HIPReading2 - 5mm Human Information Processing 91-2 The...

Info iconThis preview shows pages 1–11. Sign up to view the full content.

View Full Document Right Arrow Icon
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Background image of page 2
Background image of page 3

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Background image of page 4
Background image of page 5

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Background image of page 6
Background image of page 7

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Background image of page 8
Background image of page 9

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Background image of page 10
Background image of page 11
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: 5mm Human Information Processing 91-2 The Systems Approach (or Human as Component) In any system combining humans and machines, the greatest efficiency and utility comes from exploiting the strengths of each as much as possible. Since from a"systems viewpoint, both the human and the machine can be viewed as “black boxes” one can start to characterize both “components” as the first step to partitioning the task between the human and the machine. If we View the human as an information-processing device, it is perhaps wise to compare it against machine information processing (ie. the computer). Some of these differences are outlined in table 1.This is at best, an informal guide as it cannot possibly define all the subtleties of humans and computers. It would be wrong and misleading to apply it dogmatically. A case in point is “numeric manipulation”, where humans compare poorly against the computer. While true in the sense of explicit numeric manipulation such as adding a column of numbers, it is completely misleading when dealing with processes that we model mathematically. For example, the equations capable of controlling a robotic arm to open a door are beyond the state of the art in computing, yet in a sense, those very same equations are implicitly dealt with (in real time) by even the most innumerate human. Characteristic Human Computer General Characteristics - Performance - Adaptability - Creativity - Numeric manipulation - Handling complexity - Behaviour Problem-solving skills - Cognition - Data format sensitivity data poorly - Pattern recognition - Handling ill—defined problems - Manipulative skills M e m o ry - Reliability - Short-term capacity - Access method - Data interconnections Variable Very adaptable Very, able to deal with unexpectedness Poor, slow Good at some, poor at others Knowledge-based Strong Handles unstructured regardless of format Excellent Excellent, can synthesize Good Poor Oddly limited Content-addressable (parallel search) Rich Failure Characteristics Precise, repeatable Inflexible Uncreative, lacks initiative Accurate, fast . Good when specifiable Rule-based Weak Generally handles data well Very poor Unable Poor Super-reliable, accurate Practically unlimited Location-addressable (sequential search) Usually rigidly structured ENSW Human Information Processing 91-2 - Failure type Gradual degradation Sudden, catastrophic failure - Failure modes Partly irrational Rational - Response to delay Impatient, “deranged” Oblivious (“patient”) - Stamina Requires rest & recreation “Tireless” - Self image Requires good self image No self image Miscellaneous - Perception of universe Subjective Objective - Physical environment relatively delicate and fragile can resist adverse or hazardous environments - Sensory range not diverse, but broad diverse, dynamic range and dynamic range resolution somewhat limited - Capital/operating expense continuously costly cheap (when mass- produced) Table 1: Comparison of Human vs. Computer Characteristics (Note: strengths shown in bold) It can be seen from table 1 that humans excel at dealing with ambiguous situations requiring pattern recognition, creativity and synthesis. Computers are best at dealing with high accuracy, repeatability and numerical applications, and are easier to “harden” to harsh environments. Consequently, you the designer, should'partition the subtasks of a human/computer system to exploit the human’s creativity, and the machine’s accuracy. Measuring Human Information Processing Capabilities As psychologists will freely admit, the human mind cannot be probed directly or piecemeal like a human-built construct can. Instead, one presents the “black box” with stimuli, and measures the response of the entire sealed unit. Deducing the underlying mechanisms is extremely difficult and is analogous to (though infinitely more complex than) reverse-engineering a VLSI chip simply by probing the chip pins. The simplest stimulus-response pair to measure is response time or reaction time. Donders devised three reaction time experiments that give us a measure of human performance. The first, or Donder’s A reaction, is simple reaction time, with l stimulus and 1 response (eg. pressing a button when a light goes on, or answering a phone call) as shown in figure 1. This measures the nervous system’s conduction time — the minimum-path time through the system. The second experiment, yielding the Donder’s B reaction, presents several possible stimuli, each with its own unique response (eg. N lights matched to N