Bo's Blog

Call to the Programmer in 2006: Make computers understand people and programming will be easy (or Mental Augmentation)

2006-06-05T08:40:00.000-05:00

Computation allows us to remember things we would have forgotten otherwise. It also allows us to share these memories. Computation allows us to create imaginary memories that might exist in the future, might have existed in the past, or could be a good explanation for what is happening right now. How do we make remembering, imagining, and sharing our memories easier?

How do the problems of individuality, society, and competitive evolution relate to easily accessed historical memories, imaginary memories, and the communal sharing of these memories?

Darwin's formal theory of evolution assumes that there are multiple species and that these species have individuals. The ideas of species and individual have been shown to be pretty fuzzy and they are becoming more fuzzy as we lose individuality through the communal sharing of computational power, one of our forms of intelligence. We are quickly losing the idea of species as humans have become the only specimen in their ecological niche; what does it mean to compete against another species, with which you cannot reproduce? The evolution of computation (memories and the thought process itself) has slowly started to move onto the Internet, where a world-language is evolving and computer languages are becoming very simple to learn and use.

Processing power over the next 20 years will not come from major advances in transistor technology, although quantum and biological computation have large promise. The problem of how to program these new types of computation still lies at the feet of the programmers, and I believe that these programmers will provide the computational power gains of the next 20 years--by rethinking computation and programming. Programming the Internet will not be solved by focusing on programming language syntax (e.g. Perl versus C versus Lisp versus Python versus Ruby versus C-sharp-dot-net-with-COM-and-templates). Easily programming the Internet will not be solved by providing more powerful function calls or great libraries of function calls. We need computers to understand humans. We need computers to be able to learn the goals of individual humans. We need computers that try to help. We don't need more computers that wait to be told what to do. We need computers that can tell when we are confused and that can provide additional data or offer mental assistance. We need computers that can infer our mental states. We need computers that know what we believe about the world and what we want out of life--our goals, desires, and beliefs. We need computers that know us as communities and individuals.

Modularity requires using a hammer to find nails

2005-12-17T10:19:00.000-05:00

In an evolving modular system of algorithms there exist new growth algorithms and old decay algorithms. Hammers are used to hammer nails. If it takes you a long time to build a hammer, and you are especially proud of the fact that you built the hammer, then you will look for nails to hammer with your hammer and because you are so excited to use your hammer, your search for nails will be occasionally interspersed with other things that are not nails, and for them your hammer does not work. This process of using hammers to find nails is a necessary part of the process of building modular systems. You do want to avoid the falacy of not constantly learning how to use new hammers, you will be rewarded by the number of already existing modules you can use in a system because if you use those modules (or algorithm packages) that are near you (default on your operating system), your module will not require the installation and loading of other modules, which makes the usage of your module more costly. Therefore, we must learn to use hammers to find nails in order to better learn to evolve modular systems of algorithms.

ToDo List

2005-12-16T11:08:00.000-05:00

Have you ever made a todo list? Lists in general are a useful tool for dealing with a linear thought process. Lists that are composed of lists can be used to generate hierarchies, which are also a very useful tool for dealing with a hierarchical thought process; for example, a plan that consists of multiple steps that are each composed of substeps. I like to write computer programs that generate my todo lists. I tell the computer the important data to consider (such as the list of graduate schools to apply to with application due dates). With a little coercing, the computer will give me a sorted list of graduate schools to apply to (all of them happened to be due on December 15th so this program didn't actually do much in practice). I consider this to be programming my mental thought process. It would be interesting to see if programming interfaces could become so efficient and easy to use and learn that we all use this method of planning our large decisions (such as life plans). What keeps life exciting? The people at the center of this system. Government regulation of the types of processing that can be done within a person's own mind will become serious societal concerns if we don't think about these issues incrementally: privacy, security, and freedom for the individual.

The Brain Popper (or Brain Stack)

2005-10-03T16:07:00.000-05:00

Imagine an operating system of human processing.

We are nearing a point in brain computer interface technology where in the next few decades we will see cosmetic brain augmentation technology. Today, BCI technology is to the point where we can feed visual and auditory information into the brain; retinal and auditory implants, while not common are becoming more common among the deaf as the camera technology is not yet low power enough to enable full retinal implants (currently requires cameras on glasses). These technologies are evidence of "input" technology, but more recently we have seen a rise in so called "output" technologies; humans are now being implanted with simple controllers that are activated by the motor and premotor cortices, which is to say, one can imagine moving a muscle and a prosthetic limb motor (or a computer mouse cursor) will respond. Currently, in order to read information out of a human brain, we must put a bundle of wires that each basically sense on the order of 10 neurons each, and in order to feed information into the brain, groups of neurons must similarly be stimulated by many wires. This limitation of information flow into and out of the brain has not allowed large portions of the brain to be "read" from or "written" to, but the implications of this ability will definately drive the research in this direction. The Brain Popper is one such implicated application.

