Some Basic Notes/Discussions on Jeff Hawkin's book On Intelligence:
Jeff says that fundamentally the present artificial intelligence technologies are different from how mammalian brain works. He says that we should at least get to know some basic working principles of brain to implement them on machine. But people have been doing the work without studying the brain. All AI technologies have failed miserably in doing tasks which can be done by humans without any effort. He is defining the term 'INTELLIGENCE' in his book. Academic world has been saying that our brains are too complicated and its almost impossible to understand them. But he seems not convinced by that and he gave few examples from the past when people didn't accept the idea that Earth is spherical in shape and few sections of the society took it on to heart and started persecuting!! He says complexity is a symptom of confusion not a cause. Yes I agree that!! Its obvious that when something as complex as brain is what you are supposed to decipher confusion inherently exists. He point outs that what is wrong in Chinese Room experiment! There the person inside basically doesn't understand chinese but still he is able to do the work. But our brains are not like that they recognize patterns and we understand what it is. Inherently the way AI and Brains work are different. Ability to make predictions about future is the crux of intelligence. What he says is that AI proponents saw parallel between computation and thinking.
A human need not do anything to understand a story. One can read a story and it can be clear and can be understood by him/her. But others cannot tell from ones quiet behavior whether he/she understood or not or even if he/she knows the language. Others can ask the person who read the story whether he/she understood the story or not. But that persons understanding occurred when he read the story not just when he answered the questions.
He says Brain and AI differ on few things they are:
-->Temporal Concept in Brain
-->Importance of feedback(Note that back propagation in Neural Networks isn't feedback. It doesn't happen when some output is required back propagation happens in input stage that is during training but in brain it happens during output also)
Neural networks have no sense of time nor history. They have standard output for standard input.
AI and Neural Net communities concentration was only on behavior/output of the system. Jeff argues that behavior doesn't measure intelligence one can be intelligent at the same time he can sit idle and think about or form solutions in his brain, this doesnt mean that he is not intelligent. So our concept was intelligence was not complete, AI was not successful because of this reason.
Intelligence is not just matter of acting or behaving intelligent. Behavior is manifestation of intelligence but not central character or primary definition of being intelligent.
He argues we can create mindful machines. He brings in functionalism to explain that. He says that" According to functionalism being intelligent or having mind is purely a property of organization and has nothing inherently to do with what you are organized out of. A mind exists in any system whose constituent parts have the right causal relationship with each other but those parts can just as validly be neurons silicon chips or something else.
Neo cortex which is the seat of intelligence has 6 layers. Its new brain, which means it evolved recently and thats why mammals have different cognitive abilities than reptiles which do not have neo cortex. Medulla oblangata, cerebellum, thalamus all are parts of old brain and mostly they are involved in motor, behavioral, emotional matters. Neo cortex is involved with memory, intelligence and assisting the old brain. Cells in our brain create the mind is fact, not a hypothesis.
Sight, hearing, touch seem very different but the way the cortex processes signals from ear is same as the way it processes signals from eyes. Cortex does something universal that can be applied to any type of sensory or motor system.
Consider the fact that we have special visual areas that seems to be specifically devoted to represent written letters and digits. This doesnt mean that we are ready to process letters and digits as soon as we are born. It all depends on the environment and how the brain trains itself. Any brain if put in right environment can learn any number of languages be it spoken, sign, musical , mathematical.
He says that there is some single powerful algorithm being implemented in every part of cortex in a suitable hierarchy which is provided with input from environment. So he argues that there is no reason for intelligent machines of future to have same senses or capabilities as we. The cortical algo can be implemented in novel ways so that genuine flexible intelligence emerges outside of bio brains.
He says Turing went wrong in saying behavior is proof of intelligence.
The four attributes of neocortical memory that are fundamentally different from computer memory are:
--> Neo cortex stores sequences of patterns.
--> Neo cortex recalls patterns auto-associatively.
-->Neo cortec stores patterns in an invariant form.
--> Neo cortex stores patterns in hierarchy.
Let me explain some terms here.
Auto-associativity: Its a kind of memory which can retrieved fully when a small portion of its given as input. Perfect example will be a camouflaged soldier in photograph. At first sight you wont notice him but when told to look for soldier your brain starts looking for patterns associated with it then you recognize there is a soldier in the picture.
Invariant form: When your think of human face what you get is certain picture of what is face? How it is? You wont get any particular face not yours or your girl friends!! You get a picture which has 2 eyes a nose and a pair of ears. If you notice a face with something missing or something is extra or even something out of place than that is exception. You have an invariant picturisation of objects in your brain.
Hierarchy: As already discussed Neo cortex has 6 layers these are organized in hierarchy. If you go down the hierarchy thee are particular neurons trained to detect a particular patterns if the input from lower levels matches the trained condition of neurons they fire further taking to upper levels. So if you go to lower levels basically you will see some seemingly random spikes which cannot make any sense, at this level all information like touch, olfactory, sight, hearing all are in same form!! spikes!! processed similarly for patterns. In the upper layers its recognized if the picture you are seeing belongs to a tiger or a lion!!
Earlier power of computer was compared using CLOCK SPEED, now a days we do it by comparing RAM, NO OF CORES. I hope someday in future, not very far we will do the same by using term EXPERIENCE. Per se you can just say hey! my computer has got XXX years of experience(data), so it knows the surroundings very well and perform tasks quicker!!
To know more about Jeff's work go through Numenta started by him. They developed some algorithms which have concept of time.