switches, or one’s response to a traffic light), as shown in figure 2. The third experiment, for the Donder’s C response, presents several stimuli, only one of which is mapped to a response as shown in figure 3 (eg. one must press a particular switch if any of M out of N lights go on, or responding to a mail—call). Subtracting the A- response time from a C-response time gives an estimate of mental identification time (ie. the time needed to identify a particular stimulus), while subtracting the C-reaction time from a B—reaction time gives an estimate of the mental selection time (ie. the time needed to select one of a set of responses) (figure 4). Repeating the experiments with different sizes of M and N begins to crudely W 2 ENsc-1o4 Human Information Processing 91-2 map out some aspects of human performance. Stimulus Stimuli I O O Stimuli Response Responses 35pm“ Figure 1. Simple reaction Figure 2. The choice reaction Figure 3. The Donders time task. A light (rep- time task. Two (or more) lights C-reaction. There are resented as a circle goes are mapped to responses. several stimuli but only on and'a response (rep- Each stimulus has its own 1 of this is mapped to resented as a square) is unique response. _ a response. made. ' i I—Rnetion mm x”: rpm , ' I" . WW :— Relation limo W 2.7.1.533“ . t - a ' ldnntifiulion limo ‘ mm - b-‘i-"W" W" , ._ , > I W g-fiuclion tim- a—c-s-imionximo “2.:Eig1ure 4. An example of Donders' subtractive logic. Subtracting a—reaction time from c—reaction time (top panel) gives an estimate of mental identification time. Subtracting c—reaction time from breaction time gives an estimate of mental selection (bottom panel). The 3 different shadings show the proportion of reaction time alloted to each mental process. Figure 5. The relationship among informational quantities. Transmitted information T(S:R), the shaded area, si the intersection of the stimulus information H(S) and the response information H(R). The joint information H(S,Fl) is the union of the Venn diagram. As one would expect, the reaction time increases with the amount of information (ie. number of stimuli or responses) that needs to be processed. However, it is linearly proportional to the transmitted information, not the information of the stimuli or the response. Figure 5 shows this as a Venn diagram, where H(S) is the self-information of the stimuli, and H(R) the self—information ENsmM Human Information Processing 91-2 of the responses. Their joint information H(S,R) is the union of the two sets, or the oo—shaped outline of the diagram. The transmitted information, T(S:R), is the intersection of the two sets. The relationship between them is ‘ T(S:R) = (H(S) + H(R)) - H(S,R) but perhaps this can be made clearer by the following example. In a particular control situation, the operator must operate a set of 4 valves in response to 4 lights on the control panel. Over a series of runs, we get the following information: Stimuli Res onses Hs = 20 5 0 0 21-h = 5 15 5 O 2 bits 0 5 15 5 I68 I onse 0.25 0.5 0.25 0.30 0.20 0.5 0.521 0.464 1.9855 bits Hr=2Hi Thus, H(S) = 1.9855 bits, and H(R) = 2 bits. To obtain H(S,R), we sum up the information of all the stimulus/response pairs (ie. the contents of the SR matrix): Stimuli Res onses 1 0.20°10g2(0.20) 0.05°log2(0.05) 2 0.05-log2(0.05) 0.15-log2(0.15) 0.05010g2(0.05) 0 3 0 0.05-log2(0.05) 0.15«10g2(0.15) 0.05-log2(0.05) 4 0 0 010-10210 0.15-loz(0.15 which equals Stimuli Res onses Valve 4 0.2161 0 0.2161 0.4105 0.2161 0 0.2161 0.2161 0.4105 ‘thh-I The sum of which, H(S,R) equals 3.1086 bits. Plugging these into the equation we solve for T(S:R): . T(S:R) = (H(S) + H(R)) — H(S,R) ENsc-104 Human Information Processing 91-2 T(S:R) = (2 + 1.9855) - 3.1086 T(S:R) = 0.8769 bits (about 44% channel efficiency) Given the transmitted information, Hicks’ Law predicts the reaction time (RT) as RT = a + b-T(S:R) Here, the constant a is an offset relating to the Donders A reaction time, and b is a “fudge—fac tor” determined experimentally with individual subjects or populations of subjects. In essence, a and b compensate for the mapping between the theoretical information content and the much-larger perceptual information content (eg. a physical deck of playing cards, with the colours, graphic details, etc. contains considerably more than 5.70 bits). Much of the speed and capacity of human information processing is based on its predictive abilities, where the human’s current model of the universe is compared to the perceived universe. The difference, or error term, is used to update the model, which provides a prediction for the next iteration.1 For instance, what is wrong with the following sentence? Jack and Jill went went up the hill to fetch a a pail of milk It is not that Jill and Jack usually go up the hill to fetch a pail of water, but rather the duplication of the words “went” and “a”. The mind interprets the duplicated word as redundancy in the channel and simply absorbs it instead of regarding it as an error. Another example of not only the existence of the mind’s predictive algorithm, but also of some of the dangers it leads to, is the following aphorism. How many f’s are there in this sentence? FINISHED FILES ARE THE RE- SULT OF YEARS OF SCIENTF— IC STUDY COMBINED WITH THE EXPERIENCE OF MANY YEARS 1 This is analogous to linear predictive coding (LPC), used as a speech compression algorithm, where sender and receiver start with the same “model” or state, and only the relatively small error term is actually transmitted. 