Any questions for Rober Monarch, the instructor for the AI for Good course(s) can go here. Anyone can answer them.
Question: How many data points (text messages) were there in the clinical message handling system for mothers? If there is a tradeoff between the quality of the model and the computation time to train it, then we need to understand when a problem will start to be a performance issue. Based on that, is the performance directly related to the number of messages that are going to go into training the model or am I misunderstanding something? What is the hardware being used (a cloud resource or a standalone machine in a truck)?|
Question: Robert mentions that the customers preferred to have the model updated quickly rather than have a better model that takes longer to train. It is not clear what is going on here. Below I try to reason it out. My question is do I have it right?
When we make a model based on a Neural Network (NN) better, we can do that by changing the topology of the NN or by training with more data.
Since we are talking about model Simple being better than model Complex, and there was no mention of training on more data to make the model better, but rather, it appeared the process was to toss everything and do a better model, I assume the change is in the topology of the network, so we start over with no trained information in the NN. If the change was the amount of training data, we could continue the training where we left off.
Do I have this correct?
Question: I’m not clear on this. Robert mentions there was a trade-off between the quality of the annotations and the quality of the results. Is this caused by the change in the annotation information and its impact on training; the same change and its impact on inferences because they are also annotated; or both? I suppose I’m asking whether or not the annotations are used/important for the data used in an inference. Is that the case?
Question: Is there ever an opportunity to push back against those who are restricting access to data because the benefit of lifting those restrictions is so great that it just happens? I’m asking this because there are discussions in the USA right now (I know this is political, but as soon as one says “privacy” one is saying “political”.) If so, based on your experience, is having the data visible and understandable that big of a deal, or can the system be designed using faux data to understand how to build it (formatting, field types, etc.), and then it runs fine even though you never saw the data. Is the question clear?
The lecture you linked does not provide the exact number of data points (text messages) that were in the clinical message handling system for mothers. However, it does mention that the data set contained text messages in multiple languages, and that the team had to identify the minimum number of labels or categories that should be in the training data. This suggests that the data set was relatively large, but not necessarily massive.
The lecture also mentions that the team investigated different performance tradeoffs, such as between a state-of-the-art machine learning model or a much simpler model. This suggests that the team was aware of the tradeoff between model quality and computation time, and that they were able to find a model that met their needs in terms of both quality and performance.
The number of messages can affect the training time, and it can also affect the accuracy of the model. In general, a model with more data will be more accurate, but it will also take longer to train. The data also has to be correctly labeled.
The hardware being used to train and deploy the AI model is also an important factor. A cloud resource, such as Google Cloud Products (GCP), Amazon Web Services (AWS) or Microsoft Azure, can provide the computing power and storage needed to train and deploy large AI models. However, a standalone machine in a truck may not have the same capabilities.
Ultimately, the performance of an AI model depends on a number of factors, including the data set, the model architecture, the hardware, and the specific use case.
Howdy, Canx. Now we are getting somewhere.
So, Robert created multiple models. What they were we do not know. The more complex model may have had a dynamic topology, but that would depend on how long ago the work was done.
Was there any case where the model used had starting point weights/biases which were arrived at by training with data, or were they all randomly or otherwise picked as starting points?
Question: Aside providing insight into live cases or references, do the additional reading resources attached to the various course modules contribute to overall grading?
No, the grading is only through the “graded” items, which I believe are only the quizzes for these courses.
Labs are provided to give you a hands on experience for what you are learning, and are not “graded” items. All the reading items are optional.
Many thanks for your response. I was wondering as I am currently on the 3rd module of the AI for Good Beginner’s course and have passed the first two quizzes. I hope to follow through with the 3rd quiz by coming Monday at the latest, after which I will find a convenient time to take additional courses.
I am still trying to fully understand how the dedicated communities operate and do hope you could be of assistance as and when the need arises.
At present, I am undertaking a PhD research on Advanced Tracking System for Real-Time Realignment of Land Mobile Satellite System. I hope I will be able to find a team to work with, especially in the aspect of Model Design and Coding. if you have any advice for me, I will be truly grateful and appreciative.
I’ll try my best to give advice, but I’m no expert in this this.
I believe all the best advices and practices have already been shared by Robert himself in these courses.
It is alright. I quite understand and appreciate your feedback.