 be a sec, we should be live now still don't see the live stream why? ok, now I have to see it ok, I wasn't meant to actually click on it to just fetch the link tell me a second, just in case someone from LinkedIn wants to join in this sound is not great as well do I have a semi-professional microphone without it? the GPU is in front of my microphone so probably all the noise is getting into it I'm doing a couple of things try to stick to just one what? I see comments nope, nothing ok, just might get started maybe they will appear later yeah, YouTube changed something again ok, we'll get to fuzzy logic in a sec that's what we were doing yesterday or the day before it doesn't matter actually, at the end of the long GPT-4 session we asked to write a relevant punchline it actually did had this second sentence that I thought was not really necessary which GPT agreed, of course it did so it rewrote the punchline into saying explaining modern technology to my grandparents it's like trying to install Wi-Fi on a typewriter a lot of effort but I'm still not getting any signal ok, maybe that's not very funny I was trying to improve it I originally didn't use Wi-Fi I used the firmware update on a typewriter but yeah, so it could potentially improve the punchline but then it became worse seems like the single shot prompt worked a bit better so I might try it again sometime later so let's get back to it we have this web application it's a flask app this is what it looks like at the moment it's generating these different waveforms you can add noise on top of it and there's a fuzzy logic detector it's essentially a classifier algorithm that should, here in the output wave, detect the original input measure detection error and present unknown the moment is not really doing it so we will be trying to improve it and the problem is that the output, the classification is a print we're currently printing this output the classification of a membership function returning very small numbers so I have to fix that and we mentioned it the square classification logic output I think I changed some of those parameters but they will go over them in a bit so looking at the symmetry of the signal we're essentially looking at two things symmetry and something else, sharpness symmetry and sharpness and then for each square triangle, for each rule we should have those defined and they don't have to be high and low they can be medium, high, low change that just try test the shit don't see it on my phone for some reason suggesting me other videos so the main problem is once this output is way too low it's always below 1.3, A1 over 3, A1 fed so it's always returning square this modification you can actually do anything except adjusting that from low from high to low the symmetry of a triangle should be high no, it should be low the symmetry should be low I'm not sure exactly so symmetry is the essentially comparing the first half of the signal to the second one the triangle should be low those membership functions okay, right through the episode it's returning the noise amplitude to begin with is too high we can change that noise the max should be 1 and 2 so we can actually see the signal okay, we get 0.1 right triangle left triangle these numbers are super low okay, what is the print again? print output the output numbers are extremely low so they always be low 1.3, GPT-4 is thinking forever let's try Github Copilot as well to select the whole text it seems that this should be the fuzzy logic function where the absolve is that it could be due to the way the fuzzy membership function is ruled the find specification is down one possible solution is to adjust the fuzzy membership function and rules, for example you might want to adjust the ranges of the membership function or the conditions to actually change symmetry high, high, low 4 square, the symmetry should be high now actually it doesn't make sense we have to double check change the code in X play this quickly it's like a fuzzy library fuzzy algorithms to generate triangle membership why is that? because it's a triangle known membership function logic is curved it defines how which point we don't want a triangle or a problem on them a different shape but during if we can this might be complicated but can we plot the membership function I'm wondering if there is a way to display the fuzzy logic parameters because that should be like an explainable AI X, AI the other question is can GPT just design the model for us currently it's not working at all doing anything the website is super slow something is up back to the question that everyone answers that yes it says that one possible solution is adjusting the fuzzy membership functions and rules then it's adjusting some of the parameters but it's not actually saying why we decided to use these specific parameters so we need to be able to select those that we wanted to also before I never used the triangle a membership function so now I decided to use that to actually do it based on yeah that's what they currently defined as you need to understand the data you're working on here of course I do and what you want your fuzzy logic system to achieve for example if you find that your system is so sensitive you might want to adjust the symmetry low high function to be less deep okay and 5.25 suggesting the same numbers a membership function symmetry low in this triangle so it used to be minus 1, minus 1, 0 sharpness and that how to use this instead using some of how much as the rules I am not sure but these rules the second line that doesn't make sense we need the rules for the other say for minus 1 symmetry