Jeff says that fundamentally the present artificial intelligence technologies are different from how mammalian brain works. He says that we should at least get to know some basic working principles of brain to implement them on machine. But people have been doing the work without studying the brain. All AI technologies have failed miserably in doing tasks which can be done by humans without any effort. He is defining the term 'INTELLIGENCE' in his book. Academic world has been saying that our brains are too complicated and its almost impossible to understand them. But he seems not convinced by that and he gave few examples from the past when people didn't accept the idea that Earth is spherical in shape and few sections of the society took it on to heart and started persecuting!! He says complexity is a symptom of confusion not a cause. Yes I agree that!! Its obvious that when something as complex as brain is what you are supposed to decipher confusion inherently exists. He point outs that what is wrong in Chinese Room experiment! There the person inside basically doesn't understand chinese but still he is able to do the work. But our brains are not like that they recognize patterns and we understand what it is. Inherently the way AI and Brains work are different. Ability to make predictions about future is the crux of intelligence. What he says is that AI proponents saw parallel between computation and thinking.
A human need not do anything to understand a story. One can read a story and it can be clear and can be understood by him/her. But others cannot tell from ones quiet behavior whether he/she understood or not or even if he/she knows the language. Others can ask the person who read the story whether he/she understood the story or not. But that persons understanding occurred when he read the story not just when he answered the questions.
He says Brain and AI differ on few things they are:
-->Temporal Concept in Brain
-->Importance of feedback(Note that back propagation in Neural Networks isn't feedback. It doesn't happen when some output is required back propagation happens in input stage that is during training but in brain it happens during output also)
Neural networks have no sense of time nor history. They have standard output for standard input.
AI and Neural Net communities concentration was only on behavior/output of the system. Jeff argues that behavior doesn't measure intelligence one can be intelligent at the same time he can sit idle and think about or form solutions in his brain, this doesnt mean that he is not intelligent. So our concept was intelligence was not complete, AI was not successful because of this reason.
Intelligence is not just matter of acting or behaving intelligent. Behavior is manifestation of intelligence but not central character or primary definition of being intelligent.
He argues we can create mindful machines. He brings in functionalism to explain that. He says that" According to functionalism being intelligent or having mind is purely a property of organization and has nothing inherently to do with what you are organized out of. A mind exists in any system whose constituent parts have the right causal relationship with each other but those parts can just as validly be neurons silicon chips or something else.
Neo cortex which is the seat of intelligence has 6 layers. Its new brain, which means it evolved recently and thats why mammals have different cognitive abilities than reptiles which do not have neo cortex. Medulla oblangata, cerebellum, thalamus all are parts of old brain and mostly they are involved in motor, behavioral, emotional matters. Neo cortex is involved with memory, intelligence and assisting the old brain. Cells in our brain create the mind is fact, not a hypothesis.
Sight, hearing, touch seem very different but the way the cortex processes signals from ear is same as the way it processes signals from eyes. Cortex does something universal that can be applied to any type of sensory or motor system.
Consider the fact that we have special visual areas that seems to be specifically devoted to represent written letters and digits. This doesnt mean that we are ready to process letters and digits as soon as we are born. It all depends on the environment and how the brain trains itself. Any brain if put in right environment can learn any number of languages be it spoken, sign, musical , mathematical.
He says that there is some single powerful algorithm being implemented in every part of cortex in a suitable hierarchy which is provided with input from environment. So he argues that there is no reason for intelligent machines of future to have same senses or capabilities as we. The cortical algo can be implemented in novel ways so that genuine flexible intelligence emerges outside of bio brains.
He says Turing went wrong in saying behavior is proof of intelligence.
The four attributes of neocortical memory that are fundamentally different from computer memory are:
--> Neo cortex stores sequences of patterns.
--> Neo cortex recalls patterns auto-associatively.
-->Neo cortec stores patterns in an invariant form.
--> Neo cortex stores patterns in hierarchy.
Let me explain some terms here.
Auto-associativity: Its a kind of memory which can retrieved fully when a small portion of its given as input. Perfect example will be a camouflaged soldier in photograph. At first sight you wont notice him but when told to look for soldier your brain starts looking for patterns associated with it then you recognize there is a soldier in the picture.
Invariant form: When your think of human face what you get is certain picture of what is face? How it is? You wont get any particular face not yours or your girl friends!! You get a picture which has 2 eyes a nose and a pair of ears. If you notice a face with something missing or something is extra or even something out of place than that is exception. You have an invariant picturisation of objects in your brain.
Hierarchy: As already discussed Neo cortex has 6 layers these are organized in hierarchy. If you go down the hierarchy thee are particular neurons trained to detect a particular patterns if the input from lower levels matches the trained condition of neurons they fire further taking to upper levels. So if you go to lower levels basically you will see some seemingly random spikes which cannot make any sense, at this level all information like touch, olfactory, sight, hearing all are in same form!! spikes!! processed similarly for patterns. In the upper layers its recognized if the picture you are seeing belongs to a tiger or a lion!!
Earlier power of computer was compared using CLOCK SPEED, now a days we do it by comparing RAM, NO OF CORES. I hope someday in future, not very far we will do the same by using term EXPERIENCE. Per se you can just say hey! my computer has got XXX years of experience(data), so it knows the surroundings very well and perform tasks quicker!!
To know more about Jeff's work go through Numenta started by him. They developed some algorithms which have concept of time.
No comments:
Post a Comment