Holcombe lectures Flashcards
What is reductionism?
• Reductionism- the practice of analysing and describing a complex phenomenon in terms of its constituents, especially when this is said to provide a sufficient explanation
What is the main feature of a good explanation?
o Explanation often requires an account of how the parts work together (leads to satisfaction of explanation)
What level of explanation did Barlow think should be used to explain how the brain produces behaviour and experience?
• Explaining how the brain produces behaviour and experience
o Barlow’s neuron doctrine- how components give rise to perceptual experience
Appropriate level of explanation for this would be neuronal coordination (neural networks)
What is emergence? What is its impact on computational neuroscience?
• Emergence- to relate to something like an unexplained or unexplainable appearance of an entity or property and/or something which is not “reducible” to well-defined interactions of other entities
o The relation between parts is important
When and what the parts are made/made of is not very important
o Behaviour of the system emerges from the relation between the parts
o Computational neuroscience is based on the idea that behaviour and experience are emergent
What are the four main characteristics of computational modelling?
• Computational modelling
o Assumes behaviour and experience are emergent
o Is not entirely satisfied until we can build something that accomplishes the target behaviour
o Seeks simple explanations of behaviour and experience (as simple as possible while satisfying requirement above)
o Copies aspects of the brain as building blocks for model
What basic building block does computational modelling use to model the brain? Give example, reason and limitation
Typically uses a simplified neuron as its basic building block
• Simplified models typically require prior knowledge of the components
• Have full understanding of monosynaptic stretch reflex and a successful model
• Must understand many interactions for better explanation
What is the aim of modelling and simulation?
o Build something which mimics the thing that is trying to be understood
Can confirm explanation and allow for further exploration of level of understanding
• If a model can be built that models the subject, then the theory has a chance of being right
What is the sign of a good model?
o Doesn’t duplicate- strips away irrelevant detail
Only represent relationships between the parts- not as concerned as to what the parts are
What models were used to explain the brain over time? Why?
The mind is like a wax tablet (Plato, 1st century)
• Impressionability
• Wax could be too hard, soft, or full of impurities
Clock (17th century)
• Spurred the scientific revolution
• The mind is like an automaton
Clockwork automation
• Descartes: nerves act like hydraulic pipes
Computer (20th century)
• Biased our understanding of the brain
• Computer metaphor:
o Computers are programmed with a series of steps, and different functions done by different modules of code that interact and execute in a series of steps
• Different parts of the brain do different things- led to box and arrow theories
21st century-Copying the brain itself
• Need to reduce the complexity of the brain
• Cognitive neuroscience should succeed by finding level of abstraction by finding diagrams that capture the important aspects of what is interacting such that behaviour emerges
o However, although ion channels may be more complex than needed, brain areas are too simplistic
What is an algorithm?
• Algorithm- a specific way to achieve something
o There’s more than on way to do things
What is the aim of psychologists?
• As psychologists, want to understand human abilities
o Computers and the brain may do things in different ways
At what level can behaviour be explained?
• Explaining behaviours requires considering some aspects of neurons
o Ignore many small details so that a simplified simulation of neurons can be built
What is the connectionist model approach? How can it be used to explain behaviour?
• Connectionist model approach- consist of a number of different nodes that interact via weighted connections that can be adjusted through by the system through different ways, the most common being backpropagation of error
o Simplification: can represent the ability of neurons to make the next one fire, ignoring the details of the neurotransmitters or the membrane potentials
Why is hard to make robots behave like humans?
Robots vs humans-
• Easy to design a robot for a specific task, but extremely hard to design one for general tasks
• Robots have difficulty doing things that humans can do with ease
• Robots don’t process information like humans
What are the capabilities of single neurons?
• Singular neuron capabilities-
o Simple enough to understand fully- reflex, Pavlovian learning
Why are neuronal interactions are important?
o Important brain functions for information processing-
Excitation or inhibition at synapse
How many neurons are there in the cortex?