low sharpness low it actually makes sense understanding that and what you want to fuzzy logic system to achieve then experiment with different settings until you find them okay we want should kind of make sense that sense that line takes the rules square triangle trapezoid it's symmetry universe again yeah control f5 square square yeah this number is the output there of the fuzzy systems super low say symmetry that's the difference between a peak sharpness universe this is wrong isn't it isn't this man to be using peak sharpness a control in space and sharpness check this logic okay they're just normalizing it is it just normalizing okay but the point is these numbers the output they're still super low for output try the same normed so the output of the fuzzy logic system is always being less than one third not only less than one third extremely low numbers and yes the number will return always return a square output so the detection won't work properly the fuzzy D-fuzz function called as D-fuzzified result of an aggregate fuzzy membership function the issue might be the way the rules are defined symmetry and peak features extracted from noisy signal yeah we have to go back to how the features are extracted from the signal symmetry yeah let's print those numbers for troubleshooting default is square okay so peak sharpness kind of makes sense symmetry is way too low peak sharpness so for left triangle the peak sharpness should be higher yeah it's higher but not by much then symmetry symmetry is even lower sharpness is lower yeah those numbers are not changing much look at those functions extracting symmetry so we're dividing the signal by two first half second half reverse the second half and the symmetry measures is the inverse of the sum of the absolute differences okay that makes sense the bad feature because so if the square function was actually different should have had the minus one then going one so it's essentially if maybe if we shift and left trap so we have to change those functions let's just look at different features unlike the triangular try yeah this try math triangle membership function what are all the other options for the types of membership functions triangle membership yeah it's normally I was using traptazoid Gaussian sigmoid weekly rules can we use is we have minus one minus five zero zero can we also add three medium delete that quickly add the medium one and the traptazoid all come on what happened to github two so they overlap today it's actually you expect them to overlap right now those mean ranges are not actually being used let's comment so we have symmetry law rules now you just copy it you can just update should include symmetry and sharpness mid um yeah include symmetry mid sharpness mid yes I did what way did this circle come from does make any sense we really need gpt4 to look at this and can we start the new shit the website is down or something probably the CEO is Gonski again kind of to be expected okay the aggregated numbers also don't make sense all the same why aggregated numbers um let's see fix the whole code yeah obviously ideally would have had the visualization of the fuzzy logic rules membership functions hey we don't have it just have to deal with it well we're using a language model anyways huh okay so the problem is that these two are not being used a see the mean and sharp mean I mean are not used select the whole code because for github component is important gpt just refuses refuses to work the back online I don't know a system have detected I don't know is anyone else getting those errors okay wipe this circle I don't have a circle for current most used fuzzy logic rules include them you can create new rules the new rule circle just associated with meet symmetry and meet sharpness with circle and no I don't want that that aggregation the the classification steps I don't have a circle shape waveform and I don't need it no I do want to use the sharpness the symmetry medium so the idea is that some of those waveforms so looking at the symmetry and sharpness of the signal and now we try to add the three ranges for the two parameters so you can have low, medium and high symmetry and same low, medium and high sharpness you want the symmetry medium rules you can modify the circles for example okay so let's see if it makes any sense so for square the symmetry is high sharpness is low yeah it would have been correct if the square actually had this extra bit so if it was a symmetrical we can actually do it can adjust the square the triangle, symmetry, medium sharpness high okay maybe trapezoid symmetry high yeah so for example yeah trapezoid definitely the symmetry is definitely high sharpness okay high if you say so yes we can check triangle versus trapezoid in terms of symmetry okay let's comment that out for a sec and use this to test we expect let's see and it's control f5 this so say if we're comparing let's clear this comparing left triangle symmetry 0.03 and trapezoid symmetry 0.18 so yes it's significantly higher which is what we expect peak sharpness 0.6 for trapezoid and 0.6 for triangle okay that doesn't make sense yeah now the problem is still that this output numbers are extremely low yeah rule square writes a bit output when we had to update we don't have the circle we don't want the circle oops just happened so now we have three levels for each feature low, medium, high I'm assuming this is correct we need any more rules and no we need to fix aggregated we might we need to update this aggregated we're removing these zeros update your code include mid-symmetry and mid-sharpness we need to explain that