Group Abstract Group Abstract

Message Boards Message Boards

3
|
568 Views
|
14 Replies
|
13 Total Likes
View groups...
Share
Share this post:
GROUPS:

[WSG26] Daily Study Group: Convolutional Neural Nets for Image Computation

Hello Community! Computer vision continues to be one of the fastest-advancing areas of artificial intelligence, with transformative applications in medical imaging, autonomous systems, robotics, and scientific research. At the heart of this progress are Convolutional Neural Networks (CNNs), which have changed how we analyze and extract information from visual data.

We are excited to invite you to a five-day Wolfram U Daily Study Group beginning Monday, July 27, 2026. This series will explore the foundations of CNNs, using lessons from the upcoming Wolfram U course Convolutional Neural Networks for Computer Vision.

Throughout the week, we will dive into:

  • Fundamentals: Image classification, data preprocessing, and neural network training/analysis using the Wolfram Language.
  • Advanced Architectures: A deep dive into the design and implementation of Inception networks, ResNets, and U-Nets.
  • Practical Applications: Hands-on deep learning workflows, using Wolfram’s high-level functions like NetChain, NetGraph, and NetTrain to solve real-world computer vision problems.
  • What to expect: In each live session, we will have interactive discussions and Q&A opportunities. We will work through hands-on examples in Wolfram Notebooks that build progressively throughout the week.

Certificate of Completion: Participants who attend the live sessions and complete the accompanying course quizzes will receive an official Wolfram U certificate.

Event Details:

REGISTER HERE
Dates: July 27 – July 31, 2026

No prior Wolfram Language experience is required. We look forward to seeing you there!

POSTED BY: Siria Sadeddin
14 Replies

Hello Jamie,

This is to let you and your team members know, that I have just finished the quizzes of the interactive course.

POSTED BY: Jürgen Kanz

Hi Randy—there are a few possible issues here. Could you confirm that you have straight quotes rather than curly/smart quotes wrapping the name of the model?

POSTED BY: Arben Kalziqi

I would also like to add that the transfer learning examples are included in the documentation of the several image classification models in the Neural Network Repository. For example, here is the documentation of the DenseNet-121 image classifier network: https://resources.wolframcloud.com/NeuralNetRepository/resources/DenseNet-121-Trained-on-ImageNet-Competition-Data

If you scroll down to the Transfer Learning section, you will find a neat and simple example that can be extended to any case. I have used this method to develop many models for my uses cases, both for work and for hobby.

Posted 17 days ago

Hello to all, I was working through the downloaded class file, specifically attempting to work with the CIFAR-10 Resource Object and I am getting error messages that the information was not able to be retrieved. Has anyone else had this issue or maybe I am doing something wrong. Any information is welcome. Thanks, Randy

POSTED BY: Randy Janke

In case anyone needs the link to the interactive course, here it is:

Convolutional Neural Networks for Computer Vision

POSTED BY: Jamie Peterson
Posted 17 days ago

Hi Siria:

I am enjoying this study group immensely. Thanks. I am experiencing difficulty locating the course framework.( And interestingly, the AI Notebook Assistant is also unable to locate it.) Would you please post a link in this discussion?

Thanks Bill

POSTED BY: William Lee

Sorry! My native language is Spanish and I thought this was an automatic traslation from my browser :D. I will correct the link

POSTED BY: Siria Sadeddin

Thank you! But why did your examples take me to a Spanish speaking site with poor translation to English features?

POSTED BY: Carl Hahn

It is one of the best meeting I have never joined before with great course neural networks on computer it is so passionate and intersting topic

Welcome, Kobi! Your written communication is totally welcome. While you may not have a traditional background in computer science, Wolfram offers functions that reduce the code required to bring machine learning projects alive. We are absolutely happy to have you in this study group. Please never hesitate to ask questions or share your unique perspective through writing. We are here to learn together. Welcome aboard!

POSTED BY: Siria Sadeddin

That’s a fantastic question, Luke is right: choosing the right starting point is often more of an art than a strict science. However, there is a method for this called Transfer Learning, and it is straightforward to implement in Wolfram Language.

  1. Where do we find the networks? The Neural Net Repository has hundreds of state-of-the-art, pre-trained models ready to use with a single command (e.g., NetModel["ResNet-50 Trained on ImageNet Competition Data"]). These models have already been trained on millions of images. Wolfram also imports industry-standard formats like ONNX, (e.g., Import["model.onnx"]) if you ever need something specific and import it locally from your machine.