5mm Human Information Processing 91-2 You probably said 3, most people do. Actually, there are 6, and the missing 3 are the f ’s pronounced as v’s. In an attempt to speed the processing, the mind runs the sentence through the (auditory) linguistic centre, looking for “eff” sounds and consequently filters out the “vee” sounds. In fact, this deletion of the unexpected is one of the major contributors to aircraft accidents ——~ valuable seconds are lost while the flight crew overcome their incredulity over an emergency warning in a normally ultra-reliable environment. As long as designers are aware of these dual dangers of this predictive system, they can nonetheless improve the reliability of a user interface by making the user interface as consistent as , is reasonable —— exploiting the same predictive ability. Unfortunately, the amount of information humans can process not only varies greatly between individuals, but is sensitive to the situation (ie. calm or panicked), the number of sensory channels being used (sound, sight, feel, smell, taste, acceleration, etc), the number of stimuli using each channel (the more stimuli using a channel make it harder to discriminate between them), and the density of the information coding used by the stimuli (eg. deciphering Morse code requires more concentration than listening to a voice transmission). Consequently, there are basically no practical metrics for the designer. Humans are much better at dealing with information overload than are computers as they degrade “gracefully” or gradually get worse, instead of failing catastrophically. Humans instead start to miss information completely, and as the processing load increases, the system will try to shed ever more of the load. If this overload is chronic, they will also subvert the system by disabling (eg. disconnecting, or hiding) the information sources that they perceive as being of the least use to them. For instance, when overloaded with monitoring their aircraft instruments, scanning the sky for enemy aircraft, keeping an eye on friendly aircraft (to maintain formation), scanning the ground for SAMSZ, monitoring multiple ground controllers and- tactical channels on radio, listening to the “lock—on” tones from their air-to-air missiles and from two radar warning devices in addition to flying their aircraft at high speed and low altitude, American pilots in the Vietnam conflict started to systematically turn off virtually all of their warning equipment because they felt that they had a better chance of surviving the mission with their natural senses than they did trying to process too much information, much of it inappropriate to their needs. The system designer should, in decreasing order of priority 1. Reduce the volume of information the user must process to accomplish the task. 2 Choose information that is relevant to the task. 3. Format the information in a way that is quickly comprehended. 4. Prioritize the information so that urgent information is recognized first by the user, or provide a mechanism whereby the user can suppress non-essential information (lack of this was a major contributor to the 3-Mi1e Island catastrophe) 5. Design the system to be user-tolerant, to still operate reliably when an overloaded user subverts your creation (because you, the designer, usually only supply a subsystem and can’t control the user’s situation). 2 Surface to Air Missile. 3.30-104 Human Information Processing 91-2 Human Memory Capabilities Human memory is strange and non-uniform; very powerful on the one hand, and strangely limited on the other. There are two distinct memory systems: short-term memory used like a “scratchpad” for transient information, and long-term memory (what we commonly think of as memory). Short term memory is limited to about 7 items3, though this varies somewhat from individual to individual (referred to as the “magic number 7i2”). This is, for instance, why telephone numbers are limited to 7 digits in North America. Short term memory behaves much like dynamic RAM in that the contents must be refreshed or “rehearsed” lest they quickly fade. Unfortunately, rehearsing an item in short term memory tends to interfere with, and suppress, the other items; more so if the items are similar (eg. a list of random numbers is harder to maintain than a mixed list of names and numbers). Long term memory, conversely, has effectively unlimited storage capacity. It exhibits an almost holographic quality in that damage to part of thememory centre reduces the clarity of some memories, but doesn’t eliminate them (as would damage to, say, a disk file). Apart from “unlimited” capacity, long term memory is associative. By this we mean that accessing a specific memory (eg.your favorite dog), also brings forth a large number of associated information (e g. the fun times you used have along the river. Why the smell of the sycamores was so...). This is probably one of the major contributors to our ability to create and synthesize, as we create associations between memories4. Another important aspect is that