computation and cognition review (non-exhaustive)

Why is it difficult to fully remove bias from AI systems by simply cleaning the training data?

• Bias is introduced during testing

• AI struggles with cleaned data

• Some biases are subtle or baked into patterns

• AI relies on randomness, not structure

You're designing a robot that needs to navigate a 3D maze. Which component of Dreamer AI would help it mentally simulate possible paths before moving?

• Actor Network

• Critic Network

• World Model

• Feedback Loop

What is the purpose of self-supervised pre-training?

• It happens before any training takes place

• It prepares the model with general language skills before task-specific training

• It uses labeled data to train the model on specific prompts

• It fine-tunes the model’s tone using human feedback

How does DeepMind's use of reinforcement learning differ from early AI approaches?

• Early AI focused more on learning from reward signals

• DeepMind uses manual heuristics created by humans

• Early AI relied on hand-coded rules, and DeepMind’s AI learns from experience

• There is no real difference between the two approaches

Which statement explains how pattern completion in Hopfield networks can also lead to spurious memories?

• Hopfield networks can only recall exact matches, so spurious memories are rare

• Hopfield networks fill in gaps using nearest stored patterns, but may overgeneralize and generate incorrect outputs

• Spurious memories arise only when the network has no stored patterns

• Pattern completion is unrelated to spurious memory formation in Hopfield networks

What is the main purpose of the Perceptron algorithm?

• To group similar data points based on patterns

• To learn how to separate labeled data into categories

• To generate new patterns based on input examples

• To store and retrieve learned patterns from memory

If a neural system is using prediction to complete missing parts of a memory from noisy input, which type of generative model is it most like?

• Transformer

• Hopfield Network

• Simple Recurrent Network

• Reinforcement Learning Agent

How do Transformer Networks handle memory differently from SRNs?

• Transformers retain past information through internal states

• Transformers rebuild memory by adding each output token to the input

• Transformers predict entire sequences in one step

• Transformers store sequences in a fixed attractor pattern

In what ways did the General Problem Solver differ from AlphaGo? Select all that apply then submit your answer.

• It was designed to model human-like thinking, including errors

• It learned by analyzing millions of examples

• It emphasized rule-based operations rather than deep learning

• It focused on simulating problem-solving steps over performance

Which of the following best illustrates Marr’s computational level in AlphaGo?

• The specific rules used by AlphaGo to play Go

• AlphaGo's use of artificial neural networks

• The goal of winning the game by maximizing reward

• The use of hardware to process data