The thing about shmart IoT shtuff is: If the company that made my cheap ass "smart" electrical socket plug thingemagings goes out of business, I essentially have a very overengineered in-line power switch sitting there, pinging a site that may be bought by any random troll for cheap for all of of eternity. Once I got a job I'll invest in some Ubiquity plugs or even find the time to learn how to solder my own. With a VPN tunnel into my home network it'll be exactly as useful as before, but without the "sending random data to china" part.
You're too kind. I'm still a new channel, and I think RUclips's great and mighty Algorithm is still figuring out who to show my videos. I built it, so hopefully they will come!
Until your bachelors you will merely have covered the basics for all of this. They are importan and all, but its even more important to educate yourself with awesome content like this to see, as you already said, whats possible with stuff that you will think is useless (linear algebra and so on)
sry, just reread my own comment and almost suffered a freaking stroke. all i mean is that, you will hear yourself say "what do i need this for?". See all that crap you will cover are the fundamental basics upon which alot of more adanved stuff is built upon. I suffered the price for not paying attention in the beginning, i hope you dont get discouraged as quickly as me.
5:32 As someone working in a research lab, I felt personally attacked. In my defense, I was on a lunch break when I watched the video. Keep up the good work!
As a musician I have often dreamed of a MIDI controller where I could sing a bass line then percussion then whatever into a loop and have the software pick an instrument to play based on the sound im making and parse simple verbal commands like pause and play. Some time ago I thought AI would help me realize my dream but as I looked into it the process, huge computer / cloud requirements shot the whole idea down. Thank you for making me realize it's now not only possible but portable! Great channel. All the best.
I have nearly same lab like you. 90% of the tools is identical. I use a prusa printer too :) And i working on an AI controlled mini robot. So now i'm happy, because i found a same crazy guy like me. :)
When youtube suggestions actually works well for you! I've binged 4 of your videos so far!😂😁you got good humor and projects brother! I appreciate you! #goals
Here is a cool ideea: this glove can be used to communicate better with deaf people. You can make it to display words based on what you are doing with your hand in sign language. I think this would be pretty cool. I came from instagram by the way, keep up the good work!
Thanks a lot! That’s totally possible, especially with how flexible the machine-learning architecture is. If I trained it with words instead of letters, it might just be doable!
@@ZackFreedman Some university kids did this recently. Well I don't know if it uses ML, but they're gloves that translate sign language. It was cool, but I need to actually read the article sometime.
Magical translating ASL gloves are a perennial idea. Unfortunately, it is generally a perennial idea of hearing people, and not always well-researched with linguists and native signers... Anyone wishing to pursue this should start here: www.theatlantic.com/technology/archive/2017/11/why-sign-language-gloves-dont-help-deaf-people/545441/ grieve-smith.com/blog/2016/04/ten-reasons-why-sign-to-speech-is-not-going-to-be-practical-any-time-soon/
Dude, pleasantly surprised by the quality of your content! I came to the video expecting some guy rambling on about general ideas on the subject, and was faced with a full all-out guide on how to embed AI. Hats off!!
That project was very cool! And the TF Lite for micros looks very promising. So glad that found your channel! I'm enjoying every one of your videos! From myself I'll add that if even after all optimizations you still cannot fit your model you can try to "distile" your network. It's a technique when you train a smaller network using the big network as an input. Sometimes it works and decreases performance by tiny bit but also deceases the model size significantly.
While i wont be tackling orignial projects on a scale that you do , your videos of your projects and a few other youtubers have got me into playign with electronics again , thanks :)
When choosing a development board, also take in to consideration the hardware accelerated AI boards. - Jetson nano - Rasbarry pi wit intel movidius stick - Miaxduino - Googles coral etc
honestly if you could turn this device into a mouse pointer/replacement this could be a game changer. I currently struggle with hand pain and basically shouldn't use any kind of mice, not even ergonomical ones and thought about alternative input methods for quite a while to help with the recovery. Might need to look into building one of these gloves to try it myself :D
New sub here... Just finished a 4 episode mini-binge and I am impressed with how you are able to present technical subjects with a humorous flare, so much so that watching episodes where I already know about the subject matter, such as circuit boards, was still bery entertaining. Here's to 100k subs! Almost there....Amazing, but well deserved, growth rate.
