It can do what?
Apparently the universe thought I was praising Microsoft too much lately and decided to toss me this one.
I don't have anything in general wrong with the tech. It is the point in the article however where it attempts to "blow your mind" by telling you that it can identify which plants are or are not edible when "you're out in the woods".
I'm going to shoot SEVERAL holes in that statement. Normally I wouldn't attack an article like this, but it is from Microsoft's official blog and I feel I have been providing too much bias in their favour of late. So, the idiocy of the statement needs to be addressed.
Firstly is the accuracy. Sure, the tech is more accurate, according to the article, than similar machines out there at the moment. But, as the "Cobberdog" example showed IN A SCRIPTED DEMO, it is still not 100% correct. I'm sorry, but we're a long ways away from me trusting such tech with discerning what foods will or will not kill me. Not a gamble I'm willing to take.
There are plants out there which are related to, and look similar to, perfectly safe to eat plants that aren't. Then, the response needs to be smart enough to tell you which part(s) of the plant are edible. Perhaps the plant typically has berries, and the berries are edible but the rest of the plant isn't. If you stumbled across the plant a time of year when it isn't bearing fruit and the application it tells you it is safe to eat... OOPS!
Then, in some cases the opposite is true. The fruit isn't safe to eat, but the roots or the leaves may be.
This is the holistic problem with attempting to use this sort of tech for ANY life saving measure. Even as a last resort. And, it applies to all sorts of areas of identification. Medicines for example can be placed in the incorrect bottles which would taint any recommendation such tech might give you on the safe usage of particular drugs. And the list goes on.
The next problem is the feasibility of the problem in general. I'm so far "out in the woods" and have been that way for so long that I need to consult my phone on what foods I can safely eat? Honestly, the odds are against you that if you're that deep in the woods for that long that A) your phone battery is still alive and B) that you actually get a cell signal.
IF I were that far out in the woods for long enough that I'm resorting to consuming unknown substances AND I still have an operational phone with a data connection... guess what? I'm gonna pull up the GPS, find out where the hell I am and call for help. I'm sure you could construct a scenario where for some reason it made sense. But it would be DAMN contrived.
So, even if you know the software is reliable enough for your purposes and you freakishly find yourself in a some strange scenario where attempting this made ANY sense. You'd still need to be concerned about battery life and have an available data connection. And that makes the feasibility of this already outlandish suggestion EVEN worse.
Not to mention, the degree of variation between established dog breeds and plants is in favour of dog breeds. There are far fewer breeds of dogs than there are of plants. It may be able to accurately label the family of plant. But, out in unknown territory you're more likely to get a false positive against whatever the most common member of a plant family is than the one you're looking at. Accurate identification also requires that service has ample photographic data available on the plant you're investigating. Dogs and cats of all breed are VERY commonly the subjects of pictures. My brother has posted more pictures of his dog online than I have of my daughter.
Which highlights yet another fault in the solution sets this is viable for. The data will intrinsically be most accurate about the most trendy/popular subjects. And those just happen to be the subjects that as a general rule of thumb, we need assistance in identifying.
I get it. It's just a tech demo, and you're happy about what you've produced and the possibilities. But, in my opinion, making up TERRIBLE and contrived use cases makes a worse impression on the software than a more trivial, but wholly viable one.
The dog example was actually good. There is a lot of pictographic data about dogs, but the average person has very little knowledge of things like different breeds and it is, while not at all important, a topic which arises from time to time. In this case it works out, because the high number of fanatics supplying photographic evidence about the subject. One dog lover can post hundreds or thousands of photos of dogs of a single, or even multiple breeds. This means relatively few people are needed to supply a sufficient number of photos to train such a service to identify them vs. those that may wish to consume the knowledge.
While there are certainly plat fanatics out there as well. From my experience there are far fewer of them than dog fanatics and far more plants out there to document than breeds of dog. And, beyond that, people are needed to associate the right pictures with the right plants and the right data about them. Not everyone taking pictures of varying forms of fichus take the time to even point out the name of the plant, let alone whether or not it will kill you.
