How a Machine Learned to Spot Depression | Hacker News Reader

How a Machine Learned to Spot Depression

npr.org

81 pointstokenadult19 comments

Whenever I see these kinds of video processing on real time, I just want to drop everything and work on that. It seems so cool! It's just like movies that we all saw while growing. Now the world is completely full of smartphones with cameras that could make great things! But somehow they don't reach us.

Might it be that sensors and processing power are still not quite there? I've just seen a Project Tango video (https://news.ycombinator.com/item?id=9634317) and it seems to suggest that.

BUT there are so much people focusing on these things that on one hand I feel that great things will come very soon (Oculus, etc.), but on the other hand, it's just a very risky move. The market is not yet created, everything is just a demo.

The other day here in HN someone commented something similar here: https://news.ycombinator.com/item?id=9627754

There is a huge risk in creating a cohesive value proposition for a big mass of people. Be reckless and you could flop like Google Glass. But once someone does that successfully, a huge market for entertainment will have just open. So my idea is to just toy with these things and gain knowledge and proficiency, but don't jump into it full time just yet. Do you think this is an intelligent (non-)move? Or just fear of the unknown? The thing is, I would also not want to move just for Fear Of Missing Out!

Why not go work for a funded startup in the area? Working full time on this problem will advance your skill like nothing else ever could, you will have a tiny chance at winning the lottery, you will meet like-minded people, and the only downside is slightly reduced pay.

> Whenever I see these kinds of video processing on real time, I just want to drop everything and work on that.

Personally, I find AI too much of a "fuzzy black box" thing to be really interested in it. I have to agree that the applications are often truly marvelous. But the fact that, for example, training neural nets is often more of an art than a science is holding me back to step into the field.

I like to work on a problem, think it through, and be 99% sure that it will work before I have even written one line of code. With AI, it is more like "let's try this and see if it works".

Personally, I find AI too much of a "fuzzy black box" thing to be really interested in it.

That's because artificial intelligence is subject to eternally moving goalposts. Any technology which we are fully able to understand ceases to be intelligent; it is relegated to the status of mere "algorithm".

We understand machine learning itself, it's the problem space that needs to be understood. That's why it is such a black box - you never know the complexity of the model you need to build before you tackle it.

Yes, the truly interesting models are hard to analyze but, to paraphrase, if intelligence was simple enough that we could understand it, we would be so simple we couldn't.

I wrote a UX tracking tool for tablets[0]. It ended up not going to market simply because you're right: processing power is not cheap enough. I started with on-board processing, but that led to substandard result. We then moved on to cloud based processing, but having cloudbased processing would increase costs a lot and we couldn't keep the pricepoint.

Lastly, such things are difficult to market and we weren't ready financially to take on such risks.

[0] - http://blog.chewxy.com/2014/04/08/eyetracking-jetpack-joyrid...

That looks cool. To get that face mesh to fit to the video in realtime like that, what kind of processing power do you need? Is this work an application of that or did they do this because now algorithms/hardware is up to the task?

Funny; we've been working on the same thing at iCouch for a while now: real time predictive analytics during a video therapy session. Even funnier is that the investor community has nearly zero interest in what we're doing. I bet if we built a real-time, gluten free, organic food ordering service, we could raise $500k in about ten minutes. Our problem is determining market size; what's the market size for technology hat can predict severe mental illness situations potentially weeks before they happen? What's the market size for tech that can prevent suicides? Since nothing else like that exists, there's no way to predict how much that would be worth in terms of an investment deck. It's frustrating. The revenue generating side of the business (our SAAS for therapists,) is doing wel enough, but certainly not well enough to be able to subsidize moon shot technology that would actually disrupt mental health diagnosis and treatment. But alas, good to see NPR is giving the issue some coverage.

It is insanely frustrating when fundraising worthiness is so tied to how good you can make 10 keynote slides. Changing the world doesn't always fit into bullet points.

Our problem is determining market size; what's the market size for technology hat can predict severe mental illness situations potentially weeks before they happen? What's the market size for tech that can prevent suicides? Since nothing else like that exists, there's no way to predict how much that would be worth in terms of an investment deck.

Look up the number of PTSD-qualified therapists in the USA (or just therapists if "PTSD-qualified" is not a meaningful distinction). Survey a random sample of them and ask if they might use technology to help diagnosis and to prevent suicide. Multiply the percentage who say yes by the total number of therapists. That's your slide deck market size.

As someone who also works on this, and thinks about startup possibilities, I understand what you're going through. Sometimes I wonder if it's possible to sustain a business like this at all or if you just have to work toward an insurance company buyout.

Maybe we should all start a therapy group!

Reminds me of:

https://simpsonswiki.com/wiki/Virtual_doctor

I'm always reminded of: https://en.wikipedia.org/wiki/ELIZA whenever anything like this comes up

The same sentence can have complete opposite meaning if it was said in an angry or calm voice. Sentiment is such an important part of human communication, that these systems will have to be incorporated in to speech recognition if they are going to be meaningful. 😉

Your comment seem phrased to imply the current Ellie system (featured in the article) doesn't read sentiment (when you write "will have to be" about incorporating sentiment- that usually reads that an existing system like Ellie doesn't have that"). From the article: "When I answer Ellie's questions, she listens. But she doesn't process the words I'm saying. She analyzes my tone. A camera tracks every detail of my facial expressions."

The comment said systems like Ellie will have to be incorporated into speech recognition, to better process the context of the words being recognised.

Christ almighty, these god damned unicode smiley faces are the bane of my existence.

very cool. Not exactly related but I have been thinking if it would be possible to teach deep neural networks lie detection from video/audio feed, picking up on cues that we would normally ignore (facial microexpressions etc)

That's a pertinent question in relation to the article submitted here. The main answer to your question is that there don't seem to be any invariant behaviors, or even reliable behaviors, associated with lying.[1] Some people are "naturally" capable of lying shamelessly, and many other people can be trained to lie while beating any lie detection system based on physical signs.

The best methods for detecting lying when (for example) conducting a police investigation or engaging in foreign intelligence is to analyze the liar's statements for inconsistencies. Statement analysis that prompts new questions in a structured interview is the best way to detect deception, because it's hard to make up a consistent set of story details in advance when making up a story.

[1] http://www.slate.com/articles/technology/future_tense/2013/0...

http://www.people.com/people/archive/article/0,,20079233,00....

http://www.amazon.com/A-Tremor-In-The-Blood/dp/0306457822

[2] http://leb.fbi.gov/2011/june/evaluating-truthfulness-and-det...

https://books.google.com/books?hl=en&lr=&id=5GjTLtibmXYC&oi=...

Reply on news.ycombinator.com