1:03 - "...as they hammer away on their 60% ortholinear keyboards..." Never have I been called out so directly in a RUclips video. You sir have earned this like and subscribe.
I have finally found a likeminded person. Thank you. I hope to follow your footsteps with the knowledge of data science machine learning and to combine it with my mechanical engineering / manufacturing experience
In my career as an assembly language programmer I use to love taking sample code from a software engineer that was megabytes of Matlab generated algorithms and reducing it down to only a few lines of assembly instructions.
Fun fact: I didn't intend the Juicero to be an AvE reference. I was just researching examples of ridiculous IOT devices, and every list had it. It's just such a ridiculous device.
@@ZackFreedman Truly, the Juicero as a product was such a bizarre concept. Really, I guess the device itself wasn't too strange, but their obscene super elitist consumerist business model certainly was ... special.
@@ZackFreedman I heard about some project using IR cameras, retroreflective tape and blu-ray lasers to zap mosquitos in midair. High-tech equivalent of mosquito netting that gets rid of those blood-suckers permanently.
Right now I am working on a project that is very similar--I am aiming to connect a touchpad to a microcontroller which will translate the touchpad inputs into letters, and print them on a computer like a keyboard. This video is perfect! I have to pretrain a network to classify characters and then cram it onto a microcontroller (my choice was a Teensy 4.1, what a coincidence). Your video is exactly what I needed. So glad I found someone else on the internet who is doing something similar to what I'm doing AND made a great video on it.
It's totally doable. You also don't need to get hung up on configuring the touchpad for absolute input, since the network doesn't care. It should actually be much easier now than when I made this video, since TFLite has developed significantly.
I’m a computer engineering major and I don’t learn how to do any of this in school, if you see this please tell me how you started and what you learn to figure out stuff like this
I had to do it in 2019 for a last semester project that I insisted to solve by embedding a ML network into an Arduino Uno. After a lot of research the only way I managed to do it was to make the matrices by hand and take the trained Tensorflow model weights and paste them into matrices. Very nice video!
I remember when Amazon came out with a paper describing a way to make big neural networks small with similar performance by quantizing down the values. Neat to see it applied! :)
This video! Amazing. Years ago I was trying to get NNs working on rPi, and it was nearly impossible as training took forever. This sidesteps it all and user the big beasts where required, and then you put it into a tiny uproc to really knock it out the park!
Love the video! Tried to do this 5 years ago for a project which indeed couldn't rely on a stable internet connection. Amazing to see how far we've come! At the time I had to use a prototype sipeed risc-v microcontroller and 3 different chinese compilers to get the job done 😅.
I had a project with edge machine learning this autumn and this video would have saved me so much time understanding it all. Was refreshing to watch this after tho, awesome!
This is way beyond my skill level but I love it. Keep up the funny vids, and don't get lured to the silly youtubers side, but I know you're too clever for that.
Man the youtube algorithm realy loves you, you are gaining like 1k-4k subscribers per day with a channel sub 100k subscribers. Im not complaining love your content :)
Man, this could be the future for deaf people to comunicate with not deaf person throught a speaker, even this can be craked down viceversa to show in a video screen( a smartphone) a video of two hands doing the sign lenguage!!! This is awesomee and very useful, thanks for putting all this work for us to be entertained.
This was a great video. I am an older computer guy that has a very small programming background (but growing) and I even followed this fairly well. Hell it was so entertaining that I will watch it a couple more times. Thanks man!!!
Oh boy. The amount of effort you put into your Projects AND videos is breathtaking! Great Work! I wish I knew enough about all of this to replicate your project but I am only a hobbyist and studies cell biology sooo...
