Please enable JavaScript.
Coggle requires JavaScript to display documents.
On Intelligence - Coggle Diagram
On Intelligence
Ch 3. The Human Brain
-
Many parts of brain are essential to human, but being human and being intelligent are different things
Early neuroscientists believed that cortex consist of many functional areas that semi-independent and specialized for spectific function
Functional regions can be higher or lower in the hierarchical operations, regardless the physical arrangements in the brain
Different senses share similar hierarchical processing, primary sensory corteies receive more signals from higher regions of cortex than the sensory organs
Synapses can be excitatory or inhibitory, regardless of the details of the synapses, the formation and strengthening (weaking) or synapses enables memories to be stored
Pyramidal neurons consist 80% of neocortex, each pyramidal neuron has more than 1000 synapses
Vernon Mountcastle (1978) published a paper titled 'An Organizing Principles for General Function', suggested that neocortex is similar in execution, and all different functional regions develop from similar initial conditions. Thus the neocortex is the universal machine that the brain utilizes
Problems with brain imaging: static picture of the brain, largely ignoring the dynamic properties.
Neocortex are highly plastic. Many rewiring experiments suggest that neocortex are inter-changable for differnt functions if rewired before the 'critical period'
Neocortex are space-hunger. Born blind or deaf people, their deprived regions are normally taken by other brain funacitons.
The input of different sensory information is largely similar. A combination of spatial and temporal spikes. The brain did not operate with pre-defined schemes but rather only on the data and gradually adapt to electrical activity firing in pattens
-
-
:warning: The examples of flexible neocortex are not appropriate. The rubber hand illusion is a binding problem, the visual cue of watching the rubber hand being flushed coincide with the actual flushing of a real hand. The illusion is not permanent and does not linger. The sensory substitution does not quite achieve the function of normal vision, the participant claimed he only uses the function as very rough guide in climbing.
:warning: Another very important issue the author does not address is the processing of signals. The neocortex received the information and send them to somewhere to provide perceptions. How do cortexes initially learn to convert the signals into meaningful subjects? The initial learning is largely unsupervised. Even if we take neocortex as the seat for universal function, the training and modifications are still unknown. Later, humans learn many things through education, this learning is largely supervised, they have a different requirement and will bring different changes to the cortex.
Ch 4. Memory
The world is an ocean of constantly changing patterns that
come lapping and crashing into your brain: Information Overflow
Human perform significant tasks within a second, less than 100 steps of operation, same as animals
:red_flag: Brain does not compute answers, but search for solutions in the database
-
:warning: Those are expectations of neocortex's functions. The storage of patterns and invariant forms is difficult to achieve at least to the current coding scheme. The hierarchical physical structure does not guarantee content structures
:warning: There is a way to achieve hierarchical structure storage: the recall process will push existing knowledge to a higher level of neurons, the further the memory goes, the higher of the hierarchy. I can code this
:warning: There is also a way to code invariant form if there if forgetting and layered re-coding. By layered re-coding, multiple memories can be stored in similar pattern space and they may overlap, while non-overlapping features will be forgetten. I can code this in the next project
:warning:Self-completion may be a side effect of auto-associative memory. The recall of friend's face may have different reasons other than the invariant form
:red_flag: The mere appearance of your friend forces
your brain to start recalling patterns associated with her
:warning: To achieve this face/music recall, a small trick is to add multiple dimensions as input to the original input, other than the location in the retina, including the distances, shapes, and lengths. This has been adopted by current facial/sound recognition methods
:warning: If the memory system operates as blind-folded as our suggestion, there should be an intelligent system that utilize this memory system, the real seat of intelligence.
Some set of cells in the brain remain active whenever recall happens - cell assembly. This is not stability of cell firing, cell assemblies can be unstable or firing at different pattern
The cortex takes the detailed, highly specific
input and converts it to an invariant form. However it is not always necessary for the brain to operate to the final form. Along the way, some decisions can be made. We may operate without reasoning. The reasoning and logic part is the key here.
Ch 1. History
The Author sets sight on the neocortex, the most recently developed part of the mammalian brain, and believed to be the seat of intelligence.
MIT approach:end-justify-the-means. Build computer programs that would match and then surpass human abilities.
Turing's test: if a computer can fool a human interrogator into thinking that it is a human, then the computer must be intelligent.
Warrent McCulloch and Walter Pitts(1943) proposed neurons could perform digital functions and connected together to perform logic operations. The brain could be built with the same analogy of wiring of digital electronic circuits. The brain could ba another kind of computer.
Eliza, a chat bot to rephrase the questions and try to fool people, failed to pass Turing test.
Block World, a close world of several boxes, the program can understand and answer questions on the relationships among the compounents.
Expert systems: data base of facts that could answer questions in a parcific domain.They played the game while not understanding the game, same as calculators perform computation without understanding the mathematics.
-
Deep Blue: supercomputer program, beats human champion on chess, Alpha-go, beats human on go.
None of the researchers in cognitive science talked about how the structure of the brain support the cognitive functions and how their theories could be implemented.
Ch 2. Neural Networks
Neural networks: the connectionist approach that based loostly on the real nervous systems, learning by back propagation, mostly three layers, but new models do exist
Verdict to neural networks: 1. brain processes dynamic steams of information, but not static input; 2. brain relies heavily on feedbacks, the neocortex has 10 times of feedback connections than feedforward connections to thalamus; 3. the model cannot explain the brain hierarchical structures (Deep learning kind of do, but rarely)
Act intelligently does not represent intelligent, a intelligent identity can be intelligent while in idle
Auto-associative memory, the Hopfield: 1. partial completion; 2. can store sequences of patterns or temporal patterns
-
-
-
-