86 billion neurons
How many connections does the average neuron have with other neurons?
10,000
What is the threshold activation rule and what does it affect?
Each connection is affected by a threshold activation rule
• Threshold activation rule determines whether neuron fires or not
Is the computer an accurate representation of how the brain works? Why/why not?
• Compare to computer
o Naïve computer-style box-and-arrow psychological theory-. Understand how network functioning differs
o The computer is a misleading metaphor-> computers become full but brains do not
• Computers vs brain-
o Computers-
Stores attributes of objects separately from each other
No interference (advantage) but no generalisation (disadvantage)
o Brain
Stores more than one item using the same set of units
• Because the network is small, interference rapidly sets in
o When learn something new, changes some of the connections we were using for old memories
Interference (disadvantage) but there is generalisation possible (advantage)
A lot of remembering is reconstructive- it’s not like retrieving a video file
Hard to know whether one is good at storing and retrieving the details from the episode vs how good you are at reconstructing and predicting the details from previous experiences
No separation between computation and memory
How is the connectionist model represented diagrammatically (especially in Simbrain)?
• Connectionist model representations
o Neurons are represented by circles
o The number inside is the neuron’s activation, perhaps its firing rate or voltage potential
o Synapses are modelled by weights or connections
o The synaptic terminal is represented by the semicircle- its colour indicates the connection strength
Red- excitatory connection
Blue-inhibitory connection
o Synaptic strength= connection strength= weight
What is the linear activation rule and how does it work?What is a disadvantage of ti?
o Linear activation rule- simply pass on the overall votes/stimulation
Adding excitation and subtracting inhibition
But linear is not enough for binary decisions
Synapses can have different weights- weighted average
What is the threshold activation rule and how does it work? What is an advantage of it?
o Threshold activation rule-
If stimulation is more than threshold, activate
Adequate for modelling of binary decisions
What do neurons representing the same location do to each other?
Mutually excite each other
What do neurons representing different locations do to each other?
Inhibit each other
What is the impact of lateral inhibition on a layer?
Lateral inhibition causes units on all layers representing a single object to eventually all become active
Neurons that do recognition also do attention- via lateral inhibition
What is posner cuing?
Posner cuing: cue pre-activates the location, helping subsequent target to win the competition sooner
What kind of connectionist network explains human rapid categorization performance?
A feedforward connectionist network explains human rapid categorization performance
How does a parallel distributed processing network work?
• Parallel distributed processing network-
o Categories emerge from overlap among all instances
o Strengthened pathways result in more rapid responses, whilst weaker pathways results in slower responses
How does the human connectionist network compare to a computer in terms of:
- Location
- Content
- Interference
- Generalisation
- Capacity limit
Connectionist network:
>Individual concept distributed widely
>Content-addressable (advantage)
—–Address memory with content and then same neurons that represent content will have connections to other subjects related to content—-> search is guided and directed by connections
>Prone to interference
>Generalises naturally (advantage)
——Wisdom
———-Expert can more rapidly activate core that they are experts in-> can lead to predictions
>Soft capacity limit (interference as memories accumulate) (advantage)
—Every time acquire new memory, subtly degrade old memories
——–If new memory activates nodes of old memories, some aspects of old memories may be strengthened
Computer: >Memories localised >Time-consuming search >No interference (advantage) >No built-in generalisation >Once full, it’s full
Compare the computer and the brain in terms of:
- Basic computation unit
- Communication between units
- Speed of messages
- Number of units
- Time for a single computation
- Storage
- Robustness
- Energy needed
- Tasks
Computer: >Transistor >Electrical wires >About 300000km per second >55 million transistors (P4) >3 Ghz: 109 >Memory, manipulation separate >Catastrophic failure from minor injury >45 watts >Chess: --Almost entirely serial -------Only use of one unit at a time --Deep Blue evaluated between 100 million and 200 million chess positions per second --During matches with Kasparov, it averaged 126 million positions per second
Brain:
>Neuron
>Synapses, transmitters, modulators, hormones
>7-430 km per hour
>More than 80 billion neurons
>About 200 Hz
>Memory, manipulation integrated and distributed memory
>Continually adaptive and graceful degradation
Information is distributed-> loss of a unit is less catastrophic
>20 watts
>Chess:
—Massively parallel
——–Can use neurons at the same time
—Chess skill: pattern recognition
Why is the brain never full?
