LIDA, (Learning IDA), adds various mode of human-like learning to the
IDA architecture (see IDA description below under Past Projects), including perceptual, episodic,
procedural and attentional learning.
LIDA Model: Experiment Replication
Touted as a cognitive model, LIDA must constitute a valid scientific model of how minds work. A major way to test this is through the replication of data from existing psychological experiments. This project involves building software agents based on the LIDA model to perform experiments, and comparing the results to those from human or animal subjects. Compatible results constitute evidence for the LIDA model. Incompatible results necessitate some updating of the model or its parameters. We've learned something. Science is working. In addition to testing the LIDA model, such work thus contributes to the identification of the internal parameters for the model including determining appropriate values. Until now, we've replicated two such experiments bearing on the timing of the action-perception cycle, one modeling the attentional blink phenomenon, and another exploring executive attention. The first three are published.
LIDA Model: MAX
A real world application of the LIDA model, MAX will be a LIDA-based software agent situated in hybrid real and digital environment composed of healthcare providers, clinical data regarding the providers' patients, and the diagnostic possibilities associated with that data. MAX will use human-style reasoning, and will communicate with the providers in natural language. It will generate diagnostic hypotheses by analyzing clinical data, and utilize clinical knowledge to investigate those hypotheses. Furthermore, MAX will be able to learn by communicating with humans, by consulting medical literature, and from its own experience.
LIDA Model: Software Framework
Intelligent software agents aiming toward general intelligence are complex systems and, as such, are difficult and time consuming to implement and to customize. A software framework is a reusable implementation of the skeleton of a software system, capturing its generic functionality. By significantly reducing the amount of effort necessary to develop customized applications, frameworks are becoming increasingly attractive for the implementation of intelligent software agents.
The LIDA software framework is a generic and customizable computational implementation of the LIDA model, programmed in Java. It allows for the relatively rapid development of LIDA controlled software agents for specific problem domains, using a declarative specification of the agent architecture by an XML file. Small pieces of the system, and even the internal implementation of whole modules, can be customized. Since biological minds operate in parallel, the Framework provides for multithreading support. Its implementation adheres to several well-established design principles, and best programming practices. The LIDA Framework is available online for research use, and has been used to produce, so far, four LIDA based software agents, resulting in two publications.
LIDA Model: From Perception to Cognition
Intelligent agents operating in complex "real-world" environments must generate their own understanding of their current situation from their sensory primitives. In some cases, for example in the visual modality in humans, the process of meaning creation and understanding is exceedingly complex. Cognitive architectures have historically eschewed considerations of the details of perception, and of reconciling high-level representations, useful for reasoning, logic, and planning, with high-dimensional, transient, noisy sensory stimuli. In this project we take a biologically-inspired view, within a broad cognitive model, LIDA, of how the abstract, invariant, and conceptual aspects of detailed, low-level sensory representation are recognized and associated. Such work fleshes out the details of LIDA's Sensory Memory, Perceptual Associative Memory, and conscious learning. Also covered is the critical relation of attentional and learning processes with this perceptual process.
LIDA Model: Vector LIDA using Extended Integer SDM
Sparse distributed memory (SDM) is a mathematical model of human long-term memory based on large binary vectors. It is distributed, auto-associative, content addressable, and noise robust. Moreover, it exhibits one-shot learning, is resilient to damage, and its contents degrade gracefully when the memory fills up. Its interesting psychological characteristics include interference, knowing when it does not know, and the tip of the tongue effect. SDM's structure is ideal for parallel processing or hardware implementation. All this makes it an attractive option for modeling memory modules in cognitive architectures and other AI applications.
Extended SDM uses a novel mechanism to add hetero-associativity to SDM, making it particularly effective for sequence learning. Integer SDM extends the memory to accept integer vectors, while retaining the benefits of the model. Built upon Extended Integer SDM, Modular Composite Representation (MCR), allows the representation of complex structures using single vectors, making them attractive for modeling internal data representations in cognitive architectures.
These technologies constitute the heart of the new Vector LIDA project that is implementing the LIDA architecture using MCR vectors for data representation. Vector LIDA should produce a more realistic and biologically plausible model, better integration with low-level perceptual processing, better scalability, and easier learning mechanisms.
In the realm of cognitive robotics, LIDA-AV aims to control an
autonomous vehicle with the LIDA technology.
Intelligent Distribution Agent addresses the Navy's problem of job
distribution using the Conscious Agent
Framework. IDA is a very complex
agent that perceives e-mails from sailors, deliberates on the
right jobs for the sailor and
negotiates with the sailor in the context of sailor's preferences and Navy's policies. This project is
funded by ONR (Office of Naval Research) and NPRST (Naval Personnel Research, Studies, and Technology).
is the first attempt to design
and implement an intelligent agent under
the framework of Bernard Baars' Global Workspace
Theory. She is a
successor to Virtual Mattie and performs the same function in the
same environment as
Mattie, to actively gather information from humans, compose announcements of
next week's seminars, and mail them each week to a
list that she keeps updated, all
without the supervision of a human.
Maes' behavior net architecture and
Hofstadter and Mitchell’s
Copycat architecture and significantly extends
"Living'' in a UNIX environment, VMattie's communication with
humans is entirely via email with no
agreed upon protocol.