So, don't get me wrong. This is a fun project and a neat little device. I don't mean to take away from it in any way.
However it is really important to consider why baby monitors are so primitive: because the cost of a false negative is huge. I didn't see any mention of this in the author's experiments (only a '>98% accuracy' note). So let's talk about this a little bit: is "accuracy" what we want? Probably not---I don't care if I get accidentally notified, but I care very much if I don't get notified when the baby is crying. So you want to weight your classifier's predictions heavily against false negatives (at the price of false positives). It would be good to make an ROC curve to characterize this behavior. More importantly though, any predictive model assumes a stationary distribution; i.e., training conditions accurately reflect test conditions. But will they in real life? What about when your neighbor's house is under construction? Can interference from chainsaws cause the model to fail to detect the baby crying? What about the dude down the street with his super loud motorcycle? What happens then? I bet the training set doesn't have situations like this.
I really, really don't want to come off like a wet blanket here. But I feel obligated to, because this is a model that directly impacts the welfare of a human, and so we should at least talk about or discuss potential drawbacks. (Again, cool weekend project, just, we need to be clear about the implications of outsourcing the decision of whether the baby is crying to a black-box model where we can't interpret what it's doing.)
> (only a '>98% accuracy' note). So let's talk about this a little bit: is "accuracy" what we want? Probably not---I don't care if I get accidentally notified, but I care very much if I don't get notified when the baby is crying
Presumably this is 98% accuracy per sample or something, not per session of crying. I wouldn't want to leave my kid completely unanswered 1/50th of the time if he actually needed me, but I think with the model as trained, regular crying would eventually get noticed and that's fine with me.
Somewhere else on the thread though somebody else called out the real downside of this model: it overfits on regular crying. Kids make all kinds of noises. Mine made a completely different, eldritch noise one time when he noticed a spider, for instance.
> the real downside of this model: it overfits on regular crying
I'm with you there, that's a big downside too, but that's not the "real" downside---there are like seven different downsides present with the data science going on here and it's hard for me to say which is the biggest issue, because they're all issues. Data science is not trivial!
This reminds me of the home-built garage door opener that popped up here a few months ago. When it comes to safety devices, there are literally man-decades of expertise that has gone into these devices. It is never a good idea to rely on something you whip up yourself over the weekend or convince anyone else that they could.
> In 2016, General Motors took the lead among automotive manufacturers by introducing the Rear Seat Reminder, a technology designed to nudge drivers to check their back seats as they exit their vehicles. It uses an audible alert and a front panel message to tell drivers to check the rear of their vehicle for occupants.
> Rear Seat Reminder technology became standard on all new Chevrolet, Buick, GMC, and Cadillac four-door sedans, SUVs, and crossovers starting with the 2019 model year, and also will be standard on all 2020 model year GM pickup trucks, said GM spokesperson Phil Lienert.
> Kia, Nissan, and Subaru offer rear-seat alert systems in many of their models, according to Car and Driver, and Hyundai announced on July 31 – National Heatstroke Day – that it planned to incorporate the technology across all models by 2022. One of Hyundai’s newest innovations is the Ultrasonic Rear Occupant Alert, in which a sensor can detect the presence of a child (or pet) and activates a loud horn if the driver leaves with the child inside.
I'm not sure where you live but in the United States, in some states, children are required by law to ride in the back seat of an automobile until they are, I believe, 8 years of age. I believe that various state and federal agency recommendations are for children to ride in the back seat of an automobile until they are 13 years of age.
Presumably, then, parents are not putting their young children who cannot operate a car door into the front seat of a car and leaving them.
But the tech exists. It's not like we're using tin cans and string and trying to replace that with a gizmo; we have the gizmo, and now we're trying to make an open-source version of it. While the goal is admirable there's no real benefit from using a less-tested Raspberry Pi project. There are less risky ways to learn the same lessons - a video walky-talky maybe.
As a father, I generally agree with what you both are saying regarding being conservative with tech choices on delicate areas such as babies, but I‘d also add that 99% of situations of my baby making random noises (that also trigger a baby monitor) are not something that are going yo have an impact on the baby long term. Most times my baby has lost the pacifier in the dark, and finds it before I reach upstairs (or annoyingly, between me waking up, and me reaching my bedroom door).