memory searches are done in parallel instead of the serial searches used by computers (ie. content-addressable vs. location- addressable memory). For example, mosaof .us “never forget a face”, and it takes about the same amount of time to recognize a face now as it did 10-20 years ago, in spite of the fact that in that time, we’ve learned vastly more faces. Computers, on the other hand, take longer and longer the more faces they learn. Information from the senses goes to long term memory via short term memory5 by a process not clearly understood, but which seems related to the rehearsal process (or strong/unusual stimuli such as emergency situations or “peak experiences”). The typical memory “response curve” is shown in figure 6. Clearly, the rate at which we can absorb information accurately is controlled by short term memory’s 7i2 item limitation. One way of increasing the available bandwidth is by encoding the information. Because short term memory is limited to the number of items and not bits, we can encode more information into each item. For instance, most people dealing with computers will remember bit patterns in octal (3 bits/item) or hexadecimal (4 bits/item) instead of as a binary pattern (1 bit/item). Similarly, we group small numbers of items together so that each group occupies one “slot” in short term memory (cg. North American telephone numbers are of the form ###—####). 3 NOTE: the measure is items and not bits. . 4 Note that Apple’s HyperCardTM program mimics this organization as any card (ie. memory) can be linked to an arbitrary number of other cards in a very free-form fashion. 5 Alas, damage to the short term memory centres not only behaves like damage to a disk, but also inhibits transfer of information from short term to long term memory. ENsc-104 Human Information Processing 91-2 Such strategies can also be used with long term memory. But there are additional techniques that can be used here. One such technique is mnemonics, where the information is stored in an expanded, highly-redundant form (like the NATO spoken alphabet, itself a mnemonic), where the stored form has an easily-remembered structure, such as a simple rhyme (eg. “1 before e except after 0”). Essentially a form of mnemonics is the method of key (or peg) words, where items to be remembered are associated with (usually rhyming) words or sounds. For example, a system in use for several hundred years uses: number sound rule 0 s or z (s for cipher or z for zero) 1 t or (1 (they have 1 vertical bar) 2 11 (an n has 2 vertical bars) 3 m (an m has 3 vertical bars) 4 r (the word four ends in r) 5 L (L is the Roman numeral for 50) 6 ch or sh (no rule, just learn it) 7 k or ng (with some imagination, k can look like 7) 8 f or v (the script f looks like an 8) 9 p or b (9 looks like a twisted and rotated p or b) To remember a number, say 307, make up a key word that uses m, s/z, k/ng — mask for instance. Remember, it is the sound and not the spelling-that is important. Both short- In lum- II Item 30 lmm 4o nuns and Lon -term "gem 2mu¢h lute-sh» Imuth _ . loo h—menfory long-ter Data from Murdock ‘ memory , (1962). . so _ . E 5 20 Item = 3 I a: each 5 M g 4o 20 20 0o s to is 20 2: so as to o o 5 10 15 2o 25 30 4 Serb! warm E _ .. .: Short-term ‘3, ~ 3 memory a .g s :4 0 5 10, is 20 25 3° 0 «5" io-* is “20‘- 630 sum Pomion ‘Serial position Figure 6. As we can see from the curve, the last items in a memorized sequence are remembered most accurately. This falls off to about 20% which remains constant except for the first few items. It appears that the initial fall-off is due to short-term memory, while the remainder of the curve represents long-term memory. ENsc-104 Human Information Processing 91-2 Another Strategy, dating from the days of Greece, is the method of places, exploits long term memory’s associative properties. Here, one memorizes a series of items by visualizing them placed in sequence in familiar surroundings. To recall then, one looks at the real surroundings (or visualizes walking through the surroundings), and at each landmark along the way, the desired item becomes available through association. For example, to remember a typical shopping list of bread, eggs and butter, one might associate them with a walk from home to the store: “A loaf of bread is blocking the front door/the sailboat on the beach is filled with eggs/the railroad train is carrying a stick of butter...” Curiously, the more absurd the association, the more likely one is to remember the items. A variation on the method of places is the method of associations, where one makes the association not to a place or places, but to events in an everyday routine: “I get up in the morning and eat bread and butter with eggs on the side...” A final approach that is most useful for long buried memories is recall as problem solving. If asked the question “What were you doing on Monday afternoon in the third week of September two years ago?,” most of us would be hard-pressed to answer. However, one can again make use of the associative nature of human memory to reconstruct the events. A typical answer might look like: “Two years ago...I would be in high school in Moose Jaw...that would be my senior year. Third week in September—that’s just after summer—that would be fall