Exam August 2023 Flashcards
What are the advantages and disadvantages of using natural language as a basis for a knowledge representation system?
Advantages: A way for humans and machine to communicate with each other. A way to avoid
using symbols and can help us get closer to achieve cognition in machine.
Disadvantages: Cases where machines can have problems interpreting humans speech when
it tries to anaylze it. Another one is when a machine misinterpret commands given by humans
and show bad results.
What were Logic Theorist and General Problem Solver, and why were they significant?
Logic Theorist:
Developed Newell, Shaw, and Simon in 1955, it was the first artificial intelligence program. Demonstrated that machines could prove mathematical theorems, showcasing early AI capabilities.
General Problem Solver (GPS):
Developed by Newell and Simon in 1957, it was a more generalized problem-solving program. Showed the potential for creating universal problem-solving systems, laying the foundation for future AI research.
Why was Hebb’s learning rule important to the initial development of Artificial Neural Network?
Hebb’s learning rule was pivotal for ANNs as it introduced a fundamental concept of synaptic plasticity, influencing the development of learning algorithms and contributing to the understanding of how artificial neurons could adapt and learn from data.
Today, it seems that LLMs are on the verge of passing the Turing test. Does it mean that today’s AI systems have achieved human cognitive ability? Reasoning about whether we should keep the Turing test as a yardstick or if the criteria need to be updated.
No I don’t think it reached humans level intelligence. I think it can imitate us very well, or even
better, in terms of computation and memorization. I think Turing test is outdated and therefore
should be updated if we still think it is possible for an AI to reach human level of cognition.
a. In general, explain what it means when we say that a search strategy is optimal.
b. In general, explain what it means when we say that a search strategy is complete.
c. Name the two uninformed search strategies that would be complete and optimal and could
be used to find the shortest path for this problem.
a. when it reaches the goal with the least amount of resources used.
b. it guarantees that there is a result, if there is any.
c. breadth first search, Dijkstra’s algorithm.
In the lecture we have discussed that “Decision Support Systems (DSS) are categorized by how
they carry out their actions. Five categories have been discussed. One was “Communication-driven DSS”. (2p)
a. List the other four types
b. Describe the major differences between these types of DSS categories
a. Data-driven, document-driven, model-driven, knowledge-driven.
b. Model-Driven DSS: Utilizes mathematical models for decision-making.
Data-Driven DSS: Relies on large volumes of data for decision support.
Document-Driven DSS: Emphasizes the retrieval and management of unstructured information.
Knowledge-Driven DSS: Incorporates expert knowledge to support decision-making.
Communication-Driven DSS: Facilitates communication and collaboration among decision-makers.
“The environment where agents operate affects the agents. Hence, the agents must be
programmed to handle the task environment.”
In the lecture we discussed several environmental characteristics.
List and describe the four characteristics of the environment in which an intelligent agent
operates (2p)
Observable: where the environment is fully realized and it knows it’s surrounding well.
Static: when things the environment remain unchanged.
Deterministic: when the goal is known and it works towards it.
Episodic: when the actions of agent in the current state is indepedent of its previous ones.
vs.:
partially observable: when the knowledge of the environment is limited and it requires
information from the outside to learn the surrounding.
dynamic: when things in the environment keep changing
stochastic: when there is no set goal and things are random
non-episodic: when actions are sequential and depend on each other.
What is a “Decision Support System” (DSS) and an expert system? What parts/components have these systems? What are the differences between these systems?
DSS: is information system that supports organizational decision-making activities. It consists of Database, knowledge base and UI. Broad support for decision-making, not domain-specific.
An expert system: is a system that emulates the decision-making ability of a human expert in a specific domain. consists of Knowledge base, inference engine, user interface, explanation module. Specialized in a specific domain, providing expert-level decision-making capabilities.
FAQ (or Frequently Asked Questions) is a collection of frequently asked questions and their answers.
Consider an FAQ agent tasked with providing information on common questions or problems.