Cool video. Some advices for NN architecture: 1. Use 1d cnn instead of dense or lstm. It will give you better performance/accuracy. If conv1d is not supported, you can try to implement it with conv2d. it shouldn't have a big impact on performance if done correctly. 2. Train nn with lower prison for less drop in accuracy after model quantization. 3. Use higher batch size (and why 1st layer have count of neurons = batch size??) and try some regularization and data augmentation. 4. Try to use prelu with shared axes 1,2 instead of relu. Attention mechanism can also give better results with the same performance. 4.1 Check if there are "dead relus" in nn and remove the corresponding neurons. 5. If "big" ARM chip is your choice, DONT use tflite, try xnnpack. Your nn will run faster and device will be colder
First time viewer and I know this video is 4mos old. Can't wait to watch more innovations! One thing I'd add if you haven't already is to add a segment of you specifically using the build as intended. THANKS FOR THE AWESOME CONTENT!
Extremely cool project. I might give something like this a shot as well to start getting more familiar with ML. Especially now that I know it's possible to pull things like that off on microcontrollers. As an embedded engineer I'm usually kinda scared to veneer in higher level abstractions where I don't know exactly what my hardware and code is doing but it feels like that's where the future lies so this "locally done ML edge computing" on microcontrollers might be a good place to start. I also liked your video about PCB evolution throughout history. I expect your channel to explode one day.
Subbed and liked, great video! After watching 40+ videos on Tensorflow(light) the last 12 months, the RUclips-AI didn't have to crunch too many numbers in order to decide that I'm probably in the target audience haha :)
No dislikes. Of course, well deserved! Awesome stuff as always. Glad you are growing so quickly and I wish you the best of luck in this youtube adventure!
The sneaky reference to the 1000$ bag squeezing cashgrabbing "robot" as an analogy of the overkills crippling nowdays technology... You made my day sir
Juicero hits different when you build products. From my end, the high-pressure juice press mechanism is incredible, and the Wi-Fi fruity DRM is the dumbest part
First of all, Zach your channel is A+ godlike froth. Second of all, check out the oscilloscope video from Smarter Every Day about drawing with sound waves if you’re not already familiar. Thank you for sharing and I look forward to all of your content yet to come.
From someone who works mainly with firmware and embedded stuff, I always just kindof wrote off machine learning as outside the realm of possibility right now. I didn't even know this was an option, which is a great additional tool in the toolbox.
Just couple days ago I've watched "Fugou Keiji: Balance:Unlimited", really got inspired by the AI they show there, so i was about to look for ways to create one or copy(not a programmer at all, but i can work with instructions), it would definitely need an input method that isn't a keyboard, and here youTube algorithm recommends just the right thing! Thank you for making this great video, i will be looking into your idea(if you don't mind, that is).
You are awesome, Zack! I mostly only understand 10% of what you are saying but the humor, you skill and the content is awesome. Keep pumping out the projects! Best luck to both of you (:
This approach would be very useful if you want to deploy lots of sensors somewhere remote and have them detect fault-related patterns without just constantly sending all the data back to a server. For example, vibrations in turbine blades, movements in railway switches, noise/vibration in underground pipelines or pipes in a huge refinery.
My dude. Very cool video. Neural networks have a high overhead to implement on an embedded system even when you quantize them and do all the fancy tensorflow magic. Looking at the 2 seconds of your training session it looks like maybe a support vector machine or a tree classifier would work. Both are pretty efficient and can be implemented in a embedded systems without major problems.
As a web developer I 100% agree with his description of web developers and assessment of our response to not using the internet to solve all your problems haha.
Please make videos regularly and don't give up, you will grow bigger someday by making this kinda quality content Pardon for my bad england, i know more than one language
You can't use any kind of microcontroller in space, most arduinos use a quartz crystal oscillator to generate the cpu clock, however that crystal only works consistently on Earth, if you get that oscillator outside of the Earth it's properties change and the frequency changes too.
