Task 4 Flashcards
What is connectionist research?
A field that models how neural networks contribute to thought, emphasizing connections among simple neuron-like structures.
What are two alternative names for connectionist research?
Neural networks and Parallel Distributed Processing (PDP).
What are the two main classes of connectionist models?
Local representations – Each neuron-like unit represents a specific concept or proposition.
Distributed representations – Concepts are distributed across multiple neuron-like units.
How do connectionist networks perform parallel constraint satisfaction?
By adjusting activation levels across many units simultaneously to find a stable, consistent state.
What are the basic components of a connectionist network?
Units – Neuron-like components that activate based on input.
Links – Connections between units, which can be excitatory (positive) or inhibitory (negative).
What is the difference between one-way and symmetric links?
One-way links – Activation flows in a single direction.
Symmetric links – Activation flows back and forth between two units.
How are concepts represented in local vs. distributed networks?
Local networks – Individual units correspond to specific concepts (e.g., “computer geek”).
Distributed networks – Concepts are spread across multiple units, making the network more flexible and robust.
What is a recurrent network?
A network where activation from output units feeds back into input units, creating cyclical processing.
How do neural networks solve problems?
Through spreading activation – units pass signals to connected units until the network settles into a stable state.
What is the concept of “relaxation” in neural networks?
The process of adjusting activation across all units until they reach a stable, consistent state.
How do connectionist models handle decision-making?
They balance positive constraints (actions/goals that support each other) and negative constraints (conflicting actions/goals).
What is an example of a real-world constraint satisfaction problem?
Scheduling university classes while avoiding conflicts with rooms, professors, and student availability.
What are the two main ways learning occurs in neural networks?
Adding new units to the network.
Changing the weights on links between units.
What is Hebbian learning?
A rule stating that “neurons that fire together, wire together”, meaning that connections between co-activated neurons strengthen over time.
Why is Hebbian learning considered unsupervised?
The network learns without a teacher, simply by reinforcing frequent co-activations.
What is backpropagation, and how does it work?
A supervised learning algorithm where errors are propagated backward through the network to adjust connection weights.
What are some limitations of backpropagation as a model of human learning?
It requires a supervisor (feedback on right/wrong answers).
It is slow, needing hundreds or thousands of examples to learn patterns.
What is pattern association in neural networks?
The process of learning to associate one input pattern with a specific output pattern (e.g., associating a word with its meaning).
What is autoassociation?
A type of pattern association where the input and output patterns are identical, helping with memory recall.
How does Hebbian learning apply to pattern association?
Connections between simultaneously active input and output units are strengthened, making recall more efficient.
What is competitive learning?
An unsupervised learning process where output units compete, and only the most active unit strengthens its connections.
What are the three phases of competitive learning?
Excitation – Input activates multiple output units.
Competition – Output units inhibit each other, with the strongest unit winning.
Weight adjustment – The winner strengthens its connection to the input, improving future recognition.
What happens if a competitive network lacks proper weight control?
One unit may become dominant, preventing other units from learning new patterns (a “winner-takes-all” effect).
What is a Brain-Computer Interface (BCI)?
A system that translates brain activity into computer commands, allowing direct brain control of external devices.
What are the two main types of BCIs?
Invasive BCIs – Electrodes are surgically implanted into the brain.
Noninvasive BCIs – Brain signals are detected from outside the scalp (e.g., EEG-based systems).
What are some applications of BCIs?
Assistive BCIs – Help people with paralysis communicate and control devices.
Rehabilitative BCIs – Aid in stroke recovery and motor function restoration.
What is graceful degradation, and how does it apply to the brain?
A gradual decline in function when parts of a system are damaged, unlike catastrophic failure in traditional computers.
How does human memory retrieval differ from traditional databases?
The brain uses content-addressable memory, meaning related information is automatically activated rather than searched sequentially.
What is parallel processing, and how does it support generalization?
The brain processes multiple stimuli at once, allowing it to recognize patterns and generalize across different experiences.
What are the main advantages of connectionist models?
They can learn from experience.
They model human memory and perception effectively.
They provide robust performance, even with incomplete or noisy data.
What are some real-world applications of neural networks?
Speech recognition
Image classification
Medical diagnosis
Cognitive modeling in AI
What is parallel constraint satisfaction?
A process where neural networks adjust activation levels until multiple constraints are simultaneously satisfied.
How does a constraint satisfaction network handle decision-making?
It spreads activation through excitation and inhibition until a stable choice is reached.
What is an example of a constraint satisfaction problem (CSP)?
University scheduling, where courses, professors, and rooms must be arranged without conflicts.
How do internal and external constraints shape decision-making?
Internal constraints – Relationships between goals and actions (e.g., studying helps pass exams).
External constraints – Prioritization factors (e.g., passing exams is more important than partying).
What is a goal priority unit in a constraint satisfaction network?
A special unit that amplifies activation for more important goals, influencing final decisions.
How does autoassociation differ from pattern association?
Autoassociation stores a pattern and recalls it when given a partial input, like memory recall.
What is pattern association in neural networks?
A process where the network learns to associate an input pattern with a specific output pattern.
What are the three phases of competitive learning?
Excitation – All output units receive activation.
Competition – Units inhibit each other until only one remains active.
Weight adjustment – The winning unit strengthens its connections to the input.
What are the five types of brain signals detected by BCIs?
Slow Cortical Potentials (SCPs) – Gradual voltage shifts over several seconds.
Sensorimotor Rhythms (SMRs) – Brain waves linked to motor control.
P300 Event-Related Potentials (ERPs) – A brain spike 300ms after a new stimulus.
Steady-State Visual Evoked Potentials (SSVEPs) – Brain response to flashing lights.
Error-Related Negativity (ERNs) – Signals when the brain detects mistakes.
What is the most commonly used BCI signal for communication?
The P300 potential, because it allows users to select items by focusing on them.
