Questions on Notes Flashcards
to become one with the knowledge
Give details on the two brain initiatives
Brain Initiative:
- broad looked into different ways of studying the brain
- $300mil
- finished
Human Brain Project:
- map enough neurons to simulate part of the brain
- $1bil
- complained for being too narrow
- ongoing
Name the types of output functions
Linear
Step
Sigmoid
Describe the 4 learning rules
Plain Hebb: if both the pre-synaptic and post-synaptic neurons fired together the weights should increase.
Post-synaptic: if the post-synaptic neuron fires the weight should increase
Pre-synaptic: if the pre-synaptic neuron fires the weight should increase
Covariance: the increase in the weight is proportional to the difference between the two neurons
Describe the types of neural networks
Perceptron:
- single layer
- linearly separable
Multilayer Perceptron:
- not linearly separable as long as a non-linear output function is used
- hidden layers
Recurrent Neural network
- creates memory
- used when there is a time series
Describe the data sets needed and why
Training set
- this is used so the NN learns the weights
Validation set
- used to prevent overfitting
Give examples of applications for NN
NetTalk
- learnt to read aloud from written text
AlphaGo
- taught itself the game of Go
What are the limitations of NN
- can be overfitted to the data
- the results can be fooled
- cyberattack of dataset
- Bais from datasets used
- Blackbox means it is not easily possible to understand the process the NN took
Detail what is required to define a cellular system
Cell space
- 1D 2D 3D hexagonal
Neighbourhood
- which cells influence each other
Transition rule
- how a cell reacts to its neighbours
Boundary conditions
- Assigned, periodic, adiabatic, reflection
Initial conditions
What is the universality of CAs
Theoretically, they can compute anything as they are capable of universal computation. The game of life can do this
Give examples for the uses of CAs
- solve mazes
- model physics
What are the limitations of CAs
- no way to predict the outcome without running it
- difficult to design the rules for a specific behaviour
Detail the 4 pillars of evolution
Population: a group of several individuals
Diversity: individuals have different characteristics
Heredity: characteristics are transmitted over generations
Selection: individuals produce more offspring then the environment can support
What are the genotype and phenotype
Genotype
- the genetic material of the organism
- selection does not directly operate
Phenotype
- manifestation of the organism
- selection operates on the phenotype
Describe the differences between natural and artificial evolution
- Fitness is a measure of the individual’s ability to solve the problem
- Selection is according to the performance criteria
- It is expected that the performance will improve over the generations
Describe the types of genetic representation
Discrete
- eg binary string mapping to several phenotypes
Sequence
- sequence of values (travelling salesman)
Real-value
- sequence of real values that represent parameters eg wing profile
Tree-based
- describes a tree with branching points and terminals
Describe the types of fitness functions
Explicit: eg lift-drag
Implicit: maximum flight time
Subjective: most aesthetic
Why is selection pressure important
too strong and there is a rapid loss in diversity which could lead to a suboptimal solution.
too slow and the solution will not converge in a reasonable timeframe
List the types of selection
Proportional
Rank based
Tournament
Describe proportional based selection
- the probability of an individual producing offspring is proportional to its fitness with regards to the group’s fitness.
- if fitness is uniform selection is weak
- if there is an outperforming individual it can result in a lack of diversity
Describe rank based selection
- individuals are ranked based on their performance
- offspring is then produced proportional to their rank or from the selected best few
Describe tournament based selection
- a random group of set size is picked and competes against one another
- the champion is then selected to produced offspring
Describe the types of inheritance
One point: a proportion of the genotype at a point is switched
uniform: every other value in the genotype is switched
Arithmetic: an average of the genotype is produced
How can performance be monitored
Fitness landscape
- observe the average fitness of the population and the maximum fitness
Diversity
- when diversity decreases the potential for further evolution decreases
Only really useful when the problem is stationary
Give examples of artificial evolution
- To design an antenna
- HyperNeat
- Golem Project
- RoboGen
Describe HyperNeat
HyperNeat is used to evolve both the weights and the topology of a NN. This has then be used to design soft robots simulations designed for fast locomotion
Describe the Golem Project
- The fitness of a robot was tested in simulation
- The stick based robots where then 3D printed and tested in real life
Describe RoboGen
- Aimed to cross the reality gap, by ensuring the same performance in simulation and in real life
- Described as having embodied cognition by designing the body and the controller at the same time
Describe the thinking behind competitive co-evolution
- may increase adaptivity by producing an evolutionary arms race
- more complex solutions may arise
- helps overcome the bootstrap problem (getting off the ground)
- competitors have opposing fitness functions
- to prevent the recycling of solutions a Hall of Fame can be created and then all the best individuals put against one another
How would one ensure cooperation
homogenous swarm with swarm level selection
List the differences in engineering and natural control
Engineering
- designed system
- limited DoF
- simplified copy of biology
- fully actuated
- fully controlled
- rigid body, model-based
- central controller
Nature
- systems are grown
- high DoF
- mostly soft
- high redundancy
- passive DoFs
- distrubuted control
What are CPG and describe their use
- Biological neural circuits that produce rhythmic patterned outputs without sensory feedback or input.
