It's About Time Neil Madden and Brian Logan University of Nottingham School of Computer Science and Information Technology Jubilee Campus, Triumph Road, Nottingham, NG8 1BB. { n e m , b s l } @cs.nott.ac.uk 1 Introduction rules that are used to spot patterns in input data, and draw conclusions based on these patterns. Conclusions are writ- A number of experimental studies, e.g., [Geldard and Sher- ten to an output bus that transmits them to the input buffers rick, 1972; Kolers and von Grunau, 1976] have suggested ¨ of other processes. In this way, data can be abstracted as it that interpretations of events can override direct sensory evi- progresses through the network of connected processes, with dence. For example, for some sequences of perceptual events different abstractions persisting for different lengths of time. of short duration, the interpretation of individual events in the We envisage that processes further up a chain (further from sequence depends on the characteristics of the sequence as the initial percepts) would have buffers spanning a larger du- a whole. This `backwards referral in time', in which later ration of time than the lower level processes, allowing the events influence the perception of earlier events, is difficult to system as a whole to remember more abstract conclusions, account for within a serial model of cognition without incor- while most of the details are forgotten. porating implausible delays (basically delaying sensory expe- The bus connection architecture allows a single conclusion rience "until all the data is in"). from a low-level process to be delivered to multiple higher- Dennett and Kinsbourne [1991; 1992] have proposed the level processes, allowing for multiple drafts to be formed Multiple Drafts theory as a way of modelling such cognitive based on the same data, potentially producing different con- processes. The Multiple Drafts theory is based on a paral- clusions or interpretations. lel, distributed view of cognition, in which large numbers of processes work independently on multiple interpretations of 3 The Cutaneous "Rabbit" Model data simultaneously. These are the multiple drafts. Eventu- ally a single draft may become dominant, but no draft is ever We have used the architecture outlined above to model a num- ber of perceptual phenomena. In this section we briefly de- entirely safe from revision. scribe one such model: the cutaneous "rabbit". Geldard and Sherrick's cutaneous "rabbit" experiments [1972; 1977] illus- 2 The Temporal Abstraction Network trate a perceptual phenomenon called sensory saltation. In Architecture the experiments a series of short `taps' (of about 2ms dura- We present a cognitive architecture for perceptual processing tion) were delivered to different locations on the arm of a which draws on aspects of the Multiple Drafts theory. Our subject -- for instance, five taps at the wrist, followed by five Temporal Abstraction Network (TAN) architecture consists between wrist and elbow, and then five more at the elbow. of a set of processes, each with its own state, represented by a Subjects reported that the taps had been more or less evenly time-limited buffer, along with a procedure for drawing infer- spaced along their arm -- as if a little rabbit was hopping up ences based on the current contents of the buffer. These pro- the arm. Variation in the interval between taps (inter-stimulus interval, I S I ) causes differences in the perceived effect. If cesses are connected together via a bus architecture, allowing the I S I exceeds approximately 200 ms the taps are perceived the conclusions drawn by one process to form the inputs to at their correct locations. With an I S I of 20 ms or less, some other processes (including themselves), see Figure 1. Each process has an input buffer with specified capacity taps `disappear', with say 15 taps being perceived as just 6. and duration. Duration is the maximum length of time ele- Our model aggregates information about individual taps ments can remain in the buffer before they are forgotten. Ca- into information about a sequence of taps (see Figure 1), al- pacity is the number of items that may be present in the buffer lowing the agent to reason about and predict the behaviour of at any given time. The duration and capacity of a buffer are an object over time. At the lowest level, the model processes independent of each other, e.g., a buffer may have large ca- sensory information to determine the presence of a single tap. pacity but short duration or small capacity but longer dura- This process has a buffer duration of about 20 ms and a ca- tion. New inputs are added to the buffer in first in first out pacity of just a single element (in this case, the `element' fashion--items arriving at a full buffer cause the oldest items is actually a collection of low-level data). If more than one in the buffer to be overwritten. tap occurs within this time-frame, then the newer tap simply overwrites any previous tap. This is consistent