Task 4-BCI, connectionism, activation functions, graceful degradation, delta rule Flashcards
BCI, connectionism, activation functions, graceful degradation, delta rule
BCI
- CONNECTS THE BRAIN WITH A COMPUTER / PROTHESIS ETC BY SIGNAL ACQUISITION (OF THE BRAIN) & SIGNAL PROCESSING
- CAN BE INVASIVE AND NON-INVASIVE
- ASSISTIVE BCI: AIMS TO REPLACE LOST FUNCTIONS (E.G.
COCHLEAR IMPLANT, ARM PROTHESIS) - REHABILITATIVE BCI: AIM TO FACILITATE RESTORATION OF BRAIN FUNCTION
feature extractor= transforms raw signals(EEG) into readable signal
control interface: gtarmsforms the signals into semantic commands
device controller: executes these demand s
Connectionism
- EMPHASIZES IMPORTANCE OF CONNECTIONS AMONG
NEURONLIKE STRUCTURES IN A MODEL (INTERCONNECTED
NETWORKS OF SIMPLE UNITS) WHICH WORK PARALLEL TO EACH
OTHER - CONSISTS OF DIFFERENT LAYERS: INPUT, HIDDEN AND OUTPUT
UNITS - LOCAL REPRESENTATION: EACH UNIT IS INDEPENDENTLY
ASSOCIATED WITH ONLY ONE REPRESENTED THING (SIMPLE,
“GRANDMOTHER-CELL) -> SYMBOLIC, units, links
DISTRIBUTED REPRESENTATION: SPECIFIC KNOWLEDGE IS REPRESENTED BY ACTIVITY OF DIFFERENT UNITS (MORE COMPLEX) LIMITATIONS: ADDITION OF NEW INFO CAN CAUSE
LOSS OF OLD INFO; LEARNING CANNOT BE IMMEDIATE -> SUB SYMBOLIC, feedforward, recurrent network
Practical applicability: backprpagation techniques have been used by engineers for prediction of stressors on materials
Connectionist models are always models of learning
Decision making : you can store decisions by learning 2 different outcomes and recalling them
ACTIVATION FUNCTIONS
THRESHOLD) LINEAR (/), SIGMOID (S) AND BINARY THRESHOLD (I)
Delta rule
METHOD USED TO CALCULATE THE DIFFERENCE BETWEEN ACTUAL AND DESIRED OUTPUT (ERROR) AND CHANGE THE
A rule for changing the weight of the connection, which will in turn change the activity level of i (Δwij = [ai (desired) – ai (obtained)] aj), where Δwij is the change in weight of the connection from unit j to unit i to be made after learning.
graceful degradation
ABILITY TO PRODUCE REASONABLE APPROXIMATIONS TO THE
CORRECT OUTPUT FOLLOWING A DAMAGE TO SINGLE UNITS
loss of a few units ist detremental
Generalisation
if recall cue is similar to the pattern, the output will produce similar response
Fault tolerance:
Even if pattern is incomplete or damaged, output will recognize it
network is robust against errors in representation
prototype extraction
: if more than 1 pattern is possible, output will produce pattern with the most activation
Autoassociator
Replaces at output the same pattern that was presented at input
Learning occurs when the weights change, so that the internal input to each unit matches the external input
Resistant to incomplete and noisy pattern (and cleans them up)
Uses the delta rule
Difference autoassociator to pattern associator:
Recurrent connections that give feedback
Same input pattern at output
Doesnt need external teacher
traditional hebbian learning
Traditional: doesnt specify how much a connection should increase and the exact conditions that need to be met for increase
Neo Hebbian learning
\: solves this traditional problem Mathematical equations (dynamical differential)weights change in their strength Nodes = instar/outstar
Differential Hebbian learning:
solves the problem that connection only increase in strength
Connections change according to the difference between the nodes activation and the incoming stimulus signal: change can be positive/ negative/ null
Drive reinforcement theory
solves the problem that the change is only linear = sigmoid curve
considers recent history of stimuli, e.g. recent trials of learning (=temporal memory)
Hippocampus
During behavior, memories are stored in Hippo
During sleep these memories are consolidated into neocortex by synchronous bursts (Hebbian Learning) Autoassociative memory: recall a memory with just a cue (subcomponent of the memory unit)
DG=>Competitive learning (sparse memories)
CA3=> recurrent connections, autoassociation
CA1=> Competition
distributed representation
attributes of concepts distributed through network => good representational and neurobiological power
Networks that learn how to represent concepts or propositions in more
complex ways and distribute over complex neuronlike structures
bachpropagation
calculates error between desired and actual level of activation, therefore changes weights
local representation
Neuronlike structures are given an identifiable interpretation in terms of
specifiable concepts + propositions
a kind of NN neutrons are specific concept like “apple” limiting representational power
pattern associator
often works with Hebb rule, learns to associate one pattern with another
sub symbolic
knowledge spread over units , learning as change of connection between chunks and production rules
symbolic learning
learning of new chunks and production rules by e.g. compilation
Both representations can be used to perform PARALLEL CONSTRAINT SATISFACTION
distributed and local representation
parallel constraint satisfact8on
processing simultaneously satisfies numerous constraints
each input is setting constraints on the final state, a design is reached when there is reasonable fit between the input and teh output
used with problem solving
Example: When putting together a school schedule one needs to take into account various constraints imposed by classroom availability and the preferences of professors and students
relaxation
aim of the network is compelded after repeatedly activating until u til it reaches stable level=> learning has eben a achieved
constraints can be satisfied in parallel by repeatedly passing activation among all the units, until after some number of cycles of activity all units have reached stable activation levels
process is called RELAXATION
problem solving
Problem Solving
Example: Is Alice outgoing or shy?
