The "UX design" terminology confused me. It seems that the author is advocating more for tools and methods that let us visualize & inspect how a given neural network comes to be based on its training data, and why it gives a particular output for a particular input.
This ties back into the "neural networks are black boxes" argument that's been floating around for a while, no?
Is there any concrete research in this field? The initiatives the author points to seem to be mainly about highlighting the problem, not addressing it.
That being said, while I agree with it in principle, in practice what would compel companies to make their neural network based systems understandable and analyzable to users? There are plenty of non deep learning based systems in my life that I already don't have access to the internals of; for instance, it'd really be great if I could actually see the set of formulas that determine my credit score, for one example amongst many.
I think a simpler start might be in order. Most developers, let alone users, don't know when, why and how, to use ML. I created (with partners) the "smart reply" designs that are visualizations of ML results that are used by Google in Allo and Hangouts and were used it various other earlier Gproperties. The goal was just to allow people to understand how ML uses probabilities to predict possible futures rather than one off single guesses. You can see and play a bit with how probability trees create outputs and play around with them. They were also designed to be implicitly participatory and multiplayer, rather than just static outputs you could either use or not use. The hope was that more sophisticated users could get a mental model around this and maybe become inspired.
I think several other "lessons" like this will be in order to better get people playing with ML driven interfaces before the average developer really starts seeing enough value to use tool sets meant to visualize the NNs.
Hi all! I'm actually the author of the article- if you have any comments feel free to leave them.
I specifically chose to address this to ux designers bx ml will be the future of product design and what I had noticed, from working at IBM Watson, was a lack of understanding across design and engineering teams how to "explain" or show what was happening in products using ux. This is actually the start of a column for fast co focusing specifically on design for AI and ml.
So how a product is processing information and serving up results has to be articulated to users- maybe that articulation is a visualization or maybe it's something else? It depends upon the product- but this was a call for thinking about how, if your product uses ml, how do you translate that for users in the design of the product? It's advocating that for ML we move away from minimal design.
Reddit did a great job when they created a new ranking algorithm. They marketed as "best"[1] and they have an informational post for nontechnical people but in the post they linked to the technical explanation[2].
These deep learning neural networks imitate the human brain and are essentially black boxes. If the trainer doesn't have a complete understanding of what the network is doing how would a UX designer? This makes no sense.
i agree and disagree with you- a trainer may not understand how or what the connections a neural network will make but any product goes through EXTENSIVE QA before being launched. thus, a trainer will see all of the different outcomes any nn will make inside of a product, they see the gamut of differences. a user may not see that.
additionally, i don't think you give UX designers enough credit. i worked alongside engineers, and met frequently with trainers to talk about what kind of results our prototypes were generating. it was up to me to help articulate what the results meant and how those connections were made.
for example- how do you make a long tail result looking different than a short tail one without making it look like a google search? how do you articulate those differences to users so it just doesn't look like a shitty searching algorithm? that is a design problem- articulating those differences visually- not an engineering problem.
I don't think the fundamental premise of this is correct although I understand the reason why it's being proposed.
Machine learning will mostly remove the need for UX expertise as it will allow us to worry less about manually interacting with machines as a way to get them to do things. Instead ML will do all the in between interactions behind the scenes and present us with results.
If you think about Machine Learning as a feedback loop where only the input and result become relevant then you are back to something more akin info graphics which used to be called information design.
A much bigger problem to solve UX wise is to think about how machine learning will design things and how humans will deal with the results.
In the shown example the result created by ML is many times better than a human could have done but for it to work it also requires the ability to take this design and implement it into real life as the result is extremely complex.
A lot of things to explore in this space but I think we do ourselves a disservice to speak about is as UX.
As long as there are users, it will be necessary (or desirable, at the very very least) to design experiences for them. The assumption you make is that "talking to a humanlike intelligence is so natural as to be trivial." But salespeople, for instance, still need to be trained in how best to talk to their customers, what to say, what not to say, etc..
To the extent that salespeople are relevant and using ML nothing UX wise has changed. Same rules still apply.
My assumption is that many of the benefits of ML can't be used by humans but have to be used by other machines to create a final "useful to human" output.
The biggest challenge I think for designers in ML is, we are very often one step behind engineers who build ML algorithm. Without a good understanding of the capacities and limitations of machine learning, it’s hard for designers to envision opportunities and be involved in the early strategic conversation
15 comments
[ 3.3 ms ] story [ 46.1 ms ] threadThis ties back into the "neural networks are black boxes" argument that's been floating around for a while, no?
Is there any concrete research in this field? The initiatives the author points to seem to be mainly about highlighting the problem, not addressing it.
That being said, while I agree with it in principle, in practice what would compel companies to make their neural network based systems understandable and analyzable to users? There are plenty of non deep learning based systems in my life that I already don't have access to the internals of; for instance, it'd really be great if I could actually see the set of formulas that determine my credit score, for one example amongst many.
https://www.oreilly.com/ideas/ideas-on-interpreting-machine-...
I think several other "lessons" like this will be in order to better get people playing with ML driven interfaces before the average developer really starts seeing enough value to use tool sets meant to visualize the NNs.
I specifically chose to address this to ux designers bx ml will be the future of product design and what I had noticed, from working at IBM Watson, was a lack of understanding across design and engineering teams how to "explain" or show what was happening in products using ux. This is actually the start of a column for fast co focusing specifically on design for AI and ml.
So how a product is processing information and serving up results has to be articulated to users- maybe that articulation is a visualization or maybe it's something else? It depends upon the product- but this was a call for thinking about how, if your product uses ml, how do you translate that for users in the design of the product? It's advocating that for ML we move away from minimal design.
[1]https://redditblog.com/2009/10/15/reddits-new-comment-sortin... [2] http://www.evanmiller.org/how-not-to-sort-by-average-rating....
additionally, i don't think you give UX designers enough credit. i worked alongside engineers, and met frequently with trainers to talk about what kind of results our prototypes were generating. it was up to me to help articulate what the results meant and how those connections were made.
for example- how do you make a long tail result looking different than a short tail one without making it look like a google search? how do you articulate those differences to users so it just doesn't look like a shitty searching algorithm? that is a design problem- articulating those differences visually- not an engineering problem.
this is what the article is advocating for.
Machine learning will mostly remove the need for UX expertise as it will allow us to worry less about manually interacting with machines as a way to get them to do things. Instead ML will do all the in between interactions behind the scenes and present us with results.
If you think about Machine Learning as a feedback loop where only the input and result become relevant then you are back to something more akin info graphics which used to be called information design.
A much bigger problem to solve UX wise is to think about how machine learning will design things and how humans will deal with the results.
https://twitter.com/Hello_World/status/861735184990961664
In the shown example the result created by ML is many times better than a human could have done but for it to work it also requires the ability to take this design and implement it into real life as the result is extremely complex.
A lot of things to explore in this space but I think we do ourselves a disservice to speak about is as UX.
My assumption is that many of the benefits of ML can't be used by humans but have to be used by other machines to create a final "useful to human" output.