Background
Before we can speak if machine intelligence we need to talk a little about human intellegence. In our day to day life an othewise mundane task that is performed smartly would immediately earn the performer of the task a tag of intelligence. In some sense intelliegence is relative, contextual and even dependent on the person’s age and other attrributes. For example, if a child is able to open a closed box by unhooking the latch, the child could be attributed as being intelligent, however the same task performed by an adult would not attract any intellectual tag. So in some sense, a task performed by a person when it is not expected of him because of his qualification or his age or any other personal trait is often envisaged as intelligence. Subsequently, the same task can be performed by different humans in different ways – even if they belong to the same “category” - in such an event the most optimized way in which the task is done is often thought of as being more intelligent than the other ways of completing the same task. In effect when everything else is same it is relative else it is a function of the attribute of the hman performing the task. In some sense the definition of intelligence is very vauge and highly subjective and dependent on the relationship between the task and the human performing the task.
If we assume the methodology adopted by the human to complete a task being correlated to intelligence or on other hand a methodology that has high entropy could be considered as being more intelligently done that the methodology that is commonly used (low entropy). Clearly, the tasks that are done easily and almost involuntarily by a human are not easy for mechines to acomplish and when a machine is able to replicate what a human would otherwise do it without blinking an eye could be termed a machine intelligence.
From very early on researchers have been chasing the holy grail of building machines that can in a given scenario and given condition do exactly what a human would do. For example consider the task of opening a door with a key. More or less the machine could be programmed to perform the task of inserting the key in the keyhole and then turing it anticlockwise to open the lock. Now instead of giving one key if a bunch of keys are given to a human, most often the human tries to insert each key and keeps track of which key he used to try with all the keys one by one until he is able to open the door. A scenario change form one key to several key in a bunch could leave a machine in confusion but not a human. Going further if the task has to be performed in the dark, the human immediately uses his other cue, namely the sense of touch to find the key hole and use the sense of touch to move from one key to another to identify which key can open the lock. Clearly this changed scenario of one key -> multiple keys -> multiple keys in darkness does not require any learning by the human, it more or less comes naturaly. However a machine has to be thaught different strategies for different scenarios to achieven the same task in addition one has to make sure there are multiple sensors (camera, touch sensor) which are readily accessible to the machine.
In brief, machine is as intelligenet as a human can make it using his intelligence! So if we can programtically push in the straightforward (to the human) methods into the machine so that the mchine performas more of less the way the human performs in different scenario, the machine is considered to be intelligent. So in some sense a human has to make the machine intelligent and one can make a machine as intelligent as the human making it is and not more!
Very often the machine intelliegence is attributred to the software aspect rather than the hardware. For example, if a walking robot trips and it manages to balance itself without falling there is no acknowledgement of intelligence though there is a significant amount of programing that needs to go into making a robot to balance itself from falling. This can be attributed to the fact that this is a task, namely ability to balance when a person trips is something that comes to human very naturally. So what humans do naturally is somthing that is not attributed to a sign of intelligence when the machine is able to do it.
Making Machines Intelligent
Machines or computers by themselves are not intelligent. It is the human who needs to programatically pour in intelligence. If we suppose that machine intelligence is something that is a measure of how close the machine is able to mimic the human under similar scenarios and conditions, one is left with a question of who is the human reference? To ovecome this one uses terms like average intelligence or IQ of certain number etc – which is still very ambigious. We will not venture into this area!
So what goes into making machines “act” intelligent? Same as what makes human intellegent, learming from own (or others) experiences, learning from own (or others) mistakes. The key is learning followed by the ability to associate the learning even with a new unseen sitiation by mapping it to a similar sitiation that might have been experienced during learning. So the learning is not based on the rotting style specific to the example but is more along the direction of being able to generalize the example.
Step in machine learning algorithms which assist in allowing the machines to function intelligently. A brief taxonomy of the machine learning algorithms is shown in Figure 1.
Figure 1: Brief Taxonomy of Machine Learning.
The machine learning algorithms which make use of annotated (by human!) data (when available) are called supervised learning algorithms and the learing algorithms which do not make use of annotations (mostly because they are not available) are called unsupervised learning algorithms. The choice of the algorithm is depending on the availability of annotated data in most applications and when more than one algorithm can be used, the choice is usually based on which of these algorithms perform better.
Application
Machine intelligence is useful in all walks of life where there is a need for man-machine interaction. Consider a man machine interaction shown in Figure 2. Consider a voice based self help system run by a car insurance company and Let us consider the scenario of a customer calling the self help system to claim insurance because of an accident. Note that the Agent is chirpy when the Customer calls in and the reaction of the self help Agent could be any of the the three, namely, R1, R2 or R3.
Figure 2: Voice Based Man Machine Interaction.
Clearly, R1 is the response of the most self help system today which will in all probability result in a poor customer experience both in terms of the spoken content and the way :-) it is spoken. A speech understanding system probably would respond as R2 where in it would express empathy /Sorry to hear this./ while continuing to be chirpy :-) . While R3 which not only understands what the Customer saying but is able to empathize with the C ustomer communicate back with an emotion :-( in tune with the emotion of the Customer. Clearly, R3 kind of response will need to recognize the emotional state of the customer. So automatic detection of emotions by machine during its spontaneous conversation with a human can lead to better user experience (Ux). The ability to recognize emotion automatically makes the machine intelligent and this is something that is learnt using machine learning algorithms.
Comments