I could go on. But I won't.
I don't have anything in general wrong with the tech. It is the point in the article however where it attempts to "blow your mind" by telling you that it can identify which plants are or are not edible when "you're out in the woods".
I'm going to shoot SEVERAL holes in that statement. Normally I wouldn't attack an article like this, but it is from Microsoft's official blog and I feel I have been providing too much bias in their favour of late. So, the idiocy of the statement needs to be addressed.
Firstly is the accuracy. Sure, the tech is more accurate, according to the article, than similar machines out there at the moment. But, as the "Cobberdog" example showed IN A SCRIPTED DEMO, it is still not 100% correct. I'm sorry, but we're a long ways away from me trusting such tech with discerning what foods will or will not kill me. Not a gamble I'm willing to take.
There are plants out there which are related to, and look similar to, perfectly safe to eat plants that aren't. Then, the response needs to be smart enough to tell you which part(s) of the plant are edible. Perhaps the plant typically has berries, and the berries are edible but the rest of the plant isn't. If you stumbled across the plant a time of year when it isn't bearing fruit and the application it tells you it is safe to eat... OOPS!
Then, in some cases the opposite is true. The fruit isn't safe to eat, but the roots or the leaves may be.
This is the holistic problem with attempting to use this sort of tech for ANY life saving measure. Even as a last resort. And, it applies to all sorts of areas of identification. Medicines for example can be placed in the incorrect bottles which would taint any recommendation such tech might give you on the safe usage of particular drugs. And the list goes on.
The next problem is the feasibility of the problem in general. I'm so far "out in the woods" and have been that way for so long that I need to consult my phone on what foods I can safely eat? Honestly, the odds are against you that if you're that deep in the woods for that long that A) your phone battery is still alive and B) that you actually get a cell signal.
IF I were that far out in the woods for long enough that I'm resorting to consuming unknown substances AND I still have an operational phone with a data connection... guess what? I'm gonna pull up the GPS, find out where the hell I am and call for help. I'm sure you could construct a scenario where for some reason it made sense. But it would be DAMN contrived.
So, even if you know the software is reliable enough for your purposes and you freakishly find yourself in a some strange scenario where attempting this made ANY sense. You'd still need to be concerned about battery life and have an available data connection. And that makes the feasibility of this already outlandish suggestion EVEN worse.
Not to mention, the degree of variation between established dog breeds and plants is in favour of dog breeds. There are far fewer breeds of dogs than there are of plants. It may be able to accurately label the family of plant. But, out in unknown territory you're more likely to get a false positive against whatever the most common member of a plant family is than the one you're looking at. Accurate identification also requires that service has ample photographic data available on the plant you're investigating. Dogs and cats of all breed are VERY commonly the subjects of pictures. My brother has posted more pictures of his dog online than I have of my daughter.
Which highlights yet another fault in the solution sets this is viable for. The data will intrinsically be most accurate about the most trendy/popular subjects. And those just happen to be the subjects that as a general rule of thumb, we need assistance in identifying.
I get it. It's just a tech demo, and you're happy about what you've produced and the possibilities. But, in my opinion, making up TERRIBLE and contrived use cases makes a worse impression on the software than a more trivial, but wholly viable one.
The dog example was actually good. There is a lot of pictographic data about dogs, but the average person has very little knowledge of things like different breeds and it is, while not at all important, a topic which arises from time to time. In this case it works out, because the high number of fanatics supplying photographic evidence about the subject. One dog lover can post hundreds or thousands of photos of dogs of a single, or even multiple breeds. This means relatively few people are needed to supply a sufficient number of photos to train such a service to identify them vs. those that may wish to consume the knowledge.
While there are certainly plat fanatics out there as well. From my experience there are far fewer of them than dog fanatics and far more plants out there to document than breeds of dog. And, beyond that, people are needed to associate the right pictures with the right plants and the right data about them. Not everyone taking pictures of varying forms of fichus take the time to even point out the name of the plant, let alone whether or not it will kill you.
I could go on. But I won't.
Comments
Post a Comment