Communicating with commercial R&D management about entering into AI

Deep Learning
#1

Hi everyone,

I figured I’d get your opinion on this as I figured I’d be likely to find people who have been in this situation and perhaps recently. Last year I was looking at the amazing achievements by AI and was frustrated that I wasn’t a part of it. So I pitched changing the direction of my PhD 2/3 the way through to work on automation and AI. Last winter I watched the fast.ai course videos and jumped into it. Few months later I had proof of concepted segmentation for an important application in the company and we are almost done drafting the scientific article! So first of all, amazing work by you fast.ai for allowing people like me to take the leap! And you too community by helping.

Fast forward some months, I have made a new method in the lab which WOULD incorporate imaging and deep-learning based segmentation. It’s a huge success but I have yet to show that the segmentation will work and they are of course interested in the development time. As they are likely to hire me to finalize development of this, I would like your opinion on how to handle this situation.

While I am confident that it will work technologically, as always, the robustness and generalizability of the solution will depend on data volume and quality. I don’t really have the experience to estimate the needed data. So when they ask me how long it will take to develop the AI, how do I approach it?

The complexity of the images and objects is relatively low. Circular dark objects. They can vary in size, shape and texture somewhat. They can also meld together, which is why I want to use deep-learning and not e.g. thresholding, as the method needs to be quite accurate I separating them. I can probably create and label 400 images a day, each containing 20-200 objects.

Either way, I am wondering how seasoned developers and companies approach the unknown in development time. Can the time be estimated well by experienced developers? Are more seasoned companies just more open to take the chance?

Any input or thoughts you have are appreciated.

1 Like
#2

Hi sebbecht Hope all is well!

One way would be to build the smallest and simplest possible prototype.
With a small number of categories and the simplest possible interface.

See the following links Share you work here - highlights, and Share your work here ✅ there are many examples where people have created small applications and hosted them online.

Having worked in R&D and Cloud providing for a number of years I have seen as many successful projects as unsuccessful projects.

I would also suggest you enlist some programmers from your university IT dept. if you need to.

If you already have a working notebook or code for your segmentation model, then creating a simple app similar to the ones in the links above shouldn’t take more than a week or two for a good programmer.

Creating the app will:

  • highlight many of the possible issues your final app may have
  • will help you convince R&D management
  • help you predict more accurately the time the complete project/app will take
  • is one the best way I have found of convincing anyone of the benefits of a particular application/project.

This approach of building something is also recommended by @jeremy.

A quote I like is:

Show them before you tell them.

As it’s much easier, when you have something tangible with your communication to persuade.
Hope this helps.

cheers mrfabulous1 :grinning: :grinning:

#3

Great point @mrfabulous1! I of course have in mind to show them the previous segmentation application and telling them that most of the code is plug-and-play. But maybe a small proof of concept would be a strong message! will definitely consider. thanks.

1 Like