrast them. Our analysis demonstrates the implications of the goals on how to perform ontology mapping and how to represent the mappings.We rigorously compare facilitating interoperability between ontologies to information integration in databases. Based on the comparison, class matching is emphasized as a critical part of facilitating interoperability. For class matching, various class similarity metrics are formalized and an algorithm that utilizes these metrics is designed. We also experimentally evaluate the effectiveness of the class similarity metrics on real-world ontologies. In order to encode the correspondences between ontologies for interoperability, we develop a novel W3C-compliant representation, named skeleton.

This paper argues for a "commonsense core" hypothesis, with emphasis on the issue of consistency in agent knowledge bases. This is part of a long-term research program, in which the hypothesis itself is being gradually refined, in light of various sorts of evidence. The gist is that a commonsense reasoning agent that would otherwise become incapacitated in the presence of inconsistent data may – by means of a modest additional error-handling “core” component – carry out more effective real-time reasoning, and also that there may be cases of interest in which the “core” is more usefully integrated into the knowledge base itself.

What commonsense knowledge do intelligent systems need, in order to recover from failures or deal with unexpected situations? It is impractical to represent predetermined solutions to deal with every unanticipated situation or provide predetermined fixes for all the different ways in which systems may fail. We contend that intelligent systems require only a finite set of anomaly-handling strategies to muddle through anomalous situations. We describe a generalized metacognition module that implements such a set of anomaly-handling strategies and that in principle can be attached to any host system to improve the robustness of that system. Several implemented studies are reported, that support our contention.

The process of rationally revising beliefs in the light of new information is a topic of great importance and long-standing interest in artificial intelligence. Moreover, significant progress has been made in understanding the philosophical, logical, and computational foundations of belief revision. However, very little research has been reported with respect to the revision of other mental states, most notably propositional attitudes such as desires and intentions. In this paper, we present a first attempt to formulate a general framework for understanding the revision of mental states. We develop an abstract belief-desire-intention model of agents, and introduce a notion of rationality for this model. We then present a series of formal postulates characterizing the processes of adding beliefs, desires, and intentions, updating costs and values, and removing beliefs, desires, and intentions. We also investigate the computational complexity of several problems involving the abstract model and comment on algorithms for revision.

Informational conflicts in operational policies cause agents to run into situations where responding based on the rules in one policy violates the same or another policy. Static checking of these conflicts is infeasible and impractical in a dynamic environment. This paper discusses a practical approach to handling policy conflicts in real-time domains within the context of a hierarchical military command and control simulated system that consists of a central command, squad leaders and squad members. All the entities in the domain function according to preset communication and action protocols in order to perform successful missions. Each entity in the domain is equipped with an instance of a metacognitive component to provide on-board/on-time analysis of actions and recommendations during the operation of the system. The metacognitive component is the Metacognitive Loop (MCL) which is a general purpose anomaly processor designed to function as a cross-domain plugin system. It continuously monitors expectations and notices when they are violated, assesses the cause of the violation and guides the host system to an appropriate response. MCL makes use of three ontologies--indications, failures and responses--to perform the notice, assess and guide phases when a conflict occurs. Conflicts in the set of rules (within a policy or between policies) manifest as expectation violations in the real world. These expectation violations trigger nodes in the indication ontology which, in turn, activate associated nodes in the failure ontology. The responding failure nodes then activate the appropriate nodes in the response ontology. Depending on which response node gets activated, the actual response may vary from ignoring the conflict to prioritizing, modifying or deleting one or more conflicting rules.

As automated systems become more complex, their propen- sity to fail in unexpected ways increases. As humans, we often notice their failures with the same ease that we recog- nize our own plans going awry. Yet the systems themselves are frequently oblivious that the function they are designed to perform is no longer being performed. This is because humans have explicit expectations – about both the system’s behavior and our own behaviors – that allow us to notice an unexpected event. In this paper, we propose a way for AI systems to generate expectations about their own behavior, monitor them, and attempt to diagnose the underlying fail- ures that cause them. Once a cause has been hypothesized, attempts at recovery can be made. The process is naturally meta-cognitive in that the system must reason about its own cognitive processes to arrive at an accurate and useful re- sponse. We present an architecture called the Meta-Cognitive Loop (MCL), which attempts to tackle robustness in cognitive systems in a domain general way, as a plug-in component to decrease the brittleness of AI systems.