AGI & Attention

1. Helgi Páll Helgason helgi@perseptio.com AGI & Attention AI researcher, Ph.D. candidate Center for Analysis and Design of Intelligent Agents, Reykjavik University Intellifest 2012

2. Schedule changes Today: Attention Friday November 2nd: Attention Tuesday November 6th: Autonomy Intellifest 2012

3. Today Importance of attention for AI Example of inspiration from human attention (cognitive psychology) Attention beyond the human level (for meta-cognitive purposes) Design of an attention mechanism Intellifest 2012

4. Attention & AI • To work on AI research, each researcher or team needs to have • A clear and explicit working definition of intelligence • A clear and explicit motivation Intellifest 2012

5. What do we mean by machine intelligence? Displaying human-like behavior? Solving computationally complex problems (in some unspecified amount of time)? Performing isolated tasks that have conventionally required humans? Adapting to a complex, dynamic environment with insufficient knowledge and resources? Intellifest 2012

6. What is our motvation? • Development and validation of psychological models? • Development and validation of neurological models? • Development and implementation of practical, flexible and versatile autonomous system? • How accurately do we want to replicate existing biological mechanisms? • To what degree are we biologically inspired? Intellifest 2012

7. Intelligence • Nature’s way of dealing with complexity under resource and time constraints • No real-world intelligence exists that does not address the passage of time head on Intellifest 2012

8. A working definition “Intelligence is the capacity of a system to adapt to its environment while operating with insufficient knowledge and resources.” • Pei Wang (Rigid flexibility: The Logic of Intelligence. Springer 2006) Intellifest 2012

9. A working definition Intelligence is a capability of information processing systems Intelligence is adaptation: The system’s solution of one problem is not only determined by the problem itself, but also prior experience Intellifest 2012

10. A working definition Insufficient knowledge -> system will usually not have the best solution at hand Insufficient resources -> system can not consider (process) every possibility nor store all information Intellifest 2012

11. Intelligence “If either time or computational resources are infinite, intelligence is irrelevant.” • Dr. Kristinn R. Thórisson Intellifest 2012

12. Attention • In the domain of intelligent systems, the management of system resources is typically called “attention” • Biological (Human) Attention: • Selective concentration on one aspect of the environment while ignoring others • Artificial Attention: • Resource management and control mechanism to assign limited system resources to processing of most relevant or important information Intellifest 2012

13. Attention Time constraints ATTENTION Abundant information Limited resources Intellifest 2012

14. Intelligence Time constraints INTELLIGENCE Abundant information Limited resources Intellifest 2012

15. What do we mean by AI? • “Narrow“ (classical) AI: • Systems explicitly designed to solve specific, reasonably well-defined problems • E.g. Deep Blue, Watson, etc. • Artificial General Intelligence (AGI) • Systems designed to autonomously learn novel tasks and adapt to changing environments Intellifest 2012

16. Attention and Narrow AI • When tasks and environments are pre-specified, we know in advance… • what information is relevant to system operation • how frequently the system has to sample information • how frequently the system has to act • the resource requirements of the system Intellifest 2012

17. Attention and Narrow AI • Substantial dynamic adaption to task not required • Information filtering can be pre-programmed if characteristics of relevant information known in advance • Resource management and processing hand-tuned for specific tasks and environments • Major reduction in complexity (compared to real-world tasks and environments) • End up with limited and closed models of the real-world Intellifest 2012

18. Attention and AGI • When tasks and environments are unknown, we do not know in advance… • what information is relevant to system operation • how frequently the system has to sample information • how frequently the system has to act • the resource requirements of the system Intellifest 2012

19. Attention and AGI • Must assume up-front: • Real world environmental complexity • All information is potentially important • Not just limited, but insufficient resources at all times • Dynamic tasks, environments and time constraints Intellifest 2012

20. Why attention? • “Narrow” AI • Substantial dynamic adaptation to task not required • Data filtering can be pre-programmed if characteristics of useful data known in advance • Lower than real world task complexity • Resource management and processing hand-tuned for specific scenarios →Attention not required (?) • AGI • Real world environmental complexity assumed up-front • Computational resources for the AI assumed to be insufficient at all times • Complexity calls for data filtering and intelligent resource allocation • Environments and tasks unknown at implementation time • Resource management must be adaptive → Demands strong focus on resource management and realtime processing Intellifest 2012

21. Resource management challenges for AGI systems • Real world is highly dynamic and complex, provides abundance of information. • System resources not only limited, but insufficient in light of amount of available information. • Range of time constraints (many of which are dictated by the environment) must be satisfied. • Unexpected events requiring immediate response may occur at any time. Intellifest 2012

22. SOAR AGI SUMMER SCHOOL 2012

23. Intellifest 2012

24. Human attention Intellifest 2012 “Everyone knows what attention is. It is the taking possession by the mind, in clear and vivid form, of one out of what seem several simultaneously possible objects or trains of thought. Focalization, concentration, of consciousness are of its essence. It implies withdrawal from some things in order to deal effectively with others, and is a condition which has a real opposite in the confused, dazed, scatterbrained state which in French is called distraction, and Zerstreutheit in German.” - William James, 1890

25. Attention in Cognitive Psychology • Modern attention research started with the “cocktail party effect” (Colin Cherry, 1953) • Number of attention models have been proposed, most belonging to two classes: • Early selection: Selection of information occurs early in sensory pipeline based on shallow, primitive processing with no or limited analysis of meaning • Late selection: Selection of information based on deep analysis of meaning, occurs late in sensory pipeline • Many early selection models contradicted by observed human behavior (e.g. cocktail party scenario) Intellifest 2012

