Student Exam questions

Question 1

cognitivist cognitive systems

Answer 1

Cognitivist cognitive systems are based on the hypothesis that cognition is a form of computation. Cognitive functions are modelled as working computer programs.

Question 2

emergent cognitive models

Answer 2

In emergent cognitive models, cognition is a continuous self-organisation process that is driven by the interaction between the agent and its environment.

Question 3

Name seven cognitive capabilities that a system must possess in order to be self-reliant and adaptive, and interact with its environment

Answer 3

Self-reliant
goal-directed
autonomous
interact with other agents

Perception & Action

interpretation
sensing
anticipation
adaption

Adaption
reaction
learning
anomaly detection

Question 4

Name the four different lobes in the cerebrum and explain what their main cognitive

Answer 4

Frontal lobe: short-term memory, action planning, movement control
Parietal lobe: somatic sensation, body image
Temporal lobe: hearing, learning, memory, emotions
Occipital lobe: vision

Question 5

Assume that you have a doll company.

Answer 5

Uncanny valley phenomenon states that a steady increase in human-likeness does not yield a steady increase in the familiarity to humans.

Question 6

10. Define what a cognitive system is.

Answer 6

A cognitive system is an autonomous system that can perceive its environment, learn from experience, anticipate the outcome of events, has goal-oriented behaviour and can adapt to changing circumstances.

Question 7

Define cognition

Answer 7

Cognition is the process of knowing and includes perception and judgement. It includes all processes of consciousness by which knowledge is accumulated like perceiving, recognizing. It is the experience of knowing and different from feeling or willing. It refers to the mind and the body.

Question 8

components of a cognitive system

Answer 8

Environment (situatedness)
Body (embodiment)
Brain (constraints)
Other Agents (interaction)

Question 9

cycle of cognitive processing

Answer 9

perception -> cognition (anticipation, assimilation, adaption) -> action

Question 10

Explain what the neurorobotics platform is and explain the advantages of its usage

Answer 10

The neurorobotics platform is a multidisciplinary approach and tries to unify neuroscience, robotics, and AI. You have simulated robots in a virtual environment controlled by NN. The advantage is that you don’t have to have a physical robot and you can thus save time and money.

Question 11

Name two software tools that are integrated in the neurorobotics platform

Answer 11

Gazebo: body simulation for robots in virtual environments
OpenSim: modelling and simulation of movement

Question 12

Explain the function of buffers in ACT-R

Answer 12

Information about the current state of the module is exposed through named buffers. Every buffer stores one chunk and is assigned to exactly one module. Buffers serve as interfaces between modules. Buffers directly process queries about their contents.

Question 13

Explain a module in ACT-R. Name two features of modules in ACT-R.

Answer 13

Modules represent independent parallel processing units that encapsulate specific functions:
every module corresponds to a cognitive function in the brain
information processing is independent from other modules

Question 14

Explain what production is in ACT-R

Answer 14

Production is a statement of a particular contingency that control behavior

Question 15

Explain what the main purpose of the white matter in the cerebrum is

Answer 15

White matter essentially functions in affecting learning and brain functions, modulating the distribution of action potential, and coordinating communication between the different brain regions.

Question 16

Name and describe the anatomy of the surface of the cerebral cortex.

Answer 16

Gyri = crests formed by the convoluted surface of the cerebral cortex
Sulci = fissures between two neighbouring gyri
lobes = on each hemisphere there are two major sulci which divide the cerebral surface into the frontal, parietal, temporal and occipital lobe.

Question 17

Name two distinctive features of human brain that makes it stand out among other species

Answer 17

neuron morphology: the pyramidal neurons in the human cortex have the most elaborate and spine-rich dendritic trees
Electrical properties of neurons: pyramidal neurons have a low membrane capacity, which enhances signal transmission.

Question 18

Explain what the Blue Brain Project

Answer 18

Whole brain model that developed a data-driven reconstruction process for cortical microcircuits.
The digital reconstruction is based on experimental data and predictions.

Question 19

Name and describe the two subsystems of the vertebrate nervous system

Answer 19

Central Nervous System: central information processing system formed by the brain and the spinal cord. It collects and distributes data throughout the body
Peripheral Nervous System: transmits signals between sensory organs, muscles, and internal organs and the CNS

Question 20

Name two features or properties of the peripheral (autonomic) nervous system

Answer 20

The Peripheral Nervous System is responsible to self-regulation and operates unconsciously. It includes two antagonistic subsystems:

Sympathetic nervous system:
Prepares the organism for stress
Parasympathetic nervous system:
sets the body to resting state and increases digestive functions

Question 21

Map the nervous system into the perception-cognition-action diagram

Answer 21

Cognition with sub-functions are part of the brain. Perception are sensory organs and action is spinal cord and muscles.

Question 22

Explain the main purpose of the brainstem. Name the four sections that compose it.

Answer 22

The brainstem connects the cerebrum with the spinal cord.

Four sections:
Diencephalon
Midbrain
Pons
Medulla Oblongata

Question 23

Name three learning rules based on Hebbian learning

Answer 23

Basic Hebb’s rule
Correlation-based rule
Covariance-based rule

Question 24

Name five levels of modelling the dynamics of biological neurons

Answer 24

Detailed-compartment model
Reduced-compartment model
Single-compartment model
Cascade Models
Black-box-models

Question 25

describe Hodgkin-Huxley Model

Answer 25

• model contains a single compartment and is the basis for conductance-based models
• drawback is computational complexity
• -> phenomenological models are better

Question 26

Explain the meaning of the statement “Neurons that fire together, wire together”

Answer 26

The synaptic weight between two neurons increases if the two neurons activate simultaneously, and reduces if they activate separately.
Hebbian learning.

