Task 3 Flashcards
What is the main aim of cognitive science ?
- explain how people accomplish various kinds of thinking
What is meant by mental representation ? (refer to task 1)
- It is the knowledge in our mind
What opperates on mental representations ?
- Mental procedures /computational procedure which activate thoughts and reasoning
- on mental representations
Give an example regarding mental procedure and mental representation:
- The number system is = mental representation
- What kind of number system is the mental procedure (Roman or Arabic)
What is the core method of AI ?
- computational modeling
Repeat: What is meant by CRUM ?
- Thinking can best be understood in terms of representational structures in the mind and computational procedures that operate on those structures
What is meant by serial processing ?
- Proccessing one information at a time (computer)
What is meant by parallel processing ?
- Processing multiple task at a time (brain and new computers)
What is meant by connectionism ?
- Computational modeling inspired by neuronal structuresof the brain
- Using Units = Neurons
- Degree of activation = frequency of which neurons fire
What are two types of connectionism models ?
- Local representation (concepts are represented by single nodes) -> using parallel processing
- Distributed representation (concepts are represented as activation pattern across multiple nodes) -> using parallel processing
How does the feedforward model work ?
- informaion flows upwards trough out the network (it gets passed forward)
- Info is not encoded in any particular node but rather distributed over the whole network of nodes
What are the two properties of most computational (distributed) model ?
- Feedforward network
2. Recurrent Network
What is meant by a activation fucntion ?
- Mathematical equation that determines the output of a neural network
- usually between 1 and 0
- It determines which neuron should fire or not fire based on whether the neurons input is relevant for the output
- Indirectly it determines the activation of a neuron
What is meant by backpropagation ?
- it is the core algorithm of how machine learn, because backpropegation can be used to callculate the negative gradient
- it uses the delta rule to fid out error in hidden layers
- You calculate the steps backwards and start with the output layer to identify the error
What is the formula for the coast function?
- (x-y)^2…
- x = output of what the network gives
- Y = the output you want it to have
- …= the sum of it
What are three ways to adjust a neurons activity ? (How to change a problem identified by the delta rule)
- changing the bais + changing the weights + changing the activation from the previous layers
How does the backpropagation determine the gradient ?
- By knowing each desired activty of each neuron and knowing how the activites can be changed
- We can go step by step from the last layer to the first layer and adjust weights bias in order to identify the negative gradient
What is meant by the delta rule ?
- Part of backpropagation
- The rule of changing the weight of a connection
- it is using the difference between a target/goal activation and an actual obtained activation-> If this is zero than no adjustment needed
- basically using an error function
What is meant by gracefull degradation ?
- Ability of machine or network to maintain limited functionality even when a large portion of it has been destroyed -> the other neuron make up for their work
- to prevent catastrophic failure
- Slow loss of performance
- Important does not account for local distributed models
What is meant by Brain computer interface ?
- devices that enable its users to interact with computers by means of brain/neuronal activity only
What types of Brain machine interface are out there ?
- Assitive BCI
- Rehabilitative BCI
What is meant by Assitive BCI ?
- aims to substitute lost functions such as communication or motor function, via robotic prosthesis
What is meant by Rehabilitative BCI ?
- restoration (wiederherstellung) of brain function and/or behaviour by manipulation of neurophysiological activity
Are BCI scanners invasive or non invasive ? (how can brain activity be meassured ?)
- They are both
- Non invaisve = EEG/fmri
- Inavsive = implantation of electrodes
What existed before BCI ?
- HCI = Human-computer interfaces
- Ex: keyboards
and mouses
For which people is BCI gnerally being used and why ?
- Locked in syndrome and disabled people
- Because they need some sort of reliable muscular control such as eye movement
- The
primary users of BCI systems are individuals with mild to severe muscular handicaps (not mental disorder) - Spinla cord injuries and strokes
Which method is used by the most BCI methods ?
- EEG (Steady state potentials)
2. Imaging techniques
What is the overall goal of a BCI system ?
- to interact with a device
Why is an EEG still more used then invasive techniques ?
- beacuse it has high temporal resolution
- reasonable spatial resolution
- it is portable
- It is easy to handle (putting electrodes on the scalp)
Which disease might lead to the need of an BCI ?