that line and this one remain the same problem is that aggregated values are all the same could be a problem I don't know why it's loading forever as well loading forever yeah I know why it's loading forever because it's still loading that JavaScript being in here as well why? that's because I do need those zeros I don't know why it just doesn't work without them plotly mean.js is loading forever I don't know if it's because hopefully it's only because I'm live streaming at the moment okay, clear that for a sec just say we have left triangle symmetry 0.03 peak sharpness 0.52 and we have a trapezoid symmetry 0.06 so it's like double but it's still a very low number point of and sharpness increased as well which actually does not make sense is aggregated ways meant to have the same numbers so aggregated let's forget control shift C instead of regular control C which stops the application from running aggregated values always the same is GPT for back online or is it still going to be giving errors suggesting debugging I'm already doing it no, it's not working what the website is it it's still super slow normally GPT-3 is much faster it's only blocking me or someone else has this issue as well so returning sharpness low open in here make some starting some heavy troubleshooting peak sharpness symmetry then we have the rules all rules and aggregated and output save, reload doesn't work cause we need to actually run it so this is what we get for square okay it seems like those numbers are saturated so we need to add square and input signal and the amount of noise make noise zero let's make noise zero noise amplitude let's print input wave and noise amplitude that's right clear output is 2.6 the output is even smaller in the error when the noise is too low this trapezoid noise amplitude ideal if that fix those numbers they're not right actually fixing our model turning this very close to zero which is not expected this might be due to the fact that the aggregated membership function is constant across the entire universe that's a good sentence constant across the entire universe of discourse whatever that means the defacification process calculates the weight at the average of the universe being the membership grades if the membership grades are the same for all values in the universe the defacified value will be the centroid in this case the universe to fix this you need to ensure the membership functions are defined in such a way that aggregated membership function varies across the universe here's an example of how I might update your membership functions and rules symmetry okay but this will essentially just make all the symmetry medium and it's back to triangle why is it back to triangle don't like triangles as a membership function use riveting to that line problem in it instead it doesn't work getting errors for it in such a way that they overlap the fuzzy rules are defined in such a way that the aggregated membership function varies across the universe discourse it should present defacified values not close to zero contrast github co-pilot sorry microsoft people maybe to give it a go essentially doesn't make much sense yeah nothing works for all 3.5 github co-pilot doesn't make much sense don't know why it keeps riveting to triangle membership functions I took an instant how this chatbot works in terms of if I respond does it respond to the last looking at the last response or the current code or everything else those called trapezoidal trapezoidal membership function better this case trapezoidal triangle Gaussian depends on specific application of course it is if you are using trapezoidal membership functions for both low and high in a triangle for the mid-category this common approach actually this makes a lot of sense for a change okay can you actually do that can you update even on a mid one could be trapezoidal as well there's no problem okay it didn't actually adjust the code let's go with what it suggesting over here symmetry low and common that out sharpness rule where mid-med high-high low-high aggregated again just go with it for a sec and output is the same is 53 3 it's aggregated correct adding the zeros so we know this works let's actually go with that suggestion taking the original rule square it's actually printed again this after concatenation those test test this again yeah the plotly JavaScript is loading forever hopefully it's only because I'm live-streaming at the moment getting at this length and was working GPT back online and check as GPT is down or something turning it into NumPy array I wasn't working before so that wasn't an array symmetry low why are they becoming scale-less yeah you have copilot going places to a google from January yesterday partial outage so for now today service degradation in gpd it's the kind of pst time 11pm yes this was from an hour an hour and a half ago yeah the site is just hopeless there's no questions in the chat or anything might finish it up have this uh tool it is gesture gesture groove it's currently working as the JavaScript only and with some music generated with gestures might break my stream yeah that was taking a lot of bandwidth that face and hand detection wonder this one does it in JavaScript wonder if it's actually faster to allow camera 5 10 eras and that's not working why is it because of it does not work close this quickly let's just try it one last time to restart that camera or something yeah we have to try it again offline see if it makes any difference we'll see it later no music today try it again tomorrow go try bonycals.com let me know what you think in the chat see you later bye