Question 27

neuroplasticity

Answer 27

Neurons can change their morphology, which allows for development and learning

Question 28

Explain the resting membrane potential of a neuron

Answer 28

It is the equilibrium voltage state between the voltage gradient and the charge gradient

Question 29

Explain the function of the myelin sheath around nerve fibres.

Answer 29

In the myelinated segments of the axon, signals are transmitted through fast electrotonic conduction. The myelin sheath increases the resistance of the membrane. The action potential is regenerated at gaps in the myelin sheath, the nodes of Ranvier.

Question 30

Name two different types of neural plasticity and briefly describe them.

Answer 30

Functional plasticity: ability to alter and adapt the functional properties of neurons → biological processes modelled by STDP and hebbian learning.

Structural plasticity: physical creation and deletion of synapses, healing the network → not modelled so far.

Question 31

Name three methods for avoiding the problem of overfitting

Answer 31

Regularisation: add a regularisation term at the end of your objective (loss) function that punishes model complexity (i.e L2 regularisation)
Add more (diverse) training data 
Data augmentation:  when you have not enough data, transform your existing data (i.e in case of an image: crop, shift, flip, noise, salt&pepper)

Question 32

Max-Pooling

Answer 32

Max pooling (complex cells) are filters applied in Convolutional Neural Networks (CNN) that perform subsampling on the output of a layer by computing the maximum value of a small subset of features.

• reduces dimensionality
• makes network more robust

Question 33

Define “machine learning” with your own words

Answer 33

Machine Learning: field of study that gives computers the ability to learn without being explicitly programmed. A ML system is composed of a Dataset (S), a model (M), an objective (loss or reward) function (L) and an algorithm A that adjusts M based on S and L.

Question 34

Explain the advantage of converting the convolutional kernels and inputs of a neural network into the Fourier Space. Explain What the main challenge for this approach is?

Answer 34

Conversion of inputs and kernels into Fourier Space reduces the number of operations required for computing convolutions. The challenge lies in keeping the computational cost for the Fourier Transform low.

Question 35

Explain the difference between interpolation and regression.

Answer 35

Interpolation; the approximated function should match exactly with all the dataset points (not used for ML predictive tasks OR overfitting if applied to ML)

Regression: the approximated function should minimise a loss function and does not have to match with all train data points: it should generalise (used for predictive tasks).

Question 36

State the problem of using least squares for linear classification. Explain how support vector machines address this problem

Answer 36

Least Square classification is really sensitive to outliers.

generalisation error by computing a hyperplane that maximises the margin of the classifier, i.e. the smallest distance between the decision boundary and the training samples → insensitive to outliers

Question 37

Name three learning paradigms.

Answer 37

Supervised learning: data object (with features) and labels associated
Unsupervised learning: data object (with features), no labels needed.
I.e k-means clustering, image segmentation
Reinforcement learning: sequential data and an agent. An agent learns by interactions in the environment (thanks to rewards that are the results of his interactions)
Additional: Semi-supervised learning

Question 38

Occam’s Razor

Answer 38

Given two models M1 and M2 with equal performance, the simplest model (i.e less model features, less layers in a convolutional neural net…etc) is to be preferred (also called law of parsimony).

Question 39

. Name and describe the role of the two streams of visual processing

Answer 39

The dorsal stream of visual processing determines an object’s spatial location.
It determines how motor actions are carried out.

The ventral stream determines an object’s identity

Question 40

Name and describe the three conceptual processing layers in the human visual system.

Answer 40

Low-level-vision → basic image analysis
Mid-level-vision → objects, surfaces
High-level-vision → memory, intention

Question 41

Name a problem that can not be solved by a single perceptron with a linear activation function. State briefly or draw the reason for that.

Answer 41

XOR problem: the Perceptron learning rule only converges for linearly separable data sets → It therefore cannot classify the XOR dataset correctly.

Question 42

Explain the difference between on-policy and off-policy learning

Answer 42

The agent is learning on-policy when the behaviour policy and target policy are the same.

The agent is learning off-policy when the behaviour policy differs from the target policy.

Question 43

“bootstrapping value

Answer 43

Propagating value between consecutive states by iteratively exploiting the recursive relationship that is formulated by the Bellman Equation is denoted as bootstrapping.

Question 44

Name and describe the two alternating steps that are performed iteratively under the framework of Generalised Policy Iteration.

Answer 44

the general idea of letting policy evaluation and policy improvement processes interact, independent of the granularity and other details of the two processes

Question 45

operant conditioning

Answer 45

In operant conditioning a subject learns the relationship between a stimulus and its behavior. A stimulus is only presented in response to a certain action of the subject and serves as a reinforcer that increases or decreases the probability of that action.

Question 46

State the reason behind using a linear activation function in the output layer of a Deep Q-Network.

Answer 46

The goal of Deep Q-Network is to estimate the action value function which is Q value. Therefore the output of DQN shouldn’t be constrained by any activation function → it is just a linear layer.

Question 47

Name the two main types of modern neuroimaging

Answer 47

Structural and functional neuroimaging. Structural is Magnetic Resonance Imaging (MRI) and Electroencephalography (EEG). Functional is functional MRI, Positron Emission Tomography (PET) and then there is Single Positron Emission Computational Tomography (SPECT).