term. Let me see, I think I had chemistry lab on Mondays, so I was probably in the chemistry lab. Wait a minute—that would be the second week of school. I remember he stared off with the atomic table—I thought he was crazy, trying to make us remember that thing. You know, I think I can remember sitting...” As the designer, you should reduce the short term memory load on the user by either reducing the quantity in the first place, keeping the information visible (ie. eliminating the need to remember it), or by exploiting the user’s memory strategies (eg. grouping items together). These not only help reduce the short term memory load, but helps the user learn the system as it assists the transfer from short term to long term memory. Remember too, that the expert user has incorporated more of your system into their internal model (ie. they’ve memorized more of it) and as such, you can safely impose more long term memory load on them. Ideally, you want to design the user interface in such a way that the novice user can accomplish something useful with a minimum of effort (low memory load), and subtly aid them learning more and more features until they automatically become expert users. This usually requires the system to adapt to the user, gradually changing the appearance of the system as they learn it. However, even experts forget details (because rehearsal speeds long term memory recall), so on-line help should adapt from information on most-frequently used (and hence most-needed- to-be-learned) commands for the novice to information on the least—frequently used (hand hence most-likely-to—be-forgotten) commands for the expert. S u m m a ry Firstly, divide the tasks so that the humans in the system do what they are best at—pattern recognition, creativity, synthesis, “holistic” thought—and the machine does what it is best at— accuracy, repetition, number-crunching. Another way of saying it is “match the tool to the user”. ENscm Human Information Processing 91-2 Keep the information processing requirements within human limits by eliminating unnecessary information, prioritzing the information, and providing a mechanism where the user can suppress the lower-priority information. Furthermore and especially for the subsystem designer, design the system to be user-tolerant and “fail-safe” when an overloaded user starts to subvert the information presented by your system. (Ideally, this will never occur if the rules in the first sentence are followed, but the user’s environment is often beyond your control.) Similarly, keep short term memory requirements within limits by again reducing the quantity to be remembered in the first place, keeping the information available during the user’s “scratchpad” phases, and/or by exploiting memory aids such as grouping and encoding to reduce the short term memory “item count” limitation. Your system should pace and assist the novice and the expert differently, and should be maximally useable throughout the training period. Keep it simple. Keep it consistent. Appendix A: Information Theory Shannon at Bell Labs developed the field of information theory, with the goal of maximizing the efficiency of the telephone/telegraph channels. He defined information as the amount of uncertainty dispelled. For instance, given a vocabulary or set of messages/symbols (eg. the alphabet A. . .2) that can be transmitted down the channel, the larger the vocabulary, the less likely it is that the ith symbol will be received. Thus, as the vocabulary size increases, receiving the ith symbol becomes more significant. In other words, the ith symbol has more information. For memoryless channels (ie. where the reception of one symbol has no effect on the probability of which symbol would be received next), the information content is 1M) = F(pi) = '-k 10ngi Changing k and b let us choose the units. The most common is 1 and 2 respectively, giving the result in bits5. More compactly (and using bits as the unit), the information of any event (M1) can be expressed as: hi =log2(1/pi) For any reasonably long sequence (N) of a vocabulary of (M) symbols, one would expect the ith symbol (ie. Mi), which has probability (pi) to occur Npi times. The average information or entropy of the source would then be the sum of the information content of the N symbols transmitted, divided by N. Thus, entropy (H), for memoryless channels, can be calculated as: H=1/N(h1p1+h2p2+...hipi+...hmpm) H = 2(pi hpi) H = 2(pi log2(1/pi)) . (since hi = log2(l/pi)) H = ‘Zpi log2(pi) (since log( l/a) = -log(a)) 6 Note: logb(N) = log(N)/log(b) 10 ENsc-104 Human Information Processing 91-2 For example, a coin toss has an equal chance of coming up heads or tails. Thus, the entropy of a tossed coin is: H = 'zp110g2(pi) H = (0.5 * log2(0.5)) + (0.5 *10g2(0.5)) H = 1 bit If, on the other hand, the coin is slightly bent, giving it a probability of 90% of coming up heads, and 10% of coming up tails, the entropy becomes: Given: p(heads) = 0.9, p(tails) = 0.1, then H = (0.9 * log2(0.9)) + (0.1 * log2(0.1)) H = (0.9 * 0.15) + (0.1 * 3.32) H = 0.135 + 0.332 H = 0.47 bits 11 ...
View Full Document

Page1 / 11

HIPReading2 - 5mm Human Information Processing 91-2 The...

This preview shows document pages 1 - 11. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online