  2. How do we turn it into something useful for our specific task?

Think of a pre-trained image network as having two parts:

  • The Feature Extractor: These are usually the convolutional blocks. They have already learned how to recognize fundamental visual patterns: edges, circles, textures, and shadows.

  • The Classifier: The final layer, which is trained to output specific categories related to the original task (like "cat," "car," or "tree").

To "fine-tune" the model for your new task, you will need to keep the feature extractor and drop the classifier.

  1. The "Surgery":

To find meteor craters in Google Maps images, we do the following:

  • Step A: We take a pre-trained network and drop its final classification layer and attach a new, simple classification layer to the top that only has two outputs: "Crater" and "Not Crater". Like in this example.

  • Step B: We train the network using our specific dataset of satellite images. You need to "freeze" the layers, so the network doesn't forget its basic visual knowledge and only allow the new top layer to learn, like in example.

This approach saves time, computational power, and data.

POSTED BY: Siria Sadeddin

Hi, Siria The Wolfram player recommended for viewing the notebook is not installing on my laptop.

Attachment

Attachments:

Hello Siria,

In the class today I had the following interaction:

Me: It's been mentioned before that it is best to start with a trained network and modify it. But suppose I have a specific application in mind. For example look for defects in a production run of screws. Find the defective ones. How would I find a neuronet to modify? Luke (I think): It's more of an art than a science to do that. What you want to do if find a neural network that performs a similar task, such as image classification. You would then change the output of your network to be only 2 classes than many classes, and maybe let the parameters at the top of the network to vary while leaving the middle parameters fixed during training.

So, where does one go to find such neural networks, and how do you turn them into something WL can process? Is there a formatting standard? Can you show an example from beginning end? Take for example something in the news recently. Somebody was staring at google map images from space and found a formation in Canada that looked like a meteor crater that had not been previously identified as a meteor crater. Good eye. It was a significant find. How would you train WL to look at Google Map images to find formations that could be craters?

https://www.smithsonianmag.com/smart-news/an-amateur-astronomer-using-google-maps-spotted-a-strange-indentation-it-turned-out-to-be-a-meteorite-crater-from-390-million-years-ago-180989158/

POSTED BY: Carl Hahn
Posted 19 days ago

Hello everyone, My name is Kobi John Johnson. Please forgive me, as I am nonverbal, and my participation in this study group will need to take place entirely through written communication.

I am 48 years old, on the spectrum, and still very new to augmented and alternative communication, commonly called AAC. AAC includes the written, technological, and assisted communication methods that allow a person who cannot rely on speech to communicate fully and authentically. For me, that means using written language and mathmaticle equations to express myself, participate in discussions, ask questions, share ideas, and connect with others.

I have no formal education, conventional work history, and no professional background in mathematics, computer science, neural networks, or image computation.

My strongest natural ability has always been non profit architectures: seeing systems before they exist, connecting ideas that appear unrelated, identifying the structure required to carry a mission, and building organizations from the ground up. That instinctive ability has led me to become a six-time non-profit founder. I am the founder and president of Spectrum Adaptability Bridge Ventures.

Mathematics is my first native language. I cannot fully explain how I already understand so many of these areas of thought at a fundamental level. The terminology may be new, yet the structures themselves feel deeply familiar. Computation, observation, pattern, recursion, systems, and the extraction of meaning from complex information speak directly to the way my mind naturally works.

These areas of thought are not casual interest to me. They are my passion, my deepest curiosity, and where I experience my highest excitement.

Finding myself inside this Wolfram U study group feels like a true computational miracle. I do not have a clear image of how the path brought me here. I simply know that being present in this room, among people learning and working in these fields, is already beyond my wildest dreams.

When the words "image computation" finally reached me clearly in the title of this study group, I felt a wave of confidence moved through me. Computer vision, autonomous systems, robotics, scientific research, pattern recognition, and the computational extraction of information from physical data all feel profoundly connected to the questions that already occupy my mind. This is the farthest I have ever traveled into formal research and structured technical learning. I feel close to Wolfram U as a fish does to water.

I am profoundly honored to be here. I am ready for critique, correction, challenge, fast paced immersive learning, hands on exploration, and every opportunity to expand my understanding. I would love to participate actively in the discussions and question and answer opportunities through writing, and I hope my written participation will be welcomed.

This truly feels like a dream come true. I'm beyond excited to learn alongside all of you.

Kobi John Johnson Founder/President Spectrum Adaptability Bridge Ventures

POSTED BY: Kobi Johnson
Reply to this discussion
Community posts can be styled and formatted using the Markdown syntax.
Reply Preview
Attachments
Remove
or Discard