Concepts are represent by units
The problem has both positive constraints, such
as between “likes parties” and “outgoing”, and
negative constraints, such as between “outgoing”
and “shy”
Positive constraints are represented by excitatory
connections
Negative constraints are represented by inhibitory connections
An external constraint can be captured by linking units representing elements that satisfy the external constraint and is linked to it either positively or negatively
For example, the external constraint could be that we know that Alice likes programming and likes parties
A problem solution consists of when a group of units is activated by the set containing outgoing, while correctively deactivating the set containing shy
Consequently, outgoing will win over shy because outgoing is more directly connected to the external information that Alice likes parties
Constraints can be satisfied in parallel by repeatedly passing activation among all the units, until after some number of cycles of activity all units have reached stable activation levels process is called RELAXATION
SETTLING – achieving stability
planning
Constructing plans is usually a more sequential process understand in terms of rules or analogies rather than parallel processing
uses rule based systems
decision
We can understand decision making in terms of parallel constraint satisfaction The elements of a decision are Actions and Goals
Actions that facilitate a goal have positive constraints and negative restraints come from incompatible relations
positive internal constraints come from facilitation relations: if an action facilitates a goal, then the action and goal tend to go together
The external constraint on decision making comes from goal priority, in which some goals are inherently, desirable, providing a positive constraint
Once the elements and constraints have been specified for a particular decision problem, a constraint network can be formed
Units represent the various options and goals, and pluses and minuses indicate the excitatory and inhibitory links that embody the fundamental constraints
Analogy can also be useful in decision making, since a past case where something like A helped to bring about something like B may help one to see that A facilitates B But reasoning with analogies may itself depend on parallel constraint satisfaction
explanation
Explanations should be understood as activation of Prototypes encoded in networks
Example: Understanding why a particular bird has a long neck can come via activation of a set of nodes representing swan, which includes the prototypical expectation that swans have long necks
Units representing pieces of evidence are linked to a special evidence unit that activates them, and activation spreads out to other units
learning
Learning
There are two basic ways in which learning can take place in a connectionist model: Adding new units or changing the weights on the links between them
Work to date concentrates on weight learning, as is demonstrated in the Hebbian Learning, in which a link between A and B gets stronger with subsequent activation
A technique called Backpropagation adjusts the weights that connect the different units by assigning weight randomly, determining errors and propagating backwards
Assign weights randomly to the links between units
Activate input units based on features of what you want the network to learn about
Spread activation forward through the network to the hidden units and then to the output units
Determine errors by calculating the difference between the computed activation of the output units and the desired activation of the output units. For example, if activation of quiet and studies hard activated jock, this result would be an error Propagate errors backward down the links, changing the weights in such a way that the errors will be reduced
Eventually, after many examples have been presented to the network, it will correctly classify different kinds of students
Disadvantages:
Requires supervision
Tends to be slow, requiring many hundreds or thousands of examples
language
Word recognition can be understood as a parallel constraint satisfaction problem by representing hypotheses about what letters and words are present
example with cat Relaxing the network can pick the best overall interpretation of a word
cm psychological p
Connectionist models have furnished explanations of many psychological phenomena, but also suggested new ones
Backpropagation techniques have simulated many psychological processes
cm neuro p
the artificial networks are similar to brain structure in that they have simple elements that excite and inhibit each other.
