linkedin facebook twitter rss

02 Aug Artificial Time

Time and Space PerceptionTime is omnipresent – you can’t get away from it. It is woven into everything we do and say and understand. It is an inextricable element of context. I was just speaking of how the connections in our brain develop, grow and evolve over time. Representing and handling this “temporal” element is fundamental to any system that attempts to model human cognitive processes, including communicating with language.

Why do we need time, and how do we use it? Reports on information used by businesses are often segmented by time, such as weekly, monthly or annually. Commercial transactions are universally recorded with the time of the transaction as a key element. We also use time and timelines in planning and scheduling events that are to occur in the future. We need time to represent not only the historical and future sequences of things, but to represent relative durations, including process and event modeling. Appointment scheduling applications are based and centered on representations of time, but most other applications have some way of processing time in a meaningful way, if only to record when a data file was saved.

By the way – I no longer wear a watch – my devices, cars and preferred media sources all keep me apprised of the time so who needs the extra hardware!?

Understanding Context Cross-Reference
Click on these Links to other posts and glossary/bibliography references

  

Section 7 #33

AI Modeling Section Icon

 

Table of Context

Calendar ClockTime is also a key element in language, so the system that understands enough context to be able to interpret a person’s intent based on their words, will need to be able to understand the implications of the passing of time. By the way – I no longer wear a watch – my devices, cars and preferred media sources all keep me apprised of the time so who needs the extra hardware!?

Modeling Time in ANS

Attempts have been made to enable Artificial Neural Systems (ANS) to deal with temporal elements of computation by incorporating different spatial metaphors for time. In one such model, time is represented explicitly by representing the sequential order of the pattern with juxtaposed processing elements (PEs) in the pattern vector. The serial inputs are simply pushed through the network arm-in-arm, occupying the same temporal machine cycles but segregated to parallel clusters of neighboring PEs (Cottrell, Munro & Zisper, 1987; Hanson & Kegl, 1987).

ANS with two Hidden LayersBesides requiring a shift-register buffer to collect a series of data from the input stream, such an approach requires that all the data share the same dimensions (each datum in the stream must be of equal length). As there are no structures in the nervous system that seem to act like shift registers, and data is unpredictable, this method of temporal processing is not particularly suitable from the neuromorphic standpoint, although it is certainly a valid approach for simulating a temporal representation. If all MIPUS’s impulses traveled through his neural network in lock-step, he could not keep up his wild day-dreaming while serving dinner. Healthy robots need good fantasies.

The Temporal Element

The temporal element may be more active in some cognitive processes than in others. The visual system, for example, may require repeated or residual activation to simplify interpretation of motion. But the visual system may not require residual activation to process different images in rapid succession, like a slide show. When the head is rotated full circle, as in searching for a person in a large area, the absence of temporal dependency could be described as a “snapshot” effect in which all the information required for interpretation is immediately available. A fraction of a second of exposure (glance) is enough to produce a clear photographic imprint of an image in the brain. The information in that imprint is likely to be transient and self-contained; it can be complete and understandable just as a photograph is independently understandable, then completely forgotten.

A split-second exposure to language data, however, whether written or spoken, has less information value. A single loud exclamation, such as “STOP” or “FIRE – RUN” may draw a person’s attention and motivate action, but the real information value is less than the shock value. The temporal flow of text or speech data is only meaningful in sequence. The longer the exposure available, the more likely the context will be understandable, and the more meaningful it becomes. While snapshots may be fine for visual processing where the spatial element of context dominates, an active and prominent means of temporal processing is essential for abstract thinking such as language understanding.

Time and Directionality

Some of the most successful and flexible vision processing systems use multi-directional flow of E/I. In Fukushima’s neocognitron, for example, both forward and backward propagation are used (1988, p. 66). This approach is quite useful in the image-recognition domain, but it may be less useful in more abstract or non-imaginal types of processing because it uses only snapshots of input. There is no stream of input: the network only processes a single two-dimensional image in any given time frame.

Neocognitron

The backwards flow in Fukushima’s model is appealing because the effect resembles cuing, an important context setting process. When parts of the pattern are recognized, the output is sent backwards through the network, creating a kind of adaptive resonance that resembles selective attention. This helps the sensory input portion of the network focus on recognizable areas of the pattern (Fukushima, 1988, p. 67). The paths of backward and forward flow in Fukushima’s network are separate, but through interconnections between the paths, the backward selective-attention signals facilitate the forward signals, while the forward signals provide virtual gates for the backward signal flow.

The elaborate scheme embodied in the Fukushima model is consistent with the cuing principle; for image processing, it works extremely well. Because of the temporally restricted frame of input, however, it does not resemble the type of processing required for language processing, reasoning, or other cognitive activities that involve a stream of input and a volley of forward and backward constraints that govern the output. Ergo, while this can serve as a good component for a model for language understanding, it cannot handle all the processes, especially change over time, needed for robust interpretation. MIPUS, you and I will need something more robust to be able to perform more complex cognitive tasks, especially understanding human language communication.

Click below to look in each Understanding Context section


 

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.