The Brain Popper is a device that reads a critically large portion of the processing state of a person's mind state. This includes short term memory that one must "warm up" whenever they approach a new type of task, even if they are adept at this particular task. For example, this type of short term memory would include "riding a bike" or "driving a car" or "flying a kite" or "driving a remote control car" or "sorting bills" or "cooking lasagne" or "watching a TV show" or "reading email" any other task that one can imagine that takes skill and that one "mentally warms up" as they continue to perform the skill -- anything that can be interrupted and that takes some time to get back "into the groove". The idea in computer science terms is this: The Brain Popper keeps a stack of mental states that can be read and written to the working memory of an individual. The Brain Popper in common sense terms allows one to save one's mental state while performing a task, and then perform another task and then "Pop!" return to the first task with the push of a button without "warming up" to the first task again. Long gone is the era of distractions. If you ever get "into the groove" just hit the "Save" button on your Brain Popper and then you just return to that mental state whenever you want by the touch of a button. (In actuality, the Brain Popper only saves short term memory, so any changes in long term memory between saves and restores would not be accounted for, so saved states of mind would probably only work on the order of weeks or would need to be refreshed every week or so depending on how often the long term memory regarding the task changes).

Implications for "group think" and social processing strategies are only on the horizon. Feeding experiences from those on the forefront of human experience (frontlines of battlefields, outer space, old age, young age, or in general any experience that is difficult to communicate through language will also be an exciting brain read/write applcation.

Local cost gradient descent

2005-09-06T20:47:00.000-05:00

The act of moving toward where we as individuals expect the life of lesser negativity to be. It can be assumed that information is communicated in order to nudge each individual person closer to an evolutionarily useful view of a situation. Information is communicated in order to improve individual survival and reproduction through the use of a distributed information processing social network. Those who could exchange information efficiently in this model were favored and [Q.E.D.] we are efficient information exchanging individuals within a social information process.

Human Model

2005-07-28T08:05:00.000-05:00

Models of humans and human societies will become more popular because they can be communicated at the speed of light to other societies for the informative collaboration within a galaxy, allowing relatively equal human interaction for any star.

Milky Way

Recurrent Neural Networks

2005-05-12T14:40:00.000-05:00

This has just dawned on me and it is exciting for some strange low level (but inspirationally high level but as yet unexplicable) reason: recurrent neural networks can model processes such as the belief propagation algorithm for markov random fields. The weights of recurrent neural nets can probably also be tinkered with in order to train a subnet within the network to learn a feedforward function. I guess this makes sense because it is probably a turing complete system, but this is still exciting to me for some reason. I think I need to play with finite state machines a little more so that I see that this same thing cannot be simplified to just bits and more basic programming ideas.

Evolving Neural Programs

2005-04-06T20:42:00.000-05:00

Neural networks, probabilistic inference, and combinational devices form a subset of low level neural processing. These systems can all be iteratively clocked forming different representations of general processing elements. These low level process elements can be represented as frames with slots for the specific type of process and variables (neural network weights, probabilistic tables, or combinational relations). A frames' inputs can reference another frames' outputs. A frame can be duplicated and mutated in order to search for more useful representations through Hebbian similarity/recognition/learning.

How can a frame based system composed of general frame-based parallel processing elements be debugged? Who/what recognizes bugs, proposes solutions, tries solutions, remembers successful strategies, remembers failed strategies? (i.e. good ways to divide a space, simplify processes) Can these critics replace themselves through their own process of debugging? (by maybe creating imaginary copies of themselves in order to debug themselves and test the outcome against other critics?)

Why has the field of neural networks come to the dead end without becoming self-reflective? Are there examples of self-reflective neural networks?

Abstraction without losing the Details

2005-01-20T10:58:00.000-05:00

In the recent proliferation of programming languages to handle the many specific internet domains, there have been languages that provide abstract ways to represent and manupulate computers for a variety of purposes. For example, Java has provided the virtual machine abstraction, allowing programmers to write a program that runs on a large number of platforms. The virtual machine abstraction reduces the control the programmer has over the machine however, and complex algorithms or data intensive processing fail in Java because of the memory and speed processing control constraints. The general trend toward more abstract languages points out an important need in the programming community: C (the language that major operating systems are written in these days) is too "low level" to quickly, spontaneously prototype and develop operating system independent programs. The benefits of quick development in abstract languages outweigh the control/speed advantages of C/C++ for most developers. So, why don't we have langauges that allow full control of the machine with the benefits of abstraction as well? The answer for me is to write a language that combines the control and speed benefits of C++ with the clean syntax and abstraction capabilities of Lisp. I'm currently working on a language on top of Lisp that writes C++ code and perform compilation with the GNU C++ compiler. The process of compilation and execution is abstracted. I'm hoping that this will help to speed up my development time to be as fast as any other Lisp program, while having the advantage of C++ speed and control. ANSI C++ is of course operating system independent, so the programs developed could be compiled on any platform that compiles ANSI C++ code.