26. Broadbent Filter Model (Early Selection) Intellifest 2012

27. Treisman Attenuation Model(Early selection) Intellifest 2012

28. Early selection Intellifest 2012

29. Deutsch-Norman Model(Late selection) Intellifest 2012

30. Knudsen Attention Framework(Late selection) Knudsen EI: Fundamental components of attention. Annu Rev Neurosci 2007, 30:57-78. Intellifest 2012

31. Attention in Cognitive Psychology • Two types of attention: • Top-down • Deliberate, goal-driven, targets information related to the tasks being performed • Bottom-up • Reactive, targets unexpected but potentially important information Intellifest 2012

32. Inspiration from cognitive psychology Intellifest 2012 • Information competes for limited system resources • Allows processing decisions to be made as late as possible, when resource availability is known • Early selection may be a problematic paradigm • Ignoring information without analysis of meaning introduces operational risk • Two simultaneously active functions of attention can allow systems to perform tasks while remaining reactive to unexpected events. • Top-down attention may be controlled by active goals and predictions of the system to catch information related to current tasks. • Bottom-up attention can be controlled by novelty and unexpectedness of incoming information

33. Case study • General Attention Mechanisms for Cognitive Architectures • My Ph.D. project  Intellifest 2012

34. Research goals • Design a general attention mechanism intended for implementation in AGI systems (cognitive architectures) • In progress: Implementation and evaluation of resulting attention mechanism in state-of-the-art cognitive architectures • While work is biologically inspired at high-level, replication of any existing attention mechanism is not a goal Intellifest 2012

35. Design requirements • Complete • Targets all operational information (internal + external) • Top-down + Bottom-up • General • No limiting assumptions about tasks, environments or modalities • Uniform • Data from all modalities treated identically (at some level of processing) • Adaptive • Learns from experience Intellifest 2012

36. General approach to attention • Modality neutral • All modalities treated identically, at some level of processing • Including proprioception (internal modalities, self-sensing) • Architecture-independent Intellifest 2012

37. Related work • Attention functionality implemented in handful of AGI systems • E.g. NARS, LIDA, CLARION • Limitations: • Data-filtering only (control issues ignored) • External information only (internal states ignored) • Realtime processing not addressed Intellifest 2012

38. Approach • Amount of available information constantly assumed to exceed system processing capacity • Limited system resources must be focused on most relevant or important information • Requires capability to determine degree of information importance, based on: • Current operating context • Time constraints • Resource availability Intellifest 2012

39. Issues for AGI attention Intellifest 2012 • Is it sufficient to only evaluate relevance for data? • What about the relevance of system processes? • Just another “box” in the sensory pipeline? • Or something more pervasive? • Can we retrofit existing AGI architectures with attention? • Can we extend attention capabilities in useful ways for AGI systems?

40. Attention: Functional requirements Intellifest 2012 • Quantify current relevance of data • Data relevance: • Goal-related • Novelty / Unexpectedness • Quantify current relevance of processes • Process relevance: • Operational experience • Available data

41. Methodology Intellifest 2012 • Constructivist AI • “From Constructionist to Constructivist AI”, Thórisson 2009, BICA proceedings • Systems manage own growth • From manually constructed initial state (bootstrap/seed) • Methodology for building flexible AGI systems capable of autonomous self-reconfiguration at the architecture level

42. Attention for self-reconfiguration Intellifest 2012 • Internally, the system and its operation can be viewed as a dynamic and complex environment • Similar to external task environment • Meta-cognitive functions responsible for system growth must also process information selectively • Resources remain limited • Applying the same attention mechanism to external and internal environments may produce AI systems capable of performing tasks and improving own performance while being subject to realtime constraints and resources limitations.

43. Meta-cognitive operation Intellifest 2012 • Introspection, self-growth, self-improvement • As the sum of internal system activity is a vast stream of information… • applying attention to the internal environment can lend significant support for meta-cognitive operation by… • helping determine important information and processes for meta-cognitive functions… • in the same way it supports task performance in the external environment

44. Attention mechanism design • While design is architecture-independent, some requirements are necessary Intellifest 2012

45. Architectural requirements • Data-driven • All processing is triggered by the occurrence of data • Eliminates the need for fixed control loops, allowing for operation at multiple time scales and greater flexibility • Fine-grained • Data and processing units are small but numerous • Reasoning about small, simple components and their effects is significantly more tractable than for larger, more complex components Intellifest 2012

46. Architectural requirements • Predictive capabilities • Capacity to generate predictions and expectations • Necessary control data for top-down attention in addition to goals • Unified sensory pipeline • Data given identical treatment regardless of origin (external, internal) Intellifest 2012

47. Environment (Real world) Unified sensory pipeline: External (environmental) and internal data handled identically at architecturelevel Sampled data Data items Actuation devices Sensory devices Commands New data Processes Data-driven: Processes are activated only when paired with compatible data Fine-grained: Data and process objects are small and numerous

48. Predictive capabilities: Predictions are necessary control information for top-down attention Environment (Real world) Goals / Predictions Derived Attentional patterns Matching Sampled data Data items Top-down Actuation devices Sensory devices Data biasing Commands Processes Data and processess have priority values that are assigned by biasing.

49. Environment (Real world) Goals / Predictions Derived Attentional patterns Matching Sampled data Data items Top-down Actuation devices Sensory devices Data biasing Commands Bottom-up Bottom-up attentional processess Evaluation Processes

50. Environment (Real world) Goals / Predictions Derived Attentional patterns Matching Sampled data Data items Top-down Actuation devices Sensory devices Data biasing Commands Bottom-up Bottom-up attentional processess Evaluation Processes Data -> Process mapping Process biasing