I'm guessing the precompiled vs non-compiled just messed with you so hard until you figure that out. Every time I failed working with Arduino has been the damn pre-compiled file 😭. You're right on the edge buddy of figuring out how to make something really feel like they're messing with AI on half a gigabyte of RAM. Love your channel buddy
I totally agree with the sentiment that "smart things" don't need to be "internet of things".
The thing about shmart IoT shtuff is: If the company that made my cheap ass "smart" electrical socket plug thingemagings goes out of business, I essentially have a very overengineered in-line power switch sitting there, pinging a site that may be bought by any random troll for cheap for all of of eternity. Once I got a job I'll invest in some Ubiquity plugs or even find the time to learn how to solder my own. With a VPN tunnel into my home network it'll be exactly as useful as before, but without the "sending random data to china" part.
@@fredwupkensoppel8949 lol
@@fredwupkensoppel8949 that was very specific
@@fredwupkensoppel8949 that's sounds more like an anecdote
Tasmota is great 😉
Build a shoulder mounted turret that reads targeting info from the glove so when you flip someone off it shoots them with 3D printed nerf cannons
That would RULE
@@ZackFreedman I know the perfect partner for this project: Michael Reeves
The US Military has entered the chat.
DARPA has entered the chat.
@ShroudedPanda yeah...this glove could have an extra capability...electrifies people just with thinking 😬
As someone who works in full stack web development, your description of us is at least 80% accurate
Dude, this is amazing quality content. I don't know how this video only has a few hundred views.
You're too kind. I'm still a new channel, and I think RUclips's great and mighty Algorithm is still figuring out who to show my videos. I built it, so hopefully they will come!
Well it’s up to 13k views now
Agreed. Excellent content for the inquisitive hardware hacker. Subscribed.
@@Jianju69 I don't think he's hacking stuffs...
@@queenoffaygo it's up to 29k views. almost tripled in 4 days
As a brand new engineering student, your channel makes me excited to see what is possible with a little understanding. Keep the good content coming!
Agreed, we stand on the shoulder of giants with a LOT of understanding.
@@favesongslist omg why is everyone here thinking the exact same thing I'm thinking
Until your bachelors you will merely have covered the basics for all of this. They are importan and all, but its even more important to educate yourself with awesome content like this to see, as you already said, whats possible with stuff that you will think is useless (linear algebra and so on)
sry, just reread my own comment and almost suffered a freaking stroke. all i mean is that, you will hear yourself say "what do i need this for?". See all that crap you will cover are the fundamental basics upon which alot of more adanved stuff is built upon. I suffered the price for not paying attention in the beginning, i hope you dont get discouraged as quickly as me.
5:32 As someone working in a research lab, I felt personally attacked. In my defense, I was on a lunch break when I watched the video.
Keep up the good work!
You're the type of person I want my friends to be.
As a musician I have often dreamed of a MIDI controller where I could sing a bass line then percussion then whatever into a loop and have the software pick an instrument to play based on the sound im making and parse simple verbal commands like pause and play.
Some time ago I thought AI would help me realize my dream but as I looked into it the process, huge computer / cloud requirements shot the whole idea down.
Thank you for making me realize it's now not only possible but portable!
Great channel. All the best.
I have nearly same lab like you. 90% of the tools is identical. I use a prusa printer too :) And i working on an AI controlled mini robot. So now i'm happy, because i found a same crazy guy like me. :)
I see you have good taste!
Should add content
@@dull218 I'm work on it.
@@gyorsomogyi Its been a year, get on it! I demand more mad inventor content.
I understood the word "glove" and that's about as far as I've gotten.
Nice I thought I am the only one by looking at the comment section. Looks very interesting though :D
Same
Same
When youtube suggestions actually works well for you! I've binged 4 of your videos so far!😂😁you got good humor and projects brother! I appreciate you! #goals
Thanks a lot! I appreciate that. Enjoy the show!