- They result in the smooth transition of gaits with simple 1D control
- Globally stable non-linear oscillators are used for control and allow for self-stabilisation when perturbed
Give a biological and engineering example where CPG are used
- in the spine of vertebrate to control gait and gait changes
- salamander robot
Describe the process of traditional AI control and its key ideas
sensors -> perception -> modelling -> planning -> tast execution -> motor control -> actuators
every action requires a lot of processing, thinking, planning and model building
Describe the process of behaviour based control and its key ideas
/ > | manipulate the world | > \
/ ^ \
/ > | build maps | > \
/ ^ \
sensors > | explore | > actuators
\ ^ /
\ > | avoid hitting things | > /
\ ^ /
\ > | locomote | > /
- bottom-up hierarchical structure
- reactive
- layers easily added
Describe the key ideas of behaviour based control
Embodiment: the physical body plays an important role in intelligence
Situatedness: control is based on its body and the environment
No planning: the world is it’s own best model
Emergent complexity: intelligence emerges rather than being programmed
What is embodiment
The use of the physical design to reduce the computational workload taking inspiration from biological systems
What is morphological computation
Using a high-dimensional, non-linear, compliant morphological structure the complex non-linear control is carried out by the structure. All that needs to be done is to identify the linear relationship between the structure and the desired outputs
Give examples of morphological control
- coffee balloon gripper
- a swimming dead trout
- slinky
Detail the principles of swarm robotics
- the swarm can solve complex problems that an individual could not solve
- The swarm is composed of several individuals, some that may be lost or make mistakes but performance is not affected
- Individuals have local sensory information, perform simple actions and have little/no memory
Detail two types of swarm communication
Direct interaction
- touch
- wireless
Stigmergy
- communication through the environment eg pheromones
What is a challenge of swarm intelligence
Difficult to find the rules for the individuals that result in the desired swarm behaviour
List 5 process of swarm intelligence
- Flocking (Reynolds flocking)
- Pathfinding (Ant colony optimisation)
- Optimisation (particle swarm optimisation)
- Decision making
- Clustering
Describe Reynolds flocking
- Separation: Biods maintain a given distance from other biods
- Cohesion: Biod move towards the centre of mass of the neighbouring biods
- Alignment: Biods aline its angle with its neighbours
Describe Ant colony optimisation
- As they move ants deposit a pheromone
- Pheromones decay with time
- Without pheromone, there is an equal probability of choosing both paths
- Ants follow the path with the highest pheromone concentration
- The short path has higher traffic as it takes less time to complete and therefore a higher pheromone concentration
- The short path will increasingly be chosen
Describe Particle swarm optimisation
- Brave: keep moving in the same direction
- Conservative: move towards its own best previous position
- Swarm: move towards the best neighbour
Describe decision making
- Individuals base their decisions on the opinion of their neighbour
- This is based on the quantity and the duration of the opinion of their neighbours
- The opinion is advertised for longer the more correct it is
Describe clustering
- if an object is in an area of low density there is a high probability of it being picked up
- if the individual is in an area of high density there is a high probability of the object being deposited
Give 1 real-life uses of swarms
- Flying robots to produce a dynamic communication
link
Name the two types of reconfigurable robots and give an example for each
Chain: PolyBot
Lattice: A-TRON
Describe morphogenesis
Morphogenesis produces structures without the need of a global map of the structure, instead of relying on self-organisation and emergent shapes.
- morphogenes are stored on every robot. U activates itself and create V.
- V inhibits V and U.
- The reaction then diffuses through the swarm, creating patterns
List design consideration of nanobots
- size
- shape
- charge
- cargo
- material
List the pros and cons of DNA memory
Pros
- Extremely high data density (10^3 times that of an optical disk)
Opertintiy for parallel computing
- Low power
Cons
- Expensive
- Error-prone
- application specific and not easily rewritten
- Not suitable for simple computation
- Difficult to extract the data
how can the travelling salesman problem be solved using DNA
- Each location is given a name in DNA
- Each path is given a name comprised of the end of the first location and the start of the end location.
- This is then mixed in a test tube and every possible solution computed simultaneously
- The stands are then ins[pected to find the short solution which contains all the locations
Name and describe the method of computation using DNA
DNA Toehold system
- an ill-fitting strand is displaced by a strand of the correct length
Describe DNA origami
Through the design of specific sequences, DNA can be made to fold in a meaningful way. Stables are used to guide how the structure folds
Define the two types of creativity
Individualist “little c”
- creativity is a new mental combination that is expressed in the world
Sociocultural “big C”
- Creativity is the generation of a product that is judged to be novel and also appropriate, useful or valuable by a suitably knowledgable social group. novelty isn’t enough
List the creative process
- Find the problem
- Acquire the knowledge
- gather related information
- incubation
- generate ideas
- combine ideas
- Select best ideas
- Externalise ideas
Give an example of artificial creativity
GAN Picture
- Generated fake images which were then inputs into a discriminator which had been taught with a large training set
Describe the challenges of artificial creativity
- It is difficult to define what creativity is
- Representing unknow unknows is difficult
- Potentially, artificial creativity must be embodied