Critical Self Reflection

2004-11-26T13:55:00.000-05:00

There are a lot of ways to think about the world. Some are inconsistent, and others are more inconsistent. In a world in which the passage of time is measured by change, to say that something could be consistent through the passage of time would be a contradiction. Here, my stated goal is that I would like to build a system that is a useful model of the world. In order for a model to be useful, it must not be consistent because this seems to be a falacy of language under the current assumptions, but I will pose the goal to the engineer of such a system to create a system which becomes less inconsistent through time.

Children are not completely inconsistent in that they have pre-programmed reflexes for dealing with many aspects of their environment (social, perceptual, and physical). The child-parent relationship provides the social reflexes necessary to sustain a highly ineffectual and inconsistent state of initial being. A child's gestation period involves active mental organization of which only the most peripheral aspects are present within the womb.

Imagine the growth of structural representation of perceptions that self-organize into equiprobably distributed categories. These form the lowest level perceptual inputs. We must avoid the hierarchical representational falacy, which would later restrict our model to be unnecessarily compartmentalized. Instead, I propose the recognition of pairs of temporal transition states as the primitive atom of representation within the system. Methods of optimization and parallel processing using (possibly) graph clustering techniques could later be applied to this sort of model. This process will allow us to build an undirected model of perception. Probabilistic transition functions could be calculated using the Markov state model assumption that the entire system state is represented within the system without needing to remember previous states of the system. This is accomplished through state augmentation using the same representational elements just described.

Of course, the process of perceptual organization can only abstract up to a certain "useful" point determined by the processing and memory capabilities available without. In order to build a useful model of the world, the process of perceptual organization and prediction must be directed by a critically self-reflective process that critically acts in order to break down assumptions about the way the world works. This grounds out in terms of the specific state-transition functions that exist within the model that can be experimented with under different circumstances in order to falsify those transition functions and replace them with smaller piecewise replacements of the originals. This is a process that I refer to as critically self-reflective break down of assumptions.

I have used the term "useful" throughout this rant without definition and I propose this model explicitely without a definition of useful. Useful is meant to be some measurement of utility and utility is measured differently by every system and can be easily explained only for very simple systems. A criterion for utility in terms of a general intelligence would probably be best left at "pleasing the user", how this is measured will be an active area of current marketing and public polling research.

Peace, Love, and Intelligence

LifeNet Beta Demo

2004-10-22T15:17:00.000-05:00

I've just added a demonstration of LifeNet to my lab page, which uses the Commonsense data from the OpenMind project. LifeNet is a temporal inference model. The underlying probabilistic model is very simple and currently only returns very fuzzy results, so the Commonsense group is looking into many different ways to gather more temporal commonsense data, including logging information from volunteer's cell phones about their daily activities. The PlaceLab, part of the House_n project, is also a possible source of learning temporal data. Highest priority at the moment though, is to make the probabilistic backend more precise without limiting the expressivity and ease of use regarding the human language interface.

Transitive Inference

2004-10-10T16:31:00.000-05:00

There are many probabilistic methods for performing temporal inference using models such as Dynamic Bayesian Networks and Temporal Markov Random Fields, but the patterns that are hardcoded into these models about time can be abstracted to a more general representation. More general representations have the advantage in learning problems of being able to reuse learned patterns in describing complex relationships.

The more general pattern that can be abstracted from temporal inference betworks are transitive relationships. Transitive inference networks would allow inference along any number of axes, not only the temporal axis. Obvious examples of transitive inference are relating the concepts of dimensionality, such as space and time. Thinking about moving forward in a spatial dimension is similar in many ways to thinking backwards in time and the primitive patterns used to learn these relationships could be used to distinguish differences and establish a deeper understanding of similarity within patterned data.

Traditional approaches to the temporal inference problem, such as dividing the inference space into duplicate slices of the entire state space are too naive an approach for a general transitive inference problem as the number of dimensions approaches the number of relationships within the model. Size, shape, quantity, brightness, speed, all of these share in defining transition relationships among data. The model cannot simply be naivy expanded in all of these dimensions resulting in explosive requirements for memory and computational resources.

My First Blog Post

2004-10-08T17:39:00.000-05:00

I just created my first blog. Congratulations my self. Keep checking back for more exciting details!