I just binged the whole channel instead of sleeping
@@h3Xh3Xh3X it's the obvious thing to do.
LOL'ed at, "Get back to work! You've a virus to sequence!". Loving these videos.
Bro we looked identical
You both look like the ends of q tips
@@gravestoner2488 hahahaha 🤣🤣🤣
This guy is such a joyful combo of careless carefree humor and actual hardcore smarts.
lol! I am a "researcher working on sequencing a virus" watching this video for it's quality content!
Here is a cool ideea: this glove can be used to communicate better with deaf people. You can make it to display words based on what you are doing with your hand in sign language. I think this would be pretty cool. I came from instagram by the way, keep up the good work!
Thanks a lot! That’s totally possible, especially with how flexible the machine-learning architecture is. If I trained it with words instead of letters, it might just be doable!
@@ZackFreedman Some university kids did this recently. Well I don't know if it uses ML, but they're gloves that translate sign language. It was cool, but I need to actually read the article sometime.
@@eatcomics That sounds rad. If you can find it, I'd love to read it.
Magical translating ASL gloves are a perennial idea. Unfortunately, it is generally a perennial idea of hearing people, and not always well-researched with linguists and native signers...
Anyone wishing to pursue this should start here: www.theatlantic.com/technology/archive/2017/11/why-sign-language-gloves-dont-help-deaf-people/545441/ grieve-smith.com/blog/2016/04/ten-reasons-why-sign-to-speech-is-not-going-to-be-practical-any-time-soon/
Dude, pleasantly surprised by the quality of your content! I came to the video expecting some guy rambling on about general ideas on the subject, and was faced with a full all-out guide on how to embed AI. Hats off!!
The explanation, the tech, the guy from Scanners getting mind blown? I'm an instant fan. Subscription confirmed.
Humour is absolutely brilliant. More videos on edge devices please - thanks for the cocktail recipe :)
That project was very cool! And the TF Lite for micros looks very promising. So glad that found your channel! I'm enjoying every one of your videos!
From myself I'll add that if even after all optimizations you still cannot fit your model you can try to "distile" your network. It's a technique when you train a smaller network using the big network as an input. Sometimes it works and decreases performance by tiny bit but also deceases the model size significantly.
I have no idea what you're saying, but you're sense of humor is worth the brain scratching! this channel smacks of early "This old Tony". I love it!
Californian's may not know this but.... lol. that right there got you a new subscriber.
We have surprisingly bad internet access in the heart of Silicon Valley. Comcast owns everything, so you have to sell a kidney for 100 Mbps.
@@EricLaspe here in Germany, i am happy, to have WiFi, since i would need to go up the next hill, just to send a SMS
@@EricLaspe Facts
_cries in Indian_
@@javidfarhan1675 cries in African
While i wont be tackling orignial projects on a scale that you do , your videos of your projects and a few other youtubers have got me into playign with electronics again , thanks :)
Jesus Christ just subscribed to the first video that I viewed. A few years from now on and this channel is going to blow haha. Good luck!
When choosing a development board, also take in to consideration the hardware accelerated AI boards.
- Jetson nano
- Rasbarry pi wit intel movidius stick
- Miaxduino
- Googles coral
etc
They are not MCUs
honestly if you could turn this device into a mouse pointer/replacement this could be a game changer. I currently struggle with hand pain and basically shouldn't use any kind of mice, not even ergonomical ones and thought about alternative input methods for quite a while to help with the recovery.
Might need to look into building one of these gloves to try it myself :D
New sub here... Just finished a 4 episode mini-binge and I am impressed with how you are able to present technical subjects with a humorous flare, so much so that watching episodes where I already know about the subject matter, such as circuit boards, was still bery entertaining. Here's to 100k subs! Almost there....Amazing, but well deserved, growth rate.
Man I just found some high quality content here.
