Neural Networks And Artificial Brains Flashcards
What does our brain consist of?
Highly interconnected neuronal networks
How many neurons are there?
100 billion
What is the volume of the brain?
1.5 litres
How much power does the brain consume?
10W
What is 50-80% of total energy consumption in the brain used for?
Conduction of action potentials along nerve fibres and in synaptic transmission
What is the other 20-50% of the brain energy consumption used for?
Manufacturing and maintenance
How do brain transmit signals so quickly despite being so slow?
It is constructed as highly parallel networks, most neurons connect directly to many thousands of others
How does brain form so many connections?
Brain exploits it’s three-dimensional volume to pack everything by bending sheets of cells into folds and weaving connections closely together into bundles
How do silicon neurons overcome the limitations of the 2D nature of chips and circuit boards?
Exporting the very high speed of conventional electronics where the same wire can carry many different messages along the same wire
what has neurally-inspired engineers adopted to reduce power but increase speed?
Using analogue rather than digital coding
How do analogue circuits work?
Instead of coding in 0s and 1s, it codes in continuous changes in voltages
How do neurons compute and make decisions?
They transmit impulses down axons to communicate the answer to target neurons
What is a disadvantage about spike coding?
It is energetically costly
How do efficient coding maximise the information represented in a pattern of spikes?
By reducing redundancy
What is an advantage of sparse coding?
Increasing energy efficiency
What is sparse coding?
Using as small number of active neurons as possible
What is a simple artificial version of a biological network built?
A silicon retina that captures light and adapts its output automatically to changes in overall lighting condition
How does a silicon retina that capture light work?
It connects to two silicon neurons like real neurons in the visual cortex and have the job of extracting information about the angles of lines and contrast boundaries in retinal image
What are neurons in silicon retina prototype called?
Integrate and fire neurons
How do integrate and fire neurons get their name?
They add up the weighted inputs, coded as voltages that are arriving at their synapses and only fire an action potential if the voltage reaches a set threshold
What are silicon neurons like?
They are built of transistors and operate in their sub threshold range and act like cell membranes of real neurons, additional transistors provide active conductance to emulate the voltage and time-dependent current flows of real ion channels
What does artificial neural networks consistent of?
They consist of a number of simple processing units that are highly connected in a network
What is the simplest form of artificial neural network?
Feedforward associator
What does a feedforward associator have?
Layers of interconnected input and output units
How is associative memory encoded by?
Modifying the strengths of connections between layers such that when an input pattern is presented, the stored pattern associated with that pattern is retrieved
What is a more complex artificial neural network?
Recurrent neural net
What does a recurrent neural net consist of?
A single layer where every unit is interconnected and all units act as input and ouput
What does the design of recurrent neural net allows it to do?
Store patterns rather than merely pairs of items
How to decode autoassociative networks?
Recursive search for a stored pattern
In a network of 1000 units, how many patterns can be retrieved before errors in the retrieval patterns become too large?
150
What is the similarity between artificial neural networks and brain?
The way they store and process information
What kind of storage does artificial neural network have?
Content-addressable storage
How are information stored in artificial neural networks?
Information is stored in the weights of connections, the same way that synapses change their strength during learning
What are the learning rules that change artificial neural networks?
They train it by modifying the strength of connections between neurons by taking the output of the network to a given input pattern and comparing it with the desired pattern, the difference which is the error is then used to adjust the weights of the connections to achieve a closer output to the desired one and gradually reduces the error to the minimum
What is the problem of artificial neural networks?
It is simulated mathematically on digital computers and this takes time so artificial neural networks cannot operate in real time
