4 ms·

But isn't this just what ML should be good at - taking a huge number of data points and finding the patterns? Or are you saying it's not an issue of ML working

But isn't this just what ML should be good at - taking a huge number of data points and finding the patterns? Or are you saying it's not an issue of ML working poorly but rather one of there not being a good enough data set to train it on properly?

tomrod 5y ago
The gap is interpretation and application of those patterns. Building expert systems is expensive, but ML hits the low hanging fruit of showing patterns to experts out of the park.
ethbr0 5y ago
ML should be good at drawing basic conclusions. End users are misunderstanding the boundary between basic and advanced. Or, to put it another way, everyone agrees there's a difference in value and quality of output between an analyst with 1 year & 10 years of experience, right? So why are we treating ML like it should be able to solve both sorts of problems equally easily? I have faith it will get there. But it's not there yet, in a general purpose way.
- mcguire 5y ago
  Because people like Hinton are outright saying that it already is there.
azalemeth 5y ago
One of the main tasks that doctors do is take patient's vague and non-specific problems, build up a rapport with them, understand what is normal and what is not, dealing with the "irrelevant" information present at the time, and focus the results into a finite tree of possibilities. In principle, this would be a great task for an ML algorithm. It's all conditional probability. But every such system has failed to do that well -- because the "gut feeling" the doctor develops is funded by a whole host of prior information that an ML algorithm won't be trained on: what is "normal" for an IV-drug addict, or a patient with psychosis; how significant is the "I'm a bit confused" in the middle-aged man who was cycling and came off his bike and hit his head? Do the burns on the soles of this child's feet sit consistently with the story of someone who ran over a fire-pit barbecue "for fun", or is it a non-accidental injury? It's a world of shades of grey where, if you call it the wrong way, ultimately someone could die. Doctors do Bayesian maths, and they do it with priors coming from both their own personal experience as a member of society, and professional training. That is, in my ignorant opinion, the main distinction between what I do -- oft-called "academic medicine" or "academic radiology" -- and clinical medicine. The former looks at populations. The latter looks at individuals. In other words, I don't think it's even possible to codify what the data the ML algorithms should be trained on -- they're culturally specific down to the level of an individual town in some sense; and require looking at huge populations at others.
- version_five 5y ago
  > In other words, I don't think it's even possible to codify what the data the ML algorithms should be trained on -- they're culturally specific down to the level of an individual town in some sense; and require looking at huge populations at others. This is inciteful. ML is by definition generalizing, and should only be used where it's ok to generalize. There is an implicit assumption in use cases like medical diagnosis that there is a latent representation of the condition that had a much lower dimensionality than the data and that the model is trained in such a way that there are no shortcuts to generalization that miss out on any information that may be important. The second condition is the hardest to meet I believe, because even if a model could take in lots of outside factors, it probably doesn't need to to do really well in training and validation, so it doesn't. The result is models that generalize, as you say, to the population instead of the individual, and end up throwing away vital context to the personal case. I also believe this is an important consideration for many other ML applications. For example those models that predict recidivism rates. I'm sure its possible to build an accurate one, but almost certainly these models stereotypes in the way I mention above, and do not actually take the individual case into account, making them unfair to use on actual people.
- tomrod 5y ago
  On codification: I actually disagree with this, but only slightly. Imagine if instead of face to face, doctor transactions were via text. The questioning of the doctor can be monitored and patterns in decision trees observed could be codified, and weighed against the healthcare outcome (however defined). What is missing, however, is the Counterfactual Reasoning. The "why" matters. The machine cannot reduce the doctor's choice of decision trees from all possible combinations, only that which it observes the doctor perform. Tail-cases like rare genetic disorders would often be missed.
  - medvezhenok 5y ago
    Tail-case like rare genetic disorders are often missed by doctors too. I have several friends who had Lyme disease with fairly serious complications (in the Northeast) (Not that Lyme disease is that rare - it's actually much more common than is expected). Each of them got misdiagnosed for multiple years by multiple different doctors until finally getting the correct diagnosis/treatment. So every system is fallible.
jjoonathan 5y ago
Right, but "the training data is bad" is a very ML centric way of looking at the issue. It pushes all the difficult parts of the problem into the "data prep" sphere of responsibility.
- srean 5y ago
  Note that there are different ways in which data can be bad (i) image resolution not good enough, too many artifacts and noise (ii) its woefully incomplete, doctors collect and use information from other channels that aren't even in the image, regular conversations, sizing up the patient, if the doctor knows the patient for a long time then a sense of what is not normal for the patient given his/her history etc., etc. Some of the issues that have been discussed in the thread can be incorporated in to a Bayesian prior for the patient, but there is still this incompleteness issue to deal with.
  - jjoonathan 5y ago
    The first step would be to build an information collection pipeline that is in the same league as the doctors. That alone will be a monumental effort because doctors have shared human experiences to draw from and they are allowed to iteratively collect information. I'm just complaining that it seems fantastically reductive to call the absence of such a pipeline "bad data" because developing such a pipeline would be a thousand times the effort of implementing an image detection model. Maybe a million times. It will require either NLP like none we have seen before or an expert system with so much buy-in from the experts and investors that it survives the thousand rounds of iterative improvement it needs to address 99% of the requirements. Comparing issues like low resolution and noise to such a development effort seems like comparing apples to... jet fighters.
- awillen 5y ago
  How else would you describe the issue?
  - jjoonathan 5y ago
    Structural. The problem hasn't even been correctly formulated yet -- and it will take an enormous amount of work to do so.
jofer 5y ago
"There's not enough good data to train it on properly" Bingo: You're in _very_ data poor environment, compared to something like predicting a consumer's preferences in videos or identifying and segmenting a bicycle. The external data is also very qualitative and hard to encode into meaningful features.