1:03 - "...as they hammer away on their 60% ortholinear keyboards..." Never have I been called out so directly in a RUclips video. You sir have earned this like and subscribe.
10:37 I see what you did there.
I used to be completly clueless to this world and now little by little this guy has taught me so much
3:52 I never expected two italian liquors in a machine learning video...
Why the devil NOT?
I have finally found a likeminded person. Thank you. I hope to follow your footsteps with the knowledge of data science machine learning and to combine it with my mechanical engineering / manufacturing experience
"stop watching youtube, ya slacker"
aight, damn!
In my career as an assembly language programmer I use to love taking sample code from a software engineer that was megabytes of Matlab generated algorithms and reducing it down to only a few lines of assembly instructions.
I saw the juicero in either this video or the one before it, and heard you mention chooching.
I see you. Also, subbed several videos ago.
Fun fact: I didn't intend the Juicero to be an AvE reference. I was just researching examples of ridiculous IOT devices, and every list had it. It's just such a ridiculous device.
@@ZackFreedman Truly, the Juicero as a product was such a bizarre concept. Really, I guess the device itself wasn't too strange, but their obscene super elitist consumerist business model certainly was ... special.
Man, YT recommendation algorithm nailed it today. Fantastic work and love the production quality of the vid.
Please, please, please build a mosquito carpet bomber (with ai)! LMAO
Ooh, the wife is a mosquito magnet. She'd love that.
@@ZackFreedman I heard about some project using IR cameras, retroreflective tape and blu-ray lasers to zap mosquitos in midair. High-tech equivalent of mosquito netting that gets rid of those blood-suckers permanently.
@@ZackFreedman make artificial chameleon tongue mosquito killer
Now I trust the RUclips recommendation algorithm,
This is the best Chanel to start the year with! Great contents!
See, the real problem is you can't build a subscription model on embedded AI.
Right now I am working on a project that is very similar--I am aiming to connect a touchpad to a microcontroller which will translate the touchpad inputs into letters, and print them on a computer like a keyboard. This video is perfect! I have to pretrain a network to classify characters and then cram it onto a microcontroller (my choice was a Teensy 4.1, what a coincidence). Your video is exactly what I needed. So glad I found someone else on the internet who is doing something similar to what I'm doing AND made a great video on it.
It's totally doable. You also don't need to get hung up on configuring the touchpad for absolute input, since the network doesn't care. It should actually be much easier now than when I made this video, since TFLite has developed significantly.
I’m a computer engineering major and I don’t learn how to do any of this in school, if you see this please tell me how you started and what you learn to figure out stuff like this
You should have access to a machine learning class. I took a ML class during grad school. If you don’t, MIT has some open courseware stuff for ML.
This really is one of the best programming/build videos I've ever seen, will definitely be watching the rest, amazing content
Midway through I felt that there was something weird in his face and after a while I realized it was that he didn't have his screen classes on 😂😂😂
I've probably never been so glad for clicking yt recommendation as I am now. Amazing video, awesome projekt.
You are awsome! Can youtry to teach sign language to the glove and then make an integrated speaker voice them?
I had to do it in 2019 for a last semester project that I insisted to solve by embedding a ML network into an Arduino Uno.
After a lot of research the only way I managed to do it was to make the matrices by hand and take the trained Tensorflow model weights and paste them into matrices.
Very nice video!
Fun video, love the sass
Thanks my grid-shaped breakfast pastry brother
I remember when Amazon came out with a paper describing a way to make big neural networks small with similar performance by quantizing down the values. Neat to see it applied! :)
Dude, you remind me of a friendlier and funnier version of the Russian villain in Iron Man 2
This video! Amazing. Years ago I was trying to get NNs working on rPi, and it was nearly impossible as training took forever. This sidesteps it all and user the big beasts where required, and then you put it into a tiny uproc to really knock it out the park!
I'd love to see this as a device for translating ASL.
