Forgive the rudimentary nature of my question.
Is it possible to build a AI startup using Perceptilabs?
Other than domain experts from a particular field, do we need to hire developers? How do we monetize the efforts?
I have a group of medical doctors and researchers as my friends who are doing research in the area of sickle cell anemia.
How can we move forward in this direction?
We have absolutely no idea how to move forward in this direction.
Any guidance in this direction will us a lot.
In fact Perceptilabs got all of us excited when we discovered it via TwoMinutePapers. Suddenly it gave us a hope that we can do this.
Forgive the rudimentary nature of my question.
Welcome to the forum!
PerceptiLabs will help you build models in primarily the computer vision domain. It can help you quickly test out some models on data you have and start getting a POC running that you can verify, and then you can take that model into production when you are ready for it.
If you are building a startup you probably are going to want to have some sort of application (I’m guessing) outside of just the model. Getting that application up, connecting it to the model and building the relevant pipelines will likely require some software engineering, but you can get surprisingly far on just a POC step and secure enough funding to take in external developers that can help you with those steps.
So to summarize, without knowing exactly how your idea looks like, I would give it a shot and start with just having a model that you can feed some data into. We are soon looking to integrate with Gradio as well which would give you a decent POC you can show people and move forward on.
(Also, just to make sure I give you proper startup advice outside the technical part, make sure you have a market by iterating as much as possible on ideas and early POCs as you can. The better you can find that people need what you are building and the better you can make sure that there are a lot of those people, the easier your life is going to be down the road.)
Hope that helps!
Thank you Robert for taking the time to respond.
Reporting a small development.
We have been able to gather a small team to build the said pipelines.
Can we build an application like this?
Is there any tutorials (video or otherwise) or resources to achieve the above?
Can you keep training the models to make it better and improve its accuracy?
Can we export the model outside of Perceptilabs environment and keep training it?
We are currently in discussion with few domain experts from the cancer research field.
Hi again @devhash,
Congratulations on the development, very cool!
There seems to be a whole stack of features on that website, but here are some things you can do:
- Detect Pulmonary Embolism
- Alert on the detection (requires an inference script)
- Segment where the pulmonary embolism can be found in the image (I didn’t see them doing this on the website, but it’s something you can do if you want to)
This seems to be the core AI based pieces, while the rest is software around it.
For resources, here are some similar cases where we classify things in images: https://docs.perceptilabs.com/perceptilabs/use-cases/healthcare-and-medical
Here they built a full pipeline using the PL model: https://blog.perceptilabs.com/case-study-genetic-screening-usings-cnns/
We also have a good amount of videos on our YouTube channel that goes through how models are built: https://www.youtube.com/c/PerceptiLabs
And we have a blog soon coming out on a segmentation model which detects bacteria.
You can keep training the models to make them better and improve their accuracy, but be careful so you don’t overfit. Overfitting means that you are learning the training data so well that you no longer can generalize well to the real world data. You can see this happening if the validation or test accuracy start dropping (or the validation loss keeps increasing).
You can export a trained model outside PL as a standard TensorFlow model. This at the very least allows you to use it for inference (run the model on real world data), but I believe you are also able to load the TensorFlow model in a script outside PL to train it outside of the tool.
It will essentially act as any other TensorFlow model at this point and it’s up to you what you do with it.
Hope that helps and let me know if you have any other questions!
@robertl Thank Robert for the guidance.
I had a query regarding putting together an MVP using Bubble.io
We are a little short on talent resources as majority of our team members are occupied with some pressing deliverables.
So would it be possible to put together a decent MVP using a platform like Bubble and Perceptilabs?
Thanks in advance.
Things have advanced since the last post
I can’t answer about bubble, but in PerceptiLabs V0.13 a number of ways of serving a trained model were introduced - including gradio which serves into a browser tab.
You can find more info here Exporting and Deploying Models with PerceptiLabs
I have seen gradio demo’d and used it myself; I don’t know your use-case but it could be worth checking out (and don’t forget the other choices too!).
@devhash awesome that things are moving forward for you!
I’ve never used bubble.io, but as long as you can add some small script where you load and use a model then you should get something up and running.
Then when you wanna get fancier you can also look at placing the model in the cloud (using FastAPI for example as a server) and just get inferences from there.
@JulianSMoore had some great suggestions as well, starting with using the model in Gradio could be a good first step.