- Paralysis since it can disconnect the brain from the body
- No volitional movemnt possible
- > BCI can restore mobility and independence for people via translating neuronal activty into control signals for the device
Name an exmaple where BCI has been proven to work regarding EEG:
- The robotic arm
- After paralysis due to the help of a BCI system it was possible to perform 3D grap and reach movements
- signals were decoded from a small, local population of motor cortex neurons
What is meant by Locked in syndrome ?
- severely motor-disabled patients become incapable to communicate while being fully conscious and awake
Name an exmaple where BCI has been proven to work regarding fmri:
- Seplling device
- decodes blood oxygenation level-dependent (bold activity) regarding different mental tasks
- Allows translating any freely chosen answer (letter by letter)
How can participants influence Bold activity ? (Three aspects)
- Location of the signal source by performing three different mental tasks
- Signal onset delay (delaying the start of the mental task for 0s, 10s, or 20s)
- Signal duration by varying the mental task duration between 10s, 20s and 30s
- 27 different combinations
How does BCI in general work ? (look up again
- Senors (mostly electrodes) measure brain activity
- Goes to the feature extractor = raw signals into relevant signals
- Then it goes into the feature translator where signals get tranformed in logical controls
- Control interface converts logical control to semantic control
- Goes into the device controller where semantic controls get translated into physical device-specific command (maybe just draw)
Based on CRUM what would be a synonym for mental representation ?
- Data structures
- cooking ingredients
Based on CRUM what would be a synonym for computational procedure ?
- Algorithm
- cooking instructions
What is meant by a cognitive theory ?
- Explains how the brain works
- Needs mental rerpresentations and mental procedures to operate on those
What is meant by a cognitive model ?
- makes the structures (mental representations) and mental procedures more precise by interpretating them with a comupter programm which needs data structures and an algorithm
What else do we need in order to create an AI, besides a cognitive theory and a cognitive model ?
- A software (java) and a platform a computer
What kind of problems can both modles “local representations” and “distributed representation” solve ?
- parallel constraint satisfaction
What is meant by parallel constraint satisfaction ?
- simultaneously satisfy numerous constraints
What is meant by the feedforward model in the distributed represenatation model?
- information flows upward through the network
- Information is not encoded in any particular unit but distributed over the whole network (hidden layers)
- From first input unit row to last row output unit
What is meant by recurrent networks ?
- Activation from the output units feeds back into the input unit
Evaluate connectionsim in terms of the above mentioned criterias (form of computational model)
- the mental representation in connectionism lacks on:
1. Representational power since it has difficultys explaining complex logical relations
2. Neuronlogical plausability: only approximations to the behaviour of real neurons
How is in generel a artifical neural network structured ?
- multiple units
- links (either exitatory or inhibitory)
- parallel processing
- Hebbian learning rule
What is meant by an invasive BCI ?
- Surgical implantation of electrodes that
measure neurons’ activity (for relevant
behavior)
What is meant by non invasive BCI?
- Record brain signals from the scalp
- 7 types of brain signals can be detected (steady state potentials and blood oxygenation lvl)
What are the 5 assumptions of connectionism?
- Neurons integrate info
- neurons pass on info about their input levels
- Brain structures are layered
- Neurons influence each other depending on the
strength of their connection - Learning changes strength of connection
How does the recurrent network work ?
- It states that recurrent network gain activation from output units feeds back into input units (feeback signal
Why are neuronal networks so key for machine learning ?
- Damage resistance and fault tolerance (graceful degredation)
- Content addressable memory
- Constraint satisfacion
What is meant by content addressable memory ?
- If given one part of a memory, the system will
activate in a way that more info about the
memory will be recovered
What is meant by constraint satisfaction ?
- As activity flows through the network, all
neurons influence each other depending on
their weight, each input can be seen as a
constraint - The newtrok tries to satisfy as many constraint as possible
What types of activation function exist ?
- Linear
- Non linear = Binary or Sigmoid
- threshold linear
What is better linear or non linear activation function ?
Linear = Output is propotional to the input, not capable of backpropagation -> No complex input can be handled
- Non linear = Capable of backpropagation -> can handle complex input -> therfore used for deep neuronal network models
What is meant by a binary actiavation function?
- It is a threshold based activation function
- you need to surpass a certain lvl
What is meant by the coast function?
- It is the error of output in comparison to the desired output
