Комментарии:
Absolutely spot on. AI doesn’t magically appear without peogramming. When you look at who’s running the shows etc it doesn’t lead people to believe in transparency. Worry here being AI takes over from human judges and therefore is more controllable ?
ОтветитьHi Raf, first time I’ve commented since your hiatus over the summer. Though I agree that the programmer has a bias, I found the ai stats from jab-er in the first fight highly more insightful than most other reports. I don’t know if you saw it. It did tally the punches landed and thrown very well. In short, it did tell the story of the fight a lot more than the score cards etc. In some fights the bias will play a major part such as the Bivol Beterbiev fight I would guess, but in most situations I don’t think it will play a big part as the scoring should be based on who lands the more clean and effective punches.
ОтветитьI think AI can be a very useful analytical tool. But it doesn't solve the problem that people would like to be solved which is subjectivity since the AI judge is still just giving an opinion based upon data through visual observation at 2nd hand which is the same as a human judge. It may be the case that AI will be better at judging whether a blow landed or not, or was blocked but neither the human nor the AI can truly know the impact of effect of any punch. Sometimes a shot doesn't seem to have much on it but is incredibly damaging, at other times a shot looks and sounds horrific but has little effect on the person receiving it.
(That's just made me think of another point - Will the AI judge take in only visual data, will it have a microphone to pick up the sounds of a punches impact or the sound of the crowd? Both of those things can influence the opinion of the human judge for better or worse)
How will the AI be programmed to assess defensive success and how will that be weighted relative to offensive success. My instinct or suspicion would be that most humans are biased towards offence over defence and therefore whether intentionally or not the AI would be weighted the same way and have a bias towards rewarding aggression. This is obviously not just a problem of AI, since bias exists in humans which I guess is the reason you made this video. This is still one of the biggest issues with judging, ie What constitutes "effective aggression" and unfortunately that's the type fo philosophical question that may not have a definitive answer. I'm not sold on the idea of AI judges or AI as a whole. I think at this stage AI is a very good mimic but to call it intelligence I think is a stretch for the time being. Over time I think people will come to realise that these things are tools even if they are very impressive ones
People are missing the point. AI judges will be fair because their preferences and biases are fixed and known. Fighters now know exactly what’s being scored. No more imbecilic human subjectivity like muah points cuh he came forward that change from fight to fight
ОтветитьOther change I’d make. Max number of rounds are odd rather than even. Drastically lower the chances of a draw. Someone has to win
ОтветитьA further point, to play devils advocate or programmers advocate 😂Even if it were hypothetically possible for a programmer not to impart their own bias or preferences into the AI there are still major dangers of machine learning which are something like this. Imagine how different an AI judge you would get if you trained and AI on the fights of Floyd Mayweather vs. training an AI on the fights of Rocky Marciano. Those two AI's would end up with very different perceptions of what makes a good fighter or a winning fighter (I use those examples as neither lost so that removes one potential confound). One would be weighted towards defence and counterpunching to more decision wins (the later Floyd at least) the other would be weighted towards pressure fighting, taking a shot to give a bigger a shot and more knockout victories. So I would be interested how those various AI's would determine the winner of a highly technical fight like James Toney v Mike McCallum 1 where both showed incredible skill and subtlety as well as punch resistance. How would each AI interpret that? I have no idea
Another problem is not the programmers bias but picking up biasses from the pool of fights which are being analysed by the AI. Some number of those fights are going to include dodgy judging decisions which led to perceived "robberies" or results which the vast majority of people believe were unjust. The AI given a robbery in the learning phase will have to interpret the "winning" fighters actions in the fight (which most observers believed lost the fight) as superior boxing since a part of it's input will be that the fight was scored for them in reality. And so even with the best will in the world the bias of real life judges and therefore unfair results (robberies) will end up forming a part of the AI's perception so to speak. Now this is not necessarily an issue since humans are subject to the same biases but then it does undermine AI as a supposedly objective form of judging. Once again this suggests that the problem of bias in judging may not be solvable.
Your right it can’t measure how hurt someone is
ОтветитьI think the main advantage of an ai will be in consistency. This can lead to the boxing community developing clearer guidelines in terms what AI should prefer. This is a positive direction cuz its allows for actual progress while judging in boxing has been a mess for decades.
ОтветитьInteresting take. There are some huge advantages of using AI for this. While it’s biased, it’s still objective, and they should make it transparent how the AI judge scores and what they look for.
Some advantages AI could do
- the AI model won’t know who’s who, or their reputation, whereas a human judge does
- damage itself is very subjective, as one fighter may bruise more easily than others
- AI can measure other things, like the actual objective speed of a punch thrown, force it was thrown with, how much the opponents head snaps back, how much sweat goes flying off the boxers head etc
I think AI judging will be useful in the same way that compubox is useful in
I think there could be a place for it to stamp out egregiously wide judging like Canelo vs GGG 1, where it could be used to paint a clearer story of the fight than what is reflected in one of the cards (whatever intrinsic biases the AI has it ain't going to be Byrd bad). For closer fights though? Human judges need to normalise scoring 10-10 rounds.
ОтветитьDave allen clearly won judges ruled for Fisher, judging in boxing is rotten, not good omen for Usyk
Ответитьthat Dave Allen robbery made me turn it off cant be arsed with anymore boxing for awhile now
ОтветитьIf there was a board of control for all fight JUDGES where they must explain why they gave fighters rounds in very controversial fights and they can’t just take some cash and score it for a certain fighter, you wouldn’t need AI judging. But judges have no one to answer to in a sport ripe for corruption.
ОтветитьHey boxing fans! What are your boxing new years resolutions? Mine is to study and mimic Winky's high guard.
ОтветитьHi Raf, think you might be underestimating just what an “AI” can do. This term is used quite loosely these days. But, if they (jabbr) are using machine learning models (they 100% would be) they can get it to do just about any task a human can do, provided they have data to train a model from. In this case, there’s tonnes and tonnes of footage that they can use to train a model. However, it’s not as simple as just giving it the footage. They’d need to explicitly classify parts of the footage with certain things. Like, “a punch landed” and “the opponent is hurt”. They’d just go through the footage frame by frame, classify them, and train a model off the footage + classifications. This type of model is a classification model and uses a type of learning called supervised learning. In this case, us humans need to manually label data (supervise) so that a model can learn. The classification process is critical, and this is the step that may have bias. If you train a model off footage that hasn’t been correctly labeled, then the model would be biased. A bias may be someone’s definition of “hurt”.
I personally, am quite optimistic about this. Boxing is far too fast for human judges. A model (or in this case it would be many) can run in real time and be much more accurate than any human could. We just need to give it the right input data. I think it makes it potentially harder to rig fights as well as the breakdown of each round is soo detailed with something like Jabbr.
Something that would be cool, which I don’t think anyone has done, Is show a video from the Jabbr stats and break down where they agree/disagree with the stats. I’d be keen to see just how accurate it currently is. Maybe you could do this? Haha