Also there’s the scary converse: some important things do not make a noise, such as a baby suffocating in her sleep.
I have to agree - I like tinkering and hacking as much as the next dev, and I don't like to be negative, but... what is the point of AI/ML here, where a simple sounds trigger works just fine? What problem is it actually solving - I only see this approach causing problems.
Cars are every bit as dangerous as guns. But no one here seems to be in such a panic with Tesla or self-driving cars.
I feel there is a bit of undercurrent of Luddite sentiment going around here. Lots of people mentioning how "primitive" a baby monitor is or should be. But is it? It's an LCD video screen powered by modern advancements in battery tech connected wirelessly to a CCD video camera, with sensing technology to detect when a baby is crying and even provide other information such as room temp. None of this existed 20 years ago. Not to mention that between now and back then, baby monitors went through a long phase of being hot garbage. Many still are garbage.
> I feel there is a bit of undercurrent of Luddite sentiment going around here.
Oh, I'm no Luddite, don't get me wrong---I'm a machine learning researcher. I have no problem with data science. The difference between all of the complex technology you pointed out inside the baby monitor and what we're talking about here is that all of that complex technology is robust and, for the most part, well-designed! The data science work here has tons of issues, and I think the vast majority of reactions are reacting to that: machine learning and data science really can be useful... but not if you apply it really badly. Someone else commented elsewhere in the thread that a simple thresholding algorithm would be just as effective---and not suffer from the myriad potential problems present with blindly applying TensorFlow because it's cool.
I wonder if there are huge advances to be made in suppressor technology? Other than the noise, guns are basically perfect in their intended function. I guess you could try to further reduce recoil.
There have been some pretty fascinating experimental weapons over the the years. For example, the H&K G11 [0] used caseless ammunition. The potential advantages of that are no need to eject a spent casing and a soldier can carry more of it since it weighs less.
Oh, there's a lot more stuff that can be done to guns:
1) improve aim: embed wind and angle sensors, possibly even battlefield intel (position, weather, land layout) to account for any drift that might impact the bullet. Also, auto-fire if a designated target is in the crosshair (train the AI on human faces, combine with the previous sensoring). I would not be surprised if this technology will be developed rather sooner than later with armed robots, and then made smaller until it fits in a gun or at least a rifle.
2) improve/rethink propulsion. Right now almost all guns operate by some form of bullet in a casing with explosive propellant (excluding the rare caseless guns and co2/pressured gas sports guns). Railguns are already a thing at "ship scale", it will be only a matter of time until it gets scaled down to hand-held guns.
3) improve projectiles. Right now bullets are dumb pieces of metal. Why not have active bullets (e.g. subminiature rockets) or bullets laced with poisons so that even a scrape kills in the end?
4) improve... guns themselves, as a concept - think laser guns a la Star Trek, highly focused microwave, sound or other energy.
In the end humanity will always improve ways to kill each other, and all the concepts are already there in sci-fi (and in the case of poison bullets, the Russians made it a reality with the Markov murder).
I think those are all really great ideas for new inventions, some of which might eventually replace firearms. But I still think firearms are basically topped out. Anything else done to them complicates them more than it improves them.
However it is really important to consider why baby monitors are so primitive: because the cost of a false negative is huge. I didn't see any mention of this in the author's experiments (only a '>98% accuracy' note). So let's talk about this a little bit: is "accuracy" what we want? Probably not---I don't care if I get accidentally notified, but I care very much if I don't get notified when the baby is crying. So you want to weight your classifier's predictions heavily against false negatives (at the price of false positives). It would be good to make an ROC curve to characterize this behavior. More importantly though, any predictive model assumes a stationary distribution; i.e., training conditions accurately reflect test conditions. But will they in real life? What about when your neighbor's house is under construction? Can interference from chainsaws cause the model to fail to detect the baby crying? What about the dude down the street with his super loud motorcycle? What happens then? I bet the training set doesn't have situations like this.
I really, really don't want to come off like a wet blanket here. But I feel obligated to, because this is a model that directly impacts the welfare of a human, and so we should at least talk about or discuss potential drawbacks. (Again, cool weekend project, just, we need to be clear about the implications of outsourcing the decision of whether the baby is crying to a black-box model where we can't interpret what it's doing.)