Omg yes! ASL live translation gloves
Love the video! Tried to do this 5 years ago for a project which indeed couldn't rely on a stable internet connection. Amazing to see how far we've come! At the time I had to use a prototype sipeed risc-v microcontroller and 3 different chinese compilers to get the job done 😅.
I wish I could understand whats going on. Great video, I guess?
I had a project with edge machine learning this autumn and this video would have saved me so much time understanding it all.
Was refreshing to watch this after tho, awesome!
This is way beyond my skill level but I love it. Keep up the funny vids, and don't get lured to the silly youtubers side, but I know you're too clever for that.
Silly? Me? Never. This is a serious channel.
Man the youtube algorithm realy loves you, you are gaining like 1k-4k subscribers per day with a channel sub 100k subscribers. Im not complaining love your content :)
Pretty sure that they could cram GPT into this in a few years lol!
Isn't the whole point of GPT-2 and GPT-3 to have so many parameters that somehow it works?
Good stuff man. as a layman, i feel like I'm learning a lot watching your videos. my favorites are the start to finish projects.
why am i watching this at 2am? i cant even code
yet...
Man, this could be the future for deaf people to comunicate with not deaf person throught a speaker, even this can be craked down viceversa to show in a video screen( a smartphone) a video of two hands doing the sign lenguage!!! This is awesomee and very useful, thanks for putting all this work for us to be entertained.
Madlad did put the recipe in the description. I had all the ingredients and did it, it's good btw.
I have almost zero experience with coding but I found that I could almost follow all the way along. You explain things very well! Thank you!
This was a great video. I am an older computer guy that has a very small programming background (but growing) and I even followed this fairly well. Hell it was so entertaining that I will watch it a couple more times. Thanks man!!!
I don't know anything about programing yet found this so interesting and entertaining, this is Quality content bro!
Oh boy. The amount of effort you put into your Projects AND videos is breathtaking! Great Work!
I wish I knew enough about all of this to replicate your project but I am only a hobbyist and studies cell biology sooo...
Thank you for teaching me how to make a cocktail 🙂
A high quality super duper amazingly one of the best content on RUclips ❤️😎
Cool video. Some advices for NN architecture:
1. Use 1d cnn instead of dense or lstm. It will give you better performance/accuracy. If conv1d is not supported, you can try to implement it with conv2d. it shouldn't have a big impact on performance if done correctly.
2. Train nn with lower prison for less drop in accuracy after model quantization.
3. Use higher batch size (and why 1st layer have count of neurons = batch size??) and try some regularization and data augmentation.
4. Try to use prelu with shared axes 1,2 instead of relu. Attention mechanism can also give better results with the same performance.
4.1 Check if there are "dead relus" in nn and remove the corresponding neurons.
5. If "big" ARM chip is your choice, DONT use tflite, try xnnpack. Your nn will run faster and device will be colder
First time viewer and I know this video is 4mos old. Can't wait to watch more innovations! One thing I'd add if you haven't already is to add a segment of you specifically using the build as intended. THANKS FOR THE AWESOME CONTENT!
Extremely cool project. I might give something like this a shot as well to start getting more familiar with ML. Especially now that I know it's possible to pull things like that off on microcontrollers.
As an embedded engineer I'm usually kinda scared to veneer in higher level abstractions where I don't know exactly what my hardware and code is doing but it feels like that's where the future lies so this "locally done ML edge computing" on microcontrollers might be a good place to start.
I also liked your video about PCB evolution throughout history. I expect your channel to explode one day.
Subbed and liked, great video! After watching 40+ videos on Tensorflow(light) the last 12 months, the RUclips-AI didn't have to crunch too many numbers in order to decide that I'm probably in the target audience haha :)
No dislikes. Of course, well deserved! Awesome stuff as always. Glad you are growing so quickly and I wish you the best of luck in this youtube adventure!
Absolutely delighted I found your channel! Maybe the algorithm is finally giving you the recognition you deserve. Really quality content. 🤘
This is pure golden knowledge. Thanks for the video, mate, this can level up anyone's microcontroller game.