• Individual memory distributed across thousands or millions of synapses
• Each new memory may degrade many of the old memories
• But also build on old memories, strengthening elements in common while degrading others (interference)
o Forgetting is more about interference
• Never hit a limit, but constantly degrading/losing old memory while making new
• Make new synapses
• Make very few neurons (too few to help significantly)- this is not the explanation of how we make memories
What is Moravec’s paradox?
• Some things that we do, we can understand well whilst we cant understand other processes we use
o Many things that are easy to us are hard for computers e.g. walking
o Many things that are easy for computers are hard for us e.g. arithmetic
• The discovery by artificial intelligence and robotics researches that while high-level reasoning requires very little computation, sensorimotor skills require enormous computational resources
What is Polanyi’s paradox?
• Polanyi’s paradox
o Many tasks people understand tacitly and accomplish effortlessly but neither computer programmers nor anyone else can enunciate the explicit rules or procedures
o The tasks that have proved hardest to automate are those demanding flexibility, judgment and common sense- skills that we understand only tacitly
Do not understand these consciously
o Polanyi’s paradox also suggests why high-level reasoning is straightforward to computerize and certain sensorimotor skills are not. High-level reasoning uses a set of formal logical tools that were developed specifically to address problems: for example, counting, mathematics, logical deduction and encoding quantitative relationships.
o In contrast, sensorimotor skills, physical flexibility, common sense, judgment, intuition, creativity, and spoken language are capabilities that the human species evolved, rather than developed. Formalizing these skills requires reverse-engineering a set of activities that we normally accomplish using only tacit understanding
What paradoxes can explain why we’ve been slow to make robots that can perceive the world and walk around?
• We’ve been slow to make robots that can perceive the world and walk around
o Moravec’s paradox-
Logical reasoning doesn’t require much computation; perception and locomotion does
o Polanyi’s paradox
We know more than we can tell
We don’t know consciously how we do perception and locomotion
o Perception and locomotion not easily done by conventional serial computation computer architecture
Who was Rodney Brooks and what was his opinion on the current progress of AI? What direction do we need to go in for AI according to him?
• Rodney Brooks-
o Early AI researchers made a big mistake: they thought intelligence was stuff they found hard to do, when it is the things that we find easy to do
No cognition. Just sensing and action. That is all I would build and completely leave out what traditionally was thought of as the intelligence of artificial intelligence
Compare the computer and the brain in terms of maths and sensory-motor capabilities in terms of:
- Arithmetic speed
- Storage of books
- Novel capabilities
- Motor abilities
- Sensory capabilities
Computer
Can do millions of arithmetic operations in a second
Can easily store and instantly access thousands of books
Can’t even write trashy novels
Can barely walk down stairs
Had bad visual recognition, until imitated connectionist networks (still easily fooled)
Brain Can’t even do ten arithmetic operations a second Has trouble memorizing even one book Can write good novels Can walk, hop and jump Can recognise objects and people
What is dyspraxia?
• Need to connect sensory signals to movement
• Dyspraxia- disorder of movement not due to:
o Weakness
o Inability to move/paralysis
o Abnormal muscle tone or posture
o Intellectual deterioration
o Other direct motor deficits
• Dyspraxia is a perceptual problem and it is the reduced ability to coordinate, perform, plan or carry out specific movements even when there is no paralysis
What are the symptoms of people with dyspraxia?
o Perseveration- Some people with dyspraxia repeat the same movement over and over again
o Movements are either over exaggerated or appear slightly odd
