Exam 2021 and resit Flashcards
Difference between cognition and intelligence
Cognition is a
- global process at the system
- that integrates many different processing modalities.
Special cognitive skills such as intelligence, learning, memory etc are constituents and synergies of a cognitive system.
Difference btw. bio-inspired and purely computational models
Bio-inspired models: Implement cognitive functions by replicating known or hypothesized mechanisms of cognitive processing from biological organisms.
Computational models: Implement cognitive functions based on a functional view of the system without any reference to biology.
Name and explain 2 Properties of dynamical systems
Dissipation: the number of reachable states reduces over time
Non-equilibrium system: stable functions require external energy supply
Non-linearity: Complex behavior can emerge from a small set of state parameters
Collective variables: the system is represented by a small set of state variables.
3 types of cognitive architecture
symbolic, emergent, hybrid
Explain how ACT-R implements parallel information processing and how it implements serial information processing
Parallel Information Processing:
- information processing within modules can be parallel to support high-dimensional data streams
- different modules process data in parallel independently from each other
Serial Information Processing.
- every module buffer can store only a single chunk
- in each cycle of the production system exactly one rule fires
Describe the buffer test and buffer action of the following production rule:
(p type =goal> isa goal state enter-number ==> \+Manual> CMD PRESS-KEY KEY "2" )
Test if the goal buffer holds a chunk of type goal and if the state is enter-number then send a buffer request to the manual module to press to key 2
Explain the purpose of MapSpikeSource and MapSpikeSink devices of a transfer function in the NR platform
MapSpikeSink: reads out spikes from the SNN and coverts them into float values
MapSpikeSource: Creates spikes from a float value that is fed into the SNN
Names of the three main components of biological neuron
Soma, dendrites, axon
2 visual processing streams and which information is computed along them
Dorsal stream: About the location of the object (Where?) In a more recent view also about how motor functions take place (How?)
Ventral stream: About what the identity of the item ist (What?)
2 main schemes for encoding SNN and the main advantages
rate encoding: robustness against noise
time encoding: fast reaction time
State the basic principle of Hebbian learning
When axon of cell A is near enough to cell B and persistently or repeatedly takes part in firing to it then some growth process takes place in one or both cells so that A’s efficiency is increased.
Fundamental properties of Hebbian learning?
Saturation: Ovoid unbounded growth of synaptic weights
Competition-> Selection: Avoid weights to converge to the same value
Locality: weights changes dependent on local variables
Describe the statistical process that takes place when Oja’s rule is applied to a dataset
- maximize the variance of the neuron’s output
Name and describe 3 learning paradigms
- unsupervised learning: learning of statistical regularities in the input data with no labels
- supervised learning:
learn the mapping between input and output data with predefined labels - reinforcement learning: learning which action to chose in a certain situation in order to maximize an external reward signal through interaction with the environment.
return
sum of future discounted rewards in a Markov reward process
action policy
a function that describes the probability for selecting a specific action conditioned by a state pi(a|s)
state-value function
State value is the expected cumulated discounted reward following a specific policy when being a specific state
reward
numerical signal that implicitly expresses the agent’s goal by encouraging/punishing goal directed/undirected state transitions.
action-value function
q is the expected cumulative and discounted reward following a specific policy when selecting a specific action in a particular state.
approximate RL
the agent predicts values with the help of non-linear function approximators.
Dynamic Programming (bootstrapping? sampling?)
bootstraps, no sampling
-> high width of updating, low depth
Monte Carlo
sampling, no bootstrapping
low width, high depth
Temporal Difference
sampling, bootstrapping
- low width, low depth
Explain GPI
Almost every RL algorithm implements the mechanisms of two alternating steps to evaluate an optimal policy
- policy evaluation: evaluation of a value function depending on a policy
- policy improvement: improving the policy by acting greedily on the value function
Both depend on each other so both steps need to be performed iteratively until an optimal policy is found.
