Wednesday, September 23, 2009

Terry Winograd's Shift from AI to HCI

In a more recent paper [1], Terry Winograd discusses the gulf between Artificial Intelligence and Human-Computer Interaction. He mentions that AI is primarily concerned with replicating the human / human mind whereas HCI is primarily concerned with augmenting human capabilities. One question is wether or not we should use AI as a metaphor when constructing human interfaces to computers. Using the AI metaphor, the goal is for the user to attribute human characteristics to the interface, and communicate with it just as if it were a human. There is also a divide in how researches attempt to understand people. The first approach, what Winograd refers to as, the “rationalistic” approach, attempts to model humans as cognitive machines within the workings of the computer. In contrast, the second approach, the “design” approach, focuses on modeling the interactions between a person and the enveloping environment.

During his career, Winograd shifted interests and crossed the gulf between AI and HCI. In his paper, he mentions that he started his career in the AI field, then rejected the AI approach, and subsequently ended up moving to the field of HCI. He writes “I have seen this as a battle between two competing philosophies of what is most effective to do with computers”. This paper looks at some of the work Winograd has done, and illustrates his shift between the two areas.

Winograd's Ph.D. Thesis and SHRDLU
In his Ph.D. thesis entitled “Procedures as a Representation for Data in a Computer Program for Understanding Natural Language” [2], Winograd describes a software system called SHRDLU that is capable of carrying on a English conversation with its user. The system contains a simulation of a robotic arm that can rearrange colored blocks within its environment. The user enters into discourse with the system and can instruct the arm to pick up, drop, and move objects. A “heuristic understander” is used by the software to infer what each command sentence means. Linguistic information about the sentence, information from other parts of the discourse, and general information are used to interpret the commands. Furthermore, the software asks for clarification from the user if it cannot understand what the inputted sentence means.

The thesis examines the issue of talking with computers. Winograd underscores the idea that it is hard for computers and human to communicate since computers communicate in their own terms; The means of communication is not natural for the human user. More importantly, computers aren't able to use reasoning in an attempt to understand ambiguity in natural language. Computers are typically only supplied with syntactical rules and do not use semantic knowledge to understand meaning. To solve this problem, Winograd suggests giving computers the ability to use more knowledge. Computers need to have knowledge of the subject they are discussing and they must be able to assemble facts in such a way so that they can understand a sentence and respond to it. In SHRDLU knowledge is represented in a structured manner and uses a language that facilitates teaching the system about new subject domains.

SHRDLU is a rationalistic attempt to model how the human mind works. It seeks to replicate human understanding of natural language. Although this work is grounded in AI, there a clear implications for work in HCI. Interfaces that communicate naturally with their users are very familiar and have little to no learning curve. Donald Norman provides several examples of natural interfaces in his book “The Design of Future Things” [3]. One example that stands out is a tea kettle whistle. The tea kettle whistle offers natural communication that water is boiling. The user does not need to translate the sound of a tea kettle whistle from system terms into something he/she understands; it already naturally offers the affordance that the water is ready.

Thinking machines: Can there be? Are We?
In “Thinking Machines” [4], Winograd aligns his prospects for artificial intelligence with those of AI critics. The critics argue that a thinking machine is a contradiction in terms. “Computers with their cold logic, can never be creative or insightful or possess real judgement”. He asserts that the philosophy that has guided artificial intelligence research lacks depth and is a “patchwork” of rationalism and logical empiricism. The technology used in conducting artificial intelligence research is not to blame, it is the under-netting and basic tenets that require scrutiny.

Winograd supports his argument by identifying some fundamental problems inherent in AI. He discusses gaps of anticipation where in a any realistically sized domain, it is near impossible to think of all situations, and combinations of events from those situations. The hope is that the body of knowledge built into the cognitive agent will be broad enough to contain the relevant general knowledge needed for success. In most cases, the body of knowledge contributed by the human element is required; since it cannot be modeled exhaustively within the system. He also writes on the blindness of representation. This is in regards to language and interpretation of language. As expounded upon in his Ph.D. thesis, natural language processing goes far beyond grammatical and syntactic rules. The ambiguity of natural language requires a deep understanding of the subject matter as well as the context. When we de-contextualize symbols (representations) they become ambiguous and can be interpreted in varying ways. Finally, he discusses the idea of domain restriction. Since there is a chance of ambiguity in representations, AI programs must be relegated to very restricted domains. Most domains, or at least the domains the AI hopes to model, are not restricted (e.g. - medicine, engineering, law). The corollary is that AI systems can give expected results only in simplified domains.

Thinking Machines offers some interesting and compelling arguments against the “rationalist approach”. It supports the idea that augmenting human capabilities is far more feasible than attempting to model human intelligence. This is inline with the “design approach” (i.e. - placing focus on modeling interactions between the person and his/her surrounding environment.)

Stanford Human-Computer Interaction Group
Winograd currently heads up Stanford's Human-Computer Interaction Group [5]. They are working on some interesting projects grounded in design. One such project, is a hardware and software toolkit that allows designers to rapidly prototype physical interaction design. Designers can use the physical components (controllers, output devices) and the accompanying software (called the editor) to form prototypes and study their behavior. Another project, named Blueprint, integrates program examples into the development environment. Program examples are brought into the IDE through an built-in web search. The main idea behind Blueprint is that it helps to facilitate the prototyping and ideation process by allowing programmers to quickly build and compare competing designs. (more information on these projects can be found on their website (link is below))

[1] Terry Winograd, Shifting viewpoints: Artificial intelligence and human–computer interaction, Artificial Intelligence 170 (2006) 1256–1258.

[2] Winograd, Terry (1971), "Procedures as a Representation for Data in a Computer Program for Understanding Natural Language," MAC-TR-84, MIT Project MAC, 1971.

[3] Norman, Donald (2001), The Design of Future Things, New York: Basic Books, 2007.

[4] Winograd, Terry (1991), "Thinking machines: Can there be? Are We?," in James Sheehan and Morton Sosna, eds., The Boundaries of Humanity: Humans, Animals, Machines, Berkeley: University of California Press, 1991 pp. 198-223. Reprinted in D. Partridge and Y. Wilks, The Foundations of Artificial Intelligence, Cambridge: Cambridge Univ. Press, 1990, pp. 167-189.

[5] The Stanford Human-Computer Interaction Group

No comments: