BenLynch29
Well-known member
There are two main areas that need reforming at the SFA above all others in my opinion, and I believe there is a consensus among all 12 Premeirship clubs that reform is needed despite the current lack of consensus on the way forward:
(1) the performance / standard / level of officiating, and
(2) the compliance officer / appeal process.
I've already put forth my thoughts on #2 in a separate thread a couple weeks ago so I won't repeat myself here.
With regards to fixing the matchday performance of the referees and assistants, there have been a lot of suggestions, and most of them good ones. Foreign referees, making them full time, making them declare which team they support, providing better training, giving them access to VAR, among other ideas. I'm going to suggest one more idea that is a bit radical but one that I think would actually provide a better solution (as defined by fewer mistakes and much more consistency at the same time).
We live in an unprecedented era of technological progress. The use of computers in the last few decades has ushered in unparalleled productivity gains across a wide swath of the global economy, and recent advances in adjacent fields of AI, Machine Learning, and High Performance Computing (HPC) are being utilized in new and original ways to improve living standards and complete tasks with precision, speed, and accuracy not previously possible. These tools are only recently -- within the last 5 years -- being used in the field of athletics and analysis thereof, but the long term potential to replace match officials, or even to compliment them in the same way goal line technology has been utilized, is very real. We should embrace it and encourage the SFA to be on the leading edge of adoption even if such applications aren't yet perfect. Here's how it would work:
One of the biggest machine learning tasks today is image classification. In layman's terms, there's a computer model where the computer is fed a series of images that are tagged with classification data. For purposes of this discussion, all the possible classification categories are covered. The computer basically "learns" to associate certain aspects of the images to the classification categories. Then the model is fed additional images without tags and it is asked to classify the images based on what it has learned during the previous training step. This process has been used in applications such as self driving vehicles (recognizing real world objects on the road), medicine (classifying medical scans as cancerous/non-cancerous), converting scanned PDF document images into text, and many others. It has also been used to analyze sporting participants (based on their uniform number) and events (based on a sequence of images - aka a video feed). See the links below for examples. It should be noted that well constructed models are currently generating accuracy >99% when the scope of the problem is narrow, and errors in models have been reducing by factors approaching an order of magnitude each year! Even models that look at 'noisy' data and are not well defined (e.g. stock price prediction) are performing significantly better than human experts.
How this can be applied in soccer/football is to set up multiple high resolution cameras, ideally 4K or better, and feed them into a model in a synchronized way so the model can create a 3-D representation of the environment and detect the same object from multiple angles. From there we train it to classify certain events based on rules (offsides) or based on consensus judgement (violent conduct) of prior video input samples. From here, we create the output classifications based on the laws of the game. Train the model to identify: handball, ordinary foul, professional foul, violent conduct, offsides, ball out of play, ball across the goal line, etc. Basically anything that a referee would need to call today. So if the model sees a foul it could spit out and tell us that (1) a foul occurred, (2) who committed the foul against whom, (3) where on the field the foul occurred (as specified in x,y coordinates, (4) type of foul (tripping, elbow, shirt pulling, etc), and what the referee determination should be (free kick, penalty, yellow card, red card, warning, etc) based on consistent application of the decision in the training sample. Once the model can do all this, we tell it to alert an actual official on the field if certain parameters are met, similar to how goal line technology today sends an electronic signal to a watch the referee is wearing in some leagues. Note that not every detection of an event will warrant an on-filed call -- the model may detect a player in an offsides position based on x,y coordinates of all players and the ball, but if that player makes no movement toward the ball or is otherwise involved in the play, play should continue as it does today.
For purposes of model training, matches from other leagues are ideal because (1) there are many more examples across the globe for each possible classification category leading to better model accuracy, and (2) if SFA officials are the ones selected to judge the training data they won't even have the opportunity to exhibit bias for or against any Scottish club. Ideally the sample data would be split based on hard rule violations (offsides, ball out of play) that could be explicitly programmed vs judgement calls (fouls and appropriate penalty for fouls). Additionally, I would recommend that determination of fouls in the training set be exclusively outside the penalty box so that we don't train the model that certain fouls are acceptable inside the box but not outside. A foul is a foul anywhere on the field, and the consequence of the foul (free kick vs. penalty) is determined by x,y coordinates relative to the penalty box. The same would apply to red/yellow card to the offender.
The appeal of such a system is obvious (near 100% accuracy, eventually, and zero bias toward or against any team. Rangers would know Celtic isn't being favored, Celtic know that Rangers aren't being favored, and the rest of the league know that there isn't one set of rules for the Glasgow giants and another set of rules for everyone else. Also, the long term annual operating costs would actually be significantly lower as we could go from 4 matchday officials down to 1, and the training of the 1 official would be very small compared to what is needed (and arguably not occurring) today. There is an upfront cost in terms of cameras and computing resources, but I suggest that this is not prohibitive even for a smaller league such as the Scottish Premiership. Also the appeals process would be used very infrequently if the model is developed well and should be limited to cases where (1) the model itself expresses uncertainty about a particular incident -- this can actually be an included feature of the model -- or (2) the model itself clearly makes a mistake based on camera malfunction or an incident type it hasn't been modeled to classify, both of which should be rare. Lastly, one of the features of such models today is the recurring feedback loop where models continue to learn based on new training data, so any errors that are found can be recycled back into the training set.
This proposal is obviously not a short term solution based on the cost and the time to implement once approved (much less the lack of consensus around such a paradigm change in how we officiate matches!). But technology exists TODAY to implement such a solution, and I believe such a system would achieve all of the objectives that we currently request from our officials -- accuracy, impartiality, and consistency in the calls they make. Had this technology existed 200 years ago when the original laws of the game were written down, is there any doubt such a system would have been adopted?
For reference:
https://sportlogiq.com/en/
https://medium.com/tensorflow/predicting-balls-and-strikes-using-tensorflow-js-2acf1d7a447c
https://www.theringer.com/nba/2018/...ers-second-spectrum-courtvision-steve-ballmer
https://www.stats.com/football/
Most of those links above are focused on coaching or the fan experience, but the technology used is identical to what would be needed for a referee use-case.
(1) the performance / standard / level of officiating, and
(2) the compliance officer / appeal process.
I've already put forth my thoughts on #2 in a separate thread a couple weeks ago so I won't repeat myself here.
With regards to fixing the matchday performance of the referees and assistants, there have been a lot of suggestions, and most of them good ones. Foreign referees, making them full time, making them declare which team they support, providing better training, giving them access to VAR, among other ideas. I'm going to suggest one more idea that is a bit radical but one that I think would actually provide a better solution (as defined by fewer mistakes and much more consistency at the same time).
We live in an unprecedented era of technological progress. The use of computers in the last few decades has ushered in unparalleled productivity gains across a wide swath of the global economy, and recent advances in adjacent fields of AI, Machine Learning, and High Performance Computing (HPC) are being utilized in new and original ways to improve living standards and complete tasks with precision, speed, and accuracy not previously possible. These tools are only recently -- within the last 5 years -- being used in the field of athletics and analysis thereof, but the long term potential to replace match officials, or even to compliment them in the same way goal line technology has been utilized, is very real. We should embrace it and encourage the SFA to be on the leading edge of adoption even if such applications aren't yet perfect. Here's how it would work:
One of the biggest machine learning tasks today is image classification. In layman's terms, there's a computer model where the computer is fed a series of images that are tagged with classification data. For purposes of this discussion, all the possible classification categories are covered. The computer basically "learns" to associate certain aspects of the images to the classification categories. Then the model is fed additional images without tags and it is asked to classify the images based on what it has learned during the previous training step. This process has been used in applications such as self driving vehicles (recognizing real world objects on the road), medicine (classifying medical scans as cancerous/non-cancerous), converting scanned PDF document images into text, and many others. It has also been used to analyze sporting participants (based on their uniform number) and events (based on a sequence of images - aka a video feed). See the links below for examples. It should be noted that well constructed models are currently generating accuracy >99% when the scope of the problem is narrow, and errors in models have been reducing by factors approaching an order of magnitude each year! Even models that look at 'noisy' data and are not well defined (e.g. stock price prediction) are performing significantly better than human experts.
How this can be applied in soccer/football is to set up multiple high resolution cameras, ideally 4K or better, and feed them into a model in a synchronized way so the model can create a 3-D representation of the environment and detect the same object from multiple angles. From there we train it to classify certain events based on rules (offsides) or based on consensus judgement (violent conduct) of prior video input samples. From here, we create the output classifications based on the laws of the game. Train the model to identify: handball, ordinary foul, professional foul, violent conduct, offsides, ball out of play, ball across the goal line, etc. Basically anything that a referee would need to call today. So if the model sees a foul it could spit out and tell us that (1) a foul occurred, (2) who committed the foul against whom, (3) where on the field the foul occurred (as specified in x,y coordinates, (4) type of foul (tripping, elbow, shirt pulling, etc), and what the referee determination should be (free kick, penalty, yellow card, red card, warning, etc) based on consistent application of the decision in the training sample. Once the model can do all this, we tell it to alert an actual official on the field if certain parameters are met, similar to how goal line technology today sends an electronic signal to a watch the referee is wearing in some leagues. Note that not every detection of an event will warrant an on-filed call -- the model may detect a player in an offsides position based on x,y coordinates of all players and the ball, but if that player makes no movement toward the ball or is otherwise involved in the play, play should continue as it does today.
For purposes of model training, matches from other leagues are ideal because (1) there are many more examples across the globe for each possible classification category leading to better model accuracy, and (2) if SFA officials are the ones selected to judge the training data they won't even have the opportunity to exhibit bias for or against any Scottish club. Ideally the sample data would be split based on hard rule violations (offsides, ball out of play) that could be explicitly programmed vs judgement calls (fouls and appropriate penalty for fouls). Additionally, I would recommend that determination of fouls in the training set be exclusively outside the penalty box so that we don't train the model that certain fouls are acceptable inside the box but not outside. A foul is a foul anywhere on the field, and the consequence of the foul (free kick vs. penalty) is determined by x,y coordinates relative to the penalty box. The same would apply to red/yellow card to the offender.
The appeal of such a system is obvious (near 100% accuracy, eventually, and zero bias toward or against any team. Rangers would know Celtic isn't being favored, Celtic know that Rangers aren't being favored, and the rest of the league know that there isn't one set of rules for the Glasgow giants and another set of rules for everyone else. Also, the long term annual operating costs would actually be significantly lower as we could go from 4 matchday officials down to 1, and the training of the 1 official would be very small compared to what is needed (and arguably not occurring) today. There is an upfront cost in terms of cameras and computing resources, but I suggest that this is not prohibitive even for a smaller league such as the Scottish Premiership. Also the appeals process would be used very infrequently if the model is developed well and should be limited to cases where (1) the model itself expresses uncertainty about a particular incident -- this can actually be an included feature of the model -- or (2) the model itself clearly makes a mistake based on camera malfunction or an incident type it hasn't been modeled to classify, both of which should be rare. Lastly, one of the features of such models today is the recurring feedback loop where models continue to learn based on new training data, so any errors that are found can be recycled back into the training set.
This proposal is obviously not a short term solution based on the cost and the time to implement once approved (much less the lack of consensus around such a paradigm change in how we officiate matches!). But technology exists TODAY to implement such a solution, and I believe such a system would achieve all of the objectives that we currently request from our officials -- accuracy, impartiality, and consistency in the calls they make. Had this technology existed 200 years ago when the original laws of the game were written down, is there any doubt such a system would have been adopted?
For reference:
https://sportlogiq.com/en/
https://medium.com/tensorflow/predicting-balls-and-strikes-using-tensorflow-js-2acf1d7a447c
https://www.theringer.com/nba/2018/...ers-second-spectrum-courtvision-steve-ballmer
https://www.stats.com/football/
Most of those links above are focused on coaching or the fan experience, but the technology used is identical to what would be needed for a referee use-case.
