Yes. Hyper-parameters control the architecture (how many layers? Which activation? Regularization? and so on) and details about the training, parameters are the knobs within the architecture that are optimized during training. Hyper-parameters are not optimized during training of one instance of an architecture.
4 Likes
Usually with jupyter running on localhost:8888
- STANDARD: http://localhost:8888/tree
- LAB: http://localhost:8888/lab
Can’t we always model a “recommended action” as a prediction of what a human specialist would do in the same situation? The difference between actions and predictions wasn’t clear for me.
1 Like
Jason Brownlee gives a good differentiation here https://machinelearningmastery.com/difference-between-a-parameter-and-a-hyperparameter/
2 Likes
An interesting resource about bias and in general “ethical AI”:
(thanks to the FastAI/Insight Data meetup 
)
1 Like
That is the case if the models feeds its predictions to a human,… which is not always the case. Lots of companies are trying to use AI to reduce their numbers of employees…
1 Like
Are there good data labelling or annotation services? @jeremy mentioned medical imaging data - Do you know of anyone offering good labelling for such datasets, like radiology ones?
There are quite a few. Appen and scale.ai for example.
2 Likes
Algorithmic systems and human decision makers are used differently in practice, so you can’t treat them as plug-and-play interchangeable. I wrote about some key differences in this post:
What HBR Gets Wrong About Algorithms and Bias
and cover those ideas in my PyBay talk on bias:
7 Likes
There are plenty of options there. Labelbox, Amazon SageMaker, v7labs.com (focused on medical imaging, but totally usable on other things) are some I know about. Most of these services have free tier quotas.
THere are also open source tools that you can install locally for “personal” labeling tasks.
2 Likes
Can you comment more on the open source tools available for labelling/annotation?
If you want to get FDA approval, you need medical professional to annotate data. That’s why there are so few labelled datasets available.
2 Likes
The tools tend to be use case specific. For text/NLP take a look at prodigy. https://prodi.gy/
2 Likes
The theory is in all the previous iteration of the course. Note that if you are looking for a mathematical proof of why deep learning models work, you will not find one. The same way you won’t find a mathematical proof of why Tylenol works or why airplanes fly.
3 Likes
You can actually call a recommended action a decision. A model which simply a formula makes a decision using the formula e.g y = mx + b. Therefore, to answer you question action can be a decision. A prediction could be just the result of a dot product which is a multiplication and addition process with a dot product.
This is something I know about: https://github.com/opencv/cvat . Also this one is great for small projects, you can start right away: https://www.makesense.ai/
However, most of the commercial tools I mentioned are free to use up to a certain quota, which should be enough for personal projects.
1 Like