Describe the FAQ agent as “a goal‐based agent”.
Goal: Providing information to user
Perception: Receives user questions.
Knowledge Base: Contains common questions and their answers.
Decision-Making: select the most relevant answer based on the user’s input.
Action: Outputs the selected answer
Learning: Can be enhanced by updating the knowledge base with new questions
Knowledge-based agents consist of two main components: Knowledge base and inference system. Describe these two components: their definition, purpose and differences.
Knowledge Base:
Stores domain-specific information. Acts as the memory of the system, holding facts and rules.
Inference System:
Draws conclusions from the knowledge base. Processes information to make decisions or solve problems.
Differences:
KB: Static, containing encoded knowledge.
IS: Dynamic, actively using knowledge for reasoning.
Natural Language Processing
Explain the following concepts using the examples below (2p):
1. Lexical analysis
2. Syntactic analysis
3. Semantic analysis
4. Pragmatic analysis
Lexical: analyising words and understand their meaning.
Syntactic: analysing syntactic rules and grammar and see whether these rules lead into
meaningful sentence
Semantic: analysing whether NP, VP, CV have semantic meaning
Pragmatic: analysing whether the sentence above give its true meaning in that situation
without misunderstanding its meaning.
Discuss the impact of learning rate on the training process of NNs and identify the optimal learning rate and its possible range of values.
Learning rate significantly influences neural network training. On rapid base it might lead to overshooting the minimum. Too slow leads to impractical training time. There’s no one-size-fits-all optimal learning rate; it requires experimentation. Considerations like decay, monitoring performance, and using adaptive methods contribute to finding an effective learning rate. Range: Typically between 0.1 and 0.0001.
Suppose you are training a model for a classification problem, using only three inputs. Determine
suitable inputs for differentiating between chairs and tables, plus describe what would be a good
input for a text classification problem (for e.g., political vs economic)
Input 1: Does the object have 4 legs?
Input 2: Does the object have a rectangular piece on top of the 4 legs?
Input 3: Does the object have a backrest?
You can manage this to decide whether the object is a chair or a table with the help of weighted nodes in the ANN. It’s hard to choose just 3 inputs as a table can be round, a square, a rectangle and what not so I decided to limit it to the most basic tables where there are 4 corners and usually a rectangular pattern. I’ve also used my most basic view of a chair where there is 4 legs and a backrest.
For the text classification problem you need to choose specific keywords and if it contains those keywords you can make the nodes fire in regards to how strongly the words are associated to the specific subject.
Explain the validation process of a NN, focusing on the accuracy curve and error curve.
Elaborate on the concept of overfitting and how it relates to number of epochs in NNs.
An NN trains via errors and can use different learning algorithms, one of the most common ones are backpropagation where you go from the top to the bottom and finding where the “error” exists and trains the NN to “do better” with the data it is handling. Overfitting can be a big issue when you’ve trained the NN too well and it has a 0% error curve and no loss of data.
This means that it usually cannot handle new data as it only uses it’s knowledge of the known data and can therefore make mistakes with new data being put in to the NN.
Compare and contrast three types of Deep Neural Networks (DNN), highlight their key difference.
- Convolutional Neural Networks (CNNs):
Focus: Image-related tasks with convolutional layers.
Key Feature: Parameter sharing for spatial hierarchies.
2. Recurrent Neural Networks (RNNs):
Focus: Sequential data processing with recurrent connections.
Key Feature: Temporal dependencies for handling sequences.
3. Generative Adversarial Networks (GANs):
Focus: Data generation through adversarial training.
Key Feature: Generator and discriminator networks for adversarial learning.
Contrast:
Data Processing:
CNNs and RNNs: Process different data types (grid-structured vs. sequential).
GANs: Generate new data rather than processing input.
Training Paradigm:
CNNs and RNNs: Supervised training for classifications.
GANs: Adversarial learning between generator and discriminator.
Output:
CNNs and RNNs: Produce classifications or predictions.
GANs: Generate new, synthetic data samples.
