CIS 5603. Artificial Intelligence

Summary of AI

In this final lecture, we return to the broader picture of AI and revisit the four basic questions introduced in the first lecture.

1. The objective of AI

AI is ultimately inspired by Human Intelligence (HI), though it is impossible for the two to be identical in all aspects, due to differences in their "bodies" and "histories". While stressing the similarity between AI and HI, Turing's question "May not machines carry out something which ought to be described as thinking but which is very different from what a man does?" has been widely ignored (by himself, too). On which aspecct, and how much, must AI be similar to human intelligence to count as intelligence?

In general, the same object or process can be abstracted to different levels. Various levels of abstraction of HI (Structure, Behavior, Capability, Function, Principle) define different conceptions of AI, which have different objectives and consequences:

focus on achievements vs. constraints
evaluate performance vs. potential
object-level ability vs. meta-level ability

These objectives cannot replace one another. Many confusions and controversies in AI are caused by mixing them.

2. The potential and limitations of AI

The question "Can AI be achieved?" has no simple "yes/no" answer.

According to different definitions of "intelligence", AI can be "already achieved", "seems impossible", or "still hard to say".
Intelligence is widely taken as a matter of degree, and every objective of AI can be approached gradually.
The conclusion on what AI "can do" should distinguish the specific problem instances and the whole problem type, as well as between "in principle" and "in practice".
The conclusion on what AI "cannot do" should distinguish "an AI model", "an AI technique", and "the AI objective".

Examples of analysis: Three Fundamental Misconceptions of Artificial Intelligence, Emergent Abilities in Large Language Models: A Survey.

3. How to achieve intelligence

The following is a brief comparison of the two major competing schools of AI (generated by an LLM):

Aspect Symbolic (Logical) AI Connectionist (Statistical) AI

Knowledge Representation Uses explicit symbols, facts, rules, and logical relationships. Uses distributed numerical representations encoded in weights and activations.

Learning Relies primarily on human-crafted knowledge and rules. Learns automatically from data through training.

Reasoning Excels at logical reasoning, deduction, and planning. Excels at pattern recognition but provides limited explicit reasoning.

Interpretability Highly transparent and explainable. Often difficult to interpret and commonly viewed as a "black box."

Data Requirements Can operate effectively with relatively small amounts of data. Typically requires large datasets for effective training.

Generalization Limited outside the scope of encoded rules and knowledge. Often generalizes well from training examples to unseen data.

Robustness to Noise Can be sensitive to incomplete, uncertain, or noisy inputs. Generally more tolerant of noisy or imperfect data.

Knowledge Acquisition Requires extensive knowledge engineering by experts. Acquires knowledge automatically during training.

Computational Cost Inference is often computationally efficient. Training can be computationally expensive and resource-intensive.

Typical Applications Expert systems, planning systems, theorem proving, and knowledge-based systems. Computer vision, speech recognition, natural language processing, and recommendation systems.

Aspect	Symbolic (Logical) AI	Connectionist (Statistical) AI
Knowledge Representation	Uses explicit symbols, facts, rules, and logical relationships.	Uses distributed numerical representations encoded in weights and activations.
Learning	Relies primarily on human-crafted knowledge and rules.	Learns automatically from data through training.
Reasoning	Excels at logical reasoning, deduction, and planning.	Excels at pattern recognition but provides limited explicit reasoning.
Interpretability	Highly transparent and explainable.	Often difficult to interpret and commonly viewed as a "black box."
Data Requirements	Can operate effectively with relatively small amounts of data.	Typically requires large datasets for effective training.
Generalization	Limited outside the scope of encoded rules and knowledge.	Often generalizes well from training examples to unseen data.
Robustness to Noise	Can be sensitive to incomplete, uncertain, or noisy inputs.	Generally more tolerant of noisy or imperfect data.
Knowledge Acquisition	Requires extensive knowledge engineering by experts.	Acquires knowledge automatically during training.
Computational Cost	Inference is often computationally efficient.	Training can be computationally expensive and resource-intensive.
Typical Applications	Expert systems, planning systems, theorem proving, and knowledge-based systems.	Computer vision, speech recognition, natural language processing, and recommendation systems.

To achieve a "complete" or "general-purpose" AI (AGI), the following strategies have been proposed:

Hybrid: To develop individual functions first (using different theories and techniques), then to connect them together. [(AA)AI: More than the Sum of Its Parts]
Integrated: To design a cognitive architecture first, then to design its modules (using various techniques) accordingly. [40 Years of Cognitive Architectures: Core Cognitive Abilities and Practical Applications]
Unified: Using a single technique to start from a core system, then to extend and augment it incrementally. [Representative AGI Projects]

4. Social impacts and ethical issues

The ethics and social impacts of AI have been concerns from the very beginning of this research. With the remarkable progress made in recent years, this topic has attracted greater public attention. Besides the issues shared with other technologies, AI has special issues of its own, such as "autonomy" and "superiority".

Sample discoveries and discussions:

Theoretical issues in these discussions:

Different senses of "AI" correspond to very different situations on this matter. For instance, safety can be considered either as an issue of design (including training) or as an issue of (life-long) education and socialization.
It is necessary to balance among multiple values/goals. Pursuing any single goal at the expense of all others is a sure path to disaster.
No one can foresee all the consequences of a decision, but that should not be used as the excuse of ignoring known evidence. Decisions should be based on comparing positive and negative evidence, rather than merely on possibilities.

All of these issues demand long-term study and exploration. Like all other decisions, people's choices ultimately depend on their assessment of benefits and risks.

Readings

Poole and Mackworth: Chapters 18, 19
Russell and Norvig: Chapters 27, 28
Luger: Chapter 16