with the ex- Each inference process also contains a set of production 'Rabbit' run Two taps at wrist Drafts theory. However these models generally involve some serial component where "everything comes together". For instance, in CopyCat [Mitchell, 1993] multiple parallel pro- Inference Engine Inference Engine cesses operate in a stochastic manner on a single workspace, creating a single solution to a problem; in EPIC [Kieras and Buffer Buffer Meyer, 1997] there is a single central executive which acts to coordinate the parallel processes. Bus In contrast, our Temporal Abstraction Network architec- Tap run from s to e (n taps) ture has no global coordinating or integrating facility. How- ever, abandoning a single central executive process does not Inference Engine mean that information cannot be brought together locally for Feedback Loop integration. The TAN architecture is capable of local (serial) Buffer integration while maintaining multiple simultaneous drafts: information flow can diverge as easily as converge. This Bus approach is in contrast to that of Dennett and Kinsbourne who suggest that the only alternative to the Cartesian the- Tap at position, p atre is a strictly parallel architecture, where local integration is replaced by a more chaotic Pandemonium approach (e.g., Inference Engine [Kinsbourne, 1994]). In future work we plan to concentrate on extending the ar- Buffer chitecture to account for action selection, as well as expand- ing on the details of how reports are generated. One interest- Low-level sensory data ing area for future research will be to look at Libet's contro- versial experimental results [Libet, 1985] on voluntary action. Figure 1: Network for cutaneous "rabbit" experiment. References perimental results which indicate that taps occurring within [Dennett and Kinsbourne, 1992] Daniel C. Dennett and Mar- about 20 ms of each other are merged with the location of the cel Kinsbourne. Time and the observer: The where and newer tap dominating. when of consciousness in the brain. Behavioral and Brain The intermediate level processing has a buffer of duration Sciences, 15(2):183­247, 1992. of 200 ms and a capacity of 2 elements. When a tap arrives at [Dennett, 1991] Daniel C. Dennett. Consciousness Ex- an empty buffer (which can happen at most once every 20 ms) plained. Penguin, London, 1991. a new aggregate conclusion is generated, taking the position of the tap as the start and end position of the `run', and initial- [Geldard and Sherrick, 1972] F. A. Geldard and C. E. Sher- ising the count of taps in this run. This conclusion is passed to rick. The cutaneous `rabbit': A perceptual illusion. Sci- the output bus, where it is transmitted to other processes, but ence, 178:178­179, 1972. also, via a feedback loop, back to the input bus of the inter- [Geldard, 1977] F. A. Geldard. Cutaneous stimulis, vibra- mediate level process. If a subsequent tap arrives before this tory and saltatory. Journal of Investigative Dermatology, aggregate fact expires from the buffer (i.e., within 200 ms) 69:83­87, 1977. then a new conclusion is formed which adjusts the end point [Kieras and Meyer, 1997] David E. Kieras and David E. of the run to the new tap position and increases the tap count Meyer. An overview of the EPIC architecture for cognition by 1. The buffer duration ensures that any gap of 200 ms or and performance with application to human-computer in- more causes the previous `run' to be forgotten, and thus any teraction. Human-Computer Interaction, 12(4):391­438, subsequent tap will be perceived as the start of a new run, 1997. which is consistent with the experimental data. At the highest level of processing (top left in the Figure), a [Kinsbourne, 1994] Marcel Kinsbourne. Models of con- process with a two element buffer detects the end of a tap run sciousness: Serial or parallel in the brain? In M. S. Gaz- by comparing the start position of sequential tap-run inputs. zaniga, editor, The Cognitive Neurosciences, pages 1321­ It is the output from this process that is eventually used to 1330. MIT Press, Cambridge, MA, 1994. generate a report of the experience. It is important to note [Kolers and von Grunau, 1976] P. A. Kolers and M. von ¨ that although the intermediate-level buffer has a duration of Grunau. Shape and color in apparent motion. Vision Re- ¨ 200 ms it is not necessary to delay conclusions for 200 ms. search, 16:329­335, 1976. Instead, the process produces a conclusion whenever a new [Libet, 1985] Benjamin Libet. Unconscious cerebral initia- tap is felt (at most, once every 20 ms), and these conclusions tive and the role of conscious will in voluntary action. Be- can be acted upon immediately. havioral and Brain Sciences, 8:529­566, 1985. 4 Discussion [Mitchell, 1993] Melanie Mitchell. Analogy-Making as Per- ception: a computer model. MIT Press, Cambridge, MA, Th er e a r e a n u m b e r o f p ar allel m o d e ls o f co g n itio n ( e.g . 1993. CopyCat, EPIC) which have some similarities to the Multiple