Why Do I Approach Developing MOM The Way Visible By MOM SSC?

The MOM Simple Set Core (MOM SSC) is the most recent implementation of MOM. MOM is a trinity of research, development and a project driving both of them ahead. In core, MOM is the Model of Meaning plus research based on that model, aiming at representing every kind of content bare of words and tagging, only based on graphs and bare input sensors, such as 'light given', 'oxygene here', 'soft ground'. -- However, since 'there is a red light under that passenger seat, calmly blinking' is a bit more complex content, and that content is not yet developed by graph, currently MOM accepts crutches -- labels or pieces of software that signal a certain event being given, e.g 'web browser cannot render that page correctly'. As MOM improves, such crutches shall get replaced by the more flexible (and error resistant) representation of content as offered by MOM.

There are several promised benefits of that. Getting content available without words implies the the chance to render content to any language of the world. Getting there without tagging implies the chance that the machine knows of the content represented -- instead of just dealing with it but remaining unaware of what it means. That in turn implies the chance to load content ("knowledge") into any sort of machines, such as traffic lights or vacuum cleaners or cars. Whereby to load the knowledge might be much a bit quicker but needing to train any sort of neuronal network AI. -- MOM is not after implementing any sorts of artificial intelligence but heads for getting the content available. Call it a [content-addressable] memory.

That error resistant representation of content beforementioned originates from another core part of MOM, the recognition. -- Yes, that's right. MOM found recognition to be a part of memory, not of any sorts of intelligence. It's an automatic process which, however, might be supportable by training [link: "is it learning?"]: weighting the graph's edges. [It's clear to me that humans can improve their recognition, but I am not sure whether the causes of learning equal those of improving the recognition abilities of a MOM net, hence the differentiation.] Core of MOM's recognition and cause for its error resistance is that while the MOM net defines every possible feature of an item, for recognition not every such one must be given, only a few. -- Which, by the way, matches a claim recently posted by Chris Chatham: Only a few of the features of a known item result in a correct recognition of that item because there are only the yet known items out there: To discern all the items being similar, you don't need that many different features. But wait the day you encounter an in fact new item! -- You'd get it wrong, in any case. Remember the days you were familar to dogs as the only kind of pet animals? Then, encountering the first pet cat, you likely named it 'dog', din't you? Same so for any kind of flip pictures, like the one you can either see a beautiful young woman in or a rather old one. -- To get back to Chatham: On the issue of change blindness he claimed "[...] the brain is 'offloading' its memory requirements to the environment in which it exists: why bother remembering the location of objects when a quick glance will suffice?"


Along with research, MOM is a project of development. I am used to program, hence cast MOM into software is the most clear way to go. MOM, casted to software, allows for verifying the model. Also, over time, a full implementation of MOM might result, hence achieve to get handy all the chances MOM offers.

For example, the MOM Simple Set Core (MOM SSC) originally was only after implementing the MOM net, i.e. the functionality to maintain (parts of) a MOM net in computer memory (RAM). That's overcome now. Now, going further ahead, MOM SSC aims at implementing the reorganizer. That's a share of MOM which shrinks the graph by kepping the same content -- yet even revealing content which was only implicit beforehand.

Former versions of MOM parts were implemented using Perl. For reasons of readability, for MOM SSC, Ruby was chosen. Since the theoretical work on the reorganizer it was clear, the reorganizer modifies the MOM net, hence challenges the strengths of the recognizer. To get able to make the recognizer perform well even on reorganized MOM nets, I now begun to implement the reorganizer. Having it in place, research on the recognition might go into depth. Especially since having a reorganizer in place implies to get enabled to automatically test quality of recognition: Recognition on the reorganized net should provide the same results as recognition performed on the original net. Fine part is, neither reorganization nor recognition need any labels for the nodes (i.e.: no mark-up/tagging).


Upcoming milestone of the MOM SSC sub-project might be to implement the core of the reorganizer, accompanied by full duck typing approach for the MOM SSC classes, or/and by fixing all the chances for improvement, which accumulated over time since the beginnings of MOM SSC. -- Core of the reorganizer is to detect and replace sub-networks of the MOM graph that occupy (far) more nodes/edges than necessary to represent a piece of content. The replace would be to reduce these sub-networks to just as many nodes/edges as actually needed to represent the content.

      
Updates: none so far