The sneaky reference to the 1000$ bag squeezing cashgrabbing "robot" as an analogy of the overkills crippling nowdays technology... You made my day sir
Juicero hits different when you build products. From my end, the high-pressure juice press mechanism is incredible, and the Wi-Fi fruity DRM is the dumbest part
First of all, Zach your channel is A+ godlike froth. Second of all, check out the oscilloscope video from Smarter Every Day about drawing with sound waves if you’re not already familiar. Thank you for sharing and I look forward to all of your content yet to come.
Thank god, youtube algorithm sending me this marvel. Great job mate, really love the idea and the no-fuss/friendly approach/
From someone who works mainly with firmware and embedded stuff, I always just kindof wrote off machine learning as outside the realm of possibility right now. I didn't even know this was an option, which is a great additional tool in the toolbox.
this is so awesome!!! i hope to do a project like this one day.
Man, i love the content and also the presentation of it.
But I'm soo glad that this exists
Just couple days ago I've watched "Fugou Keiji: Balance:Unlimited", really got inspired by the AI they show there, so i was about to look for ways to create one or copy(not a programmer at all, but i can work with instructions), it would definitely need an input method that isn't a keyboard, and here youTube algorithm recommends just the right thing! Thank you for making this great video, i will be looking into your idea(if you don't mind, that is).
You are awesome, Zack! I mostly only understand 10% of what you are saying but the humor, you skill and the content is awesome. Keep pumping out the projects! Best luck to both of you (:
This approach would be very useful if you want to deploy lots of sensors somewhere remote and have them detect fault-related patterns without just constantly sending all the data back to a server. For example, vibrations in turbine blades, movements in railway switches, noise/vibration in underground pipelines or pipes in a huge refinery.
Jeez. That was crazy. But incredible that it can be put on such a small device, looks like I need to watch the glove vid now!
It's rad. And this video is just amateur beginner stuff, too.
Night sight turned on in the middle of the video and the programmers tan turned into jaundice.
Keep up the good work!
My dude. Very cool video. Neural networks have a high overhead to implement on an embedded system even when you quantize them and do all the fancy tensorflow magic. Looking at the 2 seconds of your training session it looks like maybe a support vector machine or a tree classifier would work. Both are pretty efficient and can be implemented in a embedded systems without major problems.
Your Thomas the train edits inspire me. So good
You are a hero bro. Andrew Ng (the hero of deep learning) will be proud of you
I'm glad I got this on my recommend. Leaving a comment for the algorithm and also for the insane content. You're an inspiration !
Loving the film on the oscilloscope
You are incredibly talented
This is an awesome project! Whenever I finish a Jupyter Notebook I can explain every line, 2 days later - hieroglyphs.
This video should’ve more views and like. Thank you for sharing this.
As a web developer I 100% agree with his description of web developers and assessment of our response to not using the internet to solve all your problems haha.
Understood about .05% of what was said but loved watching every minute of it! Keep it up brohemoth.
Please make videos regularly and don't give up, you will grow bigger someday by making this kinda quality content
Pardon for my bad england, i know more than one language
I have a mild knowledge (I'm not sure if you'd call it understanding) of what you did in this video, but the result is awesome. Great work!
I am so happy this video showed up in my feed
You can't use any kind of microcontroller in space, most arduinos use a quartz crystal oscillator to generate the cpu clock, however that crystal only works consistently on Earth, if you get that oscillator outside of the Earth it's properties change and the frequency changes too.
Huh, that's pretty cool!
I'm guessing the precompiled vs non-compiled just messed with you so hard until you figure that out. Every time I failed working with Arduino has been the damn pre-compiled file 😭. You're right on the edge buddy of figuring out how to make something really feel like they're messing with AI on half a gigabyte of RAM. Love your channel buddy
Loved it man❤️ I hope you get millions of views, cause you deserve it.💯
One of the best vids I’ve seen on the Ytube