tomrod 5y ago

The gap is interpretation and application of those patterns. Building expert systems is expensive, but ML hits the low hanging fruit of showing patterns to experts out of the park.

ethbr0 5y ago

ML should be good at drawing basic conclusions. End users are misunderstanding the boundary between basic and advanced. Or, to put it another way, everyone agrees there's a difference in value and quality of output between an analyst with 1 year & 10 years of experience, right? So why are we treating ML like it should be able to solve both sorts of problems equally easily? I have faith it will get there. But it's not there yet, in a general purpose way.

mcguire 5y ago
Because people like Hinton are outright saying that it already is there.

azalemeth 5y ago

One of the main tasks that doctors do is take patient's vague and non-specific problems, build up a rapport with them, understand what is normal and what is not, dealing with the "irrelevant" information present at the time, and focus the results into a finite tree of possibilities. In principle, this would be a great task for an ML algorithm. It's all conditional probability. But every such system has failed to do that well -- because the "gut feeling" the doctor develops is funded by a whole host of prior information that an ML algorithm won't be trained on: what is "normal" for an IV-drug addict, or a patient with psychosis; how significant is the "I'm a bit confused" in the middle-aged man who was cycling and came off his bike and hit his head? Do the burns on the soles of this child's feet sit consistently with the story of someone who ran over a fire-pit barbecue "for fun", or is it a non-accidental injury? It's a world of shades of grey where, if you call it the wrong way, ultimately someone could die. Doctors do Bayesian maths, and they do it with priors coming from both their own personal experience as a member of society, and professional training. That is, in my ignorant opinion, the main distinction between what I do -- oft-called "academic medicine" or "academic radiology" -- and clinical medicine. The former looks at populations. The latter looks at individuals. In other words, I don't think it's even possible to codify what the data the ML algorithms should be trained on -- they're culturally specific down to the level of an individual town in some sense; and require looking at huge populations at others.

version_five 5y ago
> In other words, I don't think it's even possible to codify what the data the ML algorithms should be trained on -- they're culturally specific down to the level of an individual town in some sense; and require looking at huge populations at others. This is inciteful. ML is by definition generalizing, and should only be used where it's ok to generalize. There is an implicit assumption in use cases like medical diagnosis that there is a latent representation of the condition that had a much lower dimensionality than the data and that the model is trained in such a way that there are no shortcuts to generalization that miss out on any information that may be important. The second condition is the hardest to meet I believe, because even if a model could take in lots of outside factors, it probably doesn't need to to do really well in training and validation, so it doesn't. The result is models that generalize, as you say, to the population instead of the individual, and end up throwing away vital context to the personal case. I also believe this is an important consideration for many other ML applications. For example those models that predict recidivism rates. I'm sure its possible to build an accurate one, but almost certainly these models stereotypes in the way I mention above, and do not actually take the individual case into account, making them unfair to use on actual people.
tomrod 5y ago
On codification: I actually disagree with this, but only slightly. Imagine if instead of face to face, doctor transactions were via text. The questioning of the doctor can be monitored and patterns in decision trees observed could be codified, and weighed against the healthcare outcome (however defined). What is missing, however, is the Counterfactual Reasoning. The "why" matters. The machine cannot reduce the doctor's choice of decision trees from all possible combinations, only that which it observes the doctor perform. Tail-cases like rare genetic disorders would often be missed.
- medvezhenok 5y ago
  Tail-case like rare genetic disorders are often missed by doctors too. I have several friends who had Lyme disease with fairly serious complications (in the Northeast) (Not that Lyme disease is that rare - it's actually much more common than is expected). Each of them got misdiagnosed for multiple years by multiple different doctors until finally getting the correct diagnosis/treatment. So every system is fallible.

jjoonathan 5y ago

Right, but "the training data is bad" is a very ML centric way of looking at the issue. It pushes all the difficult parts of the problem into the "data prep" sphere of responsibility.

srean 5y ago
Note that there are different ways in which data can be bad (i) image resolution not good enough, too many artifacts and noise (ii) its woefully incomplete, doctors collect and use information from other channels that aren't even in the image, regular conversations, sizing up the patient, if the doctor knows the patient for a long time then a sense of what is not normal for the patient given his/her history etc., etc. Some of the issues that have been discussed in the thread can be incorporated in to a Bayesian prior for the patient, but there is still this incompleteness issue to deal with.
- jjoonathan 5y ago
  The first step would be to build an information collection pipeline that is in the same league as the doctors. That alone will be a monumental effort because doctors have shared human experiences to draw from and they are allowed to iteratively collect information. I'm just complaining that it seems fantastically reductive to call the absence of such a pipeline "bad data" because developing such a pipeline would be a thousand times the effort of implementing an image detection model. Maybe a million times. It will require either NLP like none we have seen before or an expert system with so much buy-in from the experts and investors that it survives the thousand rounds of iterative improvement it needs to address 99% of the requirements. Comparing issues like low resolution and noise to such a development effort seems like comparing apples to... jet fighters.
awillen 5y ago
How else would you describe the issue?
- jjoonathan 5y ago
  Structural. The problem hasn't even been correctly formulated yet -- and it will take an enormous amount of work to do so.

jofer 5y ago

"There's not enough good data to train it on properly" Bingo: You're in _very_ data poor environment, compared to something like predicting a consumer's preferences in videos or identifying and segmenting a bicycle. The external data is also very qualitative and hard to encode into meaningful features.