Overview

Building an A.I. model of the Human brain

What types of processes probably exist in physical brain regions of humans?

Emotion Machine (Model-6) reflective problem solving example

The Neural Model of Mind (or NeuralMoM) project is the intersection of the fields of artificial intelligence with the most advanced computational models of mind and neuroscience with the most advanced computational models of the brain. The goal of the project is to use artificial intelligence models of problem solving, such as the 6-Layer Emotion Machine Model developed by Marvin Minsky, in order to guide our self-reflective and self-control understanding of the computational aspects of goal-oriented thought processing.

Combining Artificial Intelligence and Neuroscience

People, objects, goals and plans are common aspects of most current AI models of mind. While the field of neuroscience is currently using machine learning techniques (such as Hidden Markov, Naive Bayes, K-Nearest Neighbor, Artificial Neural Networks, and Linear Filters), these techniques can only learn very simple classes of computational thought processes. Most of these techniques focus on learning artificial reactive memories, which are basically memories that follow a specific progression in time without reporting errors or successes to higher level cognitive systems, which would allow for modular debugging, compiling, and execution of thought processes.

Functional Neural Patterns could map to Artificial Intelligence Procedures

We propose that in order to find these more complex types of thought processes, we are looking for correlations between artificially intelligent software implementations of these thought processes and biological causal (functional) relationships between active neuronal structures.

One approach to finding causal (functional) relationships between active neuronal structures is by using active inhibition and excitation of neural structures noninvasively through transcranial magnetic stimulation (TMS). One lightweight non-invasive procedure we are implementing for reading neural activity is near-infrared spectroscopy (NIRS), which can be calibrated by fMRI data. Many approaches using electromagnetic control of large numbers of single cells, such as the novel high-bandwidth I/O techniques developed by the Boyden Lab. While all of these technologies are very preliminary, they have all shown great effectiveness individually.

Common Sense Self-Reflection

We hope that using a Common Sense model of mind based on the computational primitives based on the Emotion Machine model developed by Marvin Minsky will allow us to make people able to easy self-reflect and control their own mental mental states directly without knowing anything complex about neuroscience or their physiological brains at all. One goal is to allow to user to come to the system with their own model of mind and use this to interact with the NeuralMoM, which will inherently be a learning system for Models of Mind and realtime neuroscience data.

Mental representations for experiments in mapping neuroscience to A.I.

Mapping features of reflective computation to natural features

Example: Simple Computational Model Mendel's Model of Genetics Run-time causal reflective computation is a field of computer science that allows processes to be watched by other processes as they are running. Understanding natural processes as computational models has proven to be a useful way of seeing and simulating the world around us. If the computational model is simple enough, such as Mendel's binary model of genetic inheritance, it can be simulated within an intelligent human mind, such as Mendel's mind. However, when the computational processes become complex, such as models of world economies or human minds, they become impossible for humans to mentally simulate without computers. Measuring features of the natural process of cognition as evidenced by the human brain have become more numerous recently; these include: fMRI, EEG, MEG, PET, fNIRS, and others. In addition, secondary external natural features include: EKG, EMG, GSR, and others. The ability to reflect on causal dependency traces of a computational process allows two things:      1. Begin mapping natural features to and from computational features.      2. Begin designing novel causal reflection models of cognition and learning. While natural cognitive features are abundant, providing a wealth of natural data, useful computational features have been more illusive. Examples of the most basic computational features include: (1) memory creation, (2) memory read, and (3) memory write. Tracing all causal relationships between these basic features allows tracing the context of all other programmer-defined semantic abstractions. All of these computational features create an intricate trace network of dependencies, automatically traceable and shared by many parallel threads of execution. We are experimenting with a programming language [see: Funk2 Project] that allows causal tracing to occur modularly to dynamically chosen parts of large consumer-scale software projects. The resulting causal dependency trace networks can be processed by critically causal reflective threads. Discovering more useful types of causal reflective threads for cognitive models of human learning in complex nontrivial environments is one of our goals. Example: Nontrivial Natural System Natural Features: fMRI, EEG, MEG, PET, fNIRS, EKG, EMG, GSR, etc. Example: Nontrivial Cognitive Architecture Preferentially Ordered Declarative Goal Structures A dynamic goal structure distributed throughout a network of interconnected parallel problem solving resources. Imagined Plans: Cooperative Subgoal Collections Collections of sufficient subgoal conditions for comprising modular components of the overall distributed declarative structure. Memoized Mental Resource Simulators Each actor's execution can be memoized dependent on goal structure context, which can be used for simulation without external effects. Traceable Compiled Mental Resource Actors Each sequential effect of these traceable actor is recorded, such that if any error occurs debugging processes can know which parts are responsible. Trusted Compiled Mental Resource Actors Repeated successful execution of traceable actors results in the creation of trusted actors—optimized and compiled for efficient execution PERCEIVE and ACT Natural biological humans and their very complex brains can act within similar behavioral experiments as human-designed computational models of intelligence (A.I. models). Example: Nontrivial Cognitive Environment PERCEIVE and ACT Human-designed computational models of intelligence (A.I. models) can interact with simulated problem environments of nontrivial complexity. Many features of natural processes can be measured, but it is currently difficult to correlate these features with complex models. Part of the problem of correlating complex natural systems with computational models is that very few features of computational processes are currently capable of being measured. We are writing an experimental programming language [see Funk2 Project] for the explicit purpose of measuring all causally dependent computational features of the process. With these new causal reflective computational techniques, finding more accurate and detailed computational models of natural processes will be closer to a reality because then we will be at a point when we can ask the question: "Which types of computation are good explanations of this natural process?" Causal reflective critics help a planner to learn to plan through run-time experience. Cooperative Resource Selector Learners As goals are often pursued and accomplished together these groups can be recognized and remembered for future planning deliberative simulations. History Writers Patterns in traces can be recognized and compiled into simpler representations for other critics to process. For example, all causal dependencies relevant to accomplish a specific goal can be compiled for quick retrieval later. Conflicting Resource Allocation Learners As conflicts are between groups of goals, these goals can be learned to be within mutually exclusive allocation sets. These MUTEXes can be learned critically through run-time feedback. Resource Conflict Blame Arbiters If there is a problem attempting to allocate a lower-level resource for two independent threads of execution, a critic may attempt to discover what two subgoals are to blame for this unanticipated interaction. Problem Distribution Balance Learner Many resources are limited, forcing the serial execution of some goals. We can recognize that some combinations of goals are better than others for either optimal resource distribution or a minimal time until goal completion. Pointless Process Recognizer If there is a process that executes and ultimately has no effect toward accomplishing the goals of the system, note that these processes did not need to be executed in the first place.

Neural Models of Mind Publications

Neural Models of Mind Media Presentations

Neural Models of Mind Developers