Real neural networks are much more complicated and complex than the units in artificial networks, which merely pass activation to each other
Furthermore, in local representations each unit has a specifiable conceptual or propositional interpretation, but neurons do not have such local interpretation
Artificial units leave out the chemical parts, like neurotransmitter
We can think of each artificial unit as representing neuronal group, a complex of neurons that work together to play a processing role
Many local networks use symmetric links between units, whereas synapses connecting neurons are one- way
While Hebbian learning does occur in the brain, backpropagation
cm practical application
Connectionist models of leaning and performance have had some interesting educational applications, for example knowledge required for reading
Reading is a kind of parallel constraint satisfaction where the constraints simultaneously involve spelling, meaning and context
Backpropagation techniques have been used to assist engineers in predicting the stresses and strains of materials needed for buildings
Connectionist models are widely used in intelligent systems
For example, in training networks to recognize bombs, underwater objects, and handwriting
interpret the results of medical tests and predict the occurrence of disease
ASSISTIVE BCI SYSTEMS
substitute lost functions, enable control of robotic devices or provide functional electrical stimulation
REHABILITATIVE BCI SYSTEMS
restore brain function and/or behaviour by manipulation of self-regulation of neurophysiological activity
Cortical Resource Allocation
Variable-resolution representations in the sensory cortex: spatial resolution is highest at the center of gaze Plasticity
NEURAL INTERFACE SYSTEM (NIC)
translates neuronal activity into control signals for assistive devices
grandmother cell
example of local representation in perception
Neurons selectively respond to more and more complex attributes, so there might be ‘grandmother cell’ which are so specific as they fire in recognition of your own grandmother
Hypothesis is fundamentally unsound rejected
NEURAL NETWORKS
Not all is lost if there is any deterioration in stimulus signal or loss of individual units
• REDUNDANCY – although some info might be lost, enough is still available to get the message across
Gradual deterioration in performance of a distributed system
weight
Weight = strength of connection
what can connectionist models do for us
An auto associator network can be trained to respond to collections of patterns with varying degree of correlation between them
When the input patterns being learned are highly correlated, the network can generate the central tendency or prototype that lies behind them, another form of spontaneous generalization
A single auto associator network can learn more than one prototype simultaneously, even when the concepts being learned are related to each other. Cueing with the prototype name will give recall of the correct prototype (which was never presented to the network complete)
An auto associator network can extract prototype info while also learning specific info about individual exemplars of the prototype
Thus, the network’s capability to retrieve specific info from cues (content addressability) means that, given a specific enough cue, it can retrieve the specific info of the individual exemplars from which the prototype generalization is constructed
Such behaviour is an example of a unitary memory system that can support both ‘episodic’ and ‘semantic’ memory within the same structure
connectionist modelling
Connectionist Modelling is inspired by information processing in the brain and a typical model consists of several layers of processing units
Unit can be thought of as similar to a neuron, with each layer summing info from units in the previous layer This info processing is derived from observations of the organization of the brain:
The basic computational operation in the brain involves one Neuron passing info • Related to the sum of the signals reaching it to other neurons
Learning changes the strength of the connections between neurons and thus the • Influence that one has on the other
Cognitive Processes involve the basic computation being performed in parallel by a • Larger number of neurons
Info, whether about an incoming signal or representing the network’s memory of • Past events, is distributed across many neurons on many connections
In contrast to models in Artificial Intelligence (AI) which contain a set of rules, connectionist models are said to be neurally inspired by our brain
connectionism and the Brain
Neurons integrate Information
Neurons pass Information about the Level of their Input
Brain Structure is Layered
Learning is achieved by changing StrengthStrength between Neurons
threshold linear
Real neurons have thresholds firing occurs only if net input is above threshold
sigmoid
Range of possible activity has been set from 0 to 1. When the net input is large and negative, the unit has an activity level close to 0. As the input becomes less negative the activity increases, gradually at first and then more quickly. As the net input becomes positive the rate of increase in activity slows down again, asymptoting at the maximum value which the unit can achieve. They prevent saturation and are good in noise suppression.
binary threshold
Models neurons as two state devices as either being on or off. This ensures that if the net input is below threshold, there is no activity. Once the net input exceeds the threshold, the neuron becomes activated.
DISTRIBUTED PROCESSING
In connectionist models info storage is not local, but distributed across many different connections in different parts of the system
LOCAL PROCESSING
Traditional models of cognitive processing usually assume a local representation of knowledge stored in different, independent locations
GRACEFUL DEGRADATION
ability to continue to produce a reasonable approximation to the correct answer following damage, rather than undergoing failure
Any info processing system which works in the brain must be fault tolerant, because the signals it has to work with are seldom perfect
An attractive aspect of content-addressable memory is that it is indeed Fault Tolerant (because no input unit uniquely determines the outcome)
Properties of Pattern Associators
gneralizatioh
During recall, pattern associator generalize
If a recall cue is similar to a pattern that has been learnt already, a pattern associator will produce a similar response to the new pattern as it would to the old
fault tolerance
Properties of Pattern Associators
competitive learning can be divided into three phases:
EXCITATION: Excitation of the output units proceeds in the usual fashion by summing the products of the activity of each input unit and the weights of its connection
COMPETITION: The units compete with each other and the identification of the winner may be achieved by selecting the unit with the highest activity value
WEIGHT ADJUSTMENT: Weight adjustment is only made to connections feeding into the winning output unit in order to make it more similar to the input vector for which it was the winner
goal of bcc
provide new channel or output for the brain that requires voluntary adaptive control by user Helping handicapped people
BCI system: allow user to interact with device
Interaction is enables through intermediary functional components, control signals and feedback loops
Intermediary functional components: perform specific functions in converting intent into action
Feedback loops inform each component in the system of the state of one or more components
problem machine learning
Concerns about the biological plausibility of current machine learning approaches: if our brains’ abilities are emulated by algorithms that could not possibly exist in the human brain then these artificial networks cannot inform us about the brain’s behavio
z.b.humans learn with supervisor most successful deep networks have relied on feed-forward architectures whereas the brain includes massive feedback connections
no equivalent to backprpagation
humans influenced by chemicals
competitive networks
connection. between winning in and output will be strengthened
loosing will be weekend
3 phases
excitement
competition
weight adjustment
