Teaching tool for animating large model principles in Excel

General Introduction

AI by Hand is an educational website focused on teaching Artificial Intelligence (AI) model building through Excel, created and maintained by Prof. Tom Yeh. It helps users manually implement AI algorithms such as Neural Networks, Transformer, etc. in Excel by providing a series of free spreadsheet templates and detailed tutorials. The goal of the site is to allow learners to understand the math and logic behind AI through hands-on practice, making it suitable for students, teachers, and beginners interested in AI. The content ranges from basic multilayer perceptrons (MLPs) to complex computer vision models, emphasizing the value of "hands-on" computation. Updated with Substack blogs, the site provides interactive learning resources that are popular with educators and learners around the world.

Function List

• Provide Excel templates for download: Users can download spreadsheets of predefined formulas and structures for direct use in AI model calculations.
• Support for teaching multiple AI models: Includes Excel implementations of models such as MLP, Transformer, RNN, and Backpropagation.
• Real-time update of tutorial content: New tutorials and exercises are released regularly through the Substack platform.
• Custom model parameters: The user can adjust the weights, biases and other parameters in the table and observe the changes in the calculation results.
• Video Demo Support: Some of the tutorials are accompanied by videos showing how to manipulate complex algorithms in Excel.
• open source commons: Some of the content is open-sourced on GitHub for users to freely modify and contribute.
• Interactive feedback mechanisms: Users can interact with the authors through comments or emails to suggest improvements or report bugs.

Using Help

1. Visiting the website and accessing resources

• move: Open your browser and enter the URLhttps://www.byhand.ai/t/spreadsheet, go to AI by Hand's Spreadsheet page.
• manipulate: The page displays a Google Sheets link (e.g.https://by-hand.ai/sp/tfmr), click on it to view the Excel template for the Transformer model.
• draw attention to sth.: It is recommended to sign up for a Substack account to subscribe to Tom Yeh's blog to get the latest templates and tutorials pushed out.

2. Download or copy Excel templates

• Download: On the Google Sheets page, click "File" > "Download" > "Microsoft Excel (.xlsx)" to save it locally. locally.
• Reproduction method: Click File > Make a copy, save the template to your Google Drive, and then edit it online.
• caveat: Make sure your version of Excel supports formula calculations (e.g. SUM, PRODUCT, etc.), Excel 2016 or higher is recommended.

3. Detailed operation flow of main functions

Function 1: Learning the Transformer model using Excel templates

• intend: Open the downloaded Transformer template (e.g.tfmr.xlsx), you will see multiple worksheets, including input layers, weight matrices, and output calculation regions.
• workflow::
  1. input data: Fill the "Input" worksheet with test data, e.g. a simple vector of sentences (in numeric form).
  2. Adjustment parameters: Go to the "Weights" worksheet and change the weights and biases (e.g. change a weight from 0.5 to 0.8).
  3. View Calculation Process: Switch to the "Forward" worksheet and the table will automatically calculate the forward propagation results, showing the intermediate values at each step.
  4. Check Output: View the final result in the Output worksheet to understand how the Transformer's attention mechanism affects the output.
• Featured Functions: Templates have built-in formulas (e.g., matrix multiplication MMULT) and visualization charts that allow users to intuitively observe model behavior by adjusting parameters.
• finesse: If the calculation result is abnormal, check whether the formula reference range is correct, or refer to the tutorial video on the website.

Function 2: Manual implementation of Backpropagation

• intend: Download Backpropagation template from the website (reference)https://www.byhand.ai(the Backpropagation article).
• workflow::
  1. Setting up the network structure: Enter the initial parameters of a three-layer network in the template (e.g., 2 neurons in the input layer, 3 in the hidden layer, and 1 in the output layer).
  2. Fill in the training data: Enter the sample data and desired output in the "Data" worksheet (e.g., enter [0.1, 0.2] and expect an output of 0.7).
  3. Compute forward propagation: Go to the Forward Pass worksheet and observe the output for each layer.
  4. Perform backpropagation: In the Backward Pass worksheet, the table automatically calculates the gradient based on the loss function and updates the weights.
  5. Iterative adjustments: Run steps 3 and 4 several times and observe how the weights are progressively optimized.
• Featured Functions: Through manual inputs and calculations, users can gain insight into the mathematics of backpropagation, and the templates are labeled with key formulas (e.g., ∂L/∂w).
• suggestion: For initial use, it is recommended to work step-by-step with articles on the website (e.g., Backpropagation tutorial on October 7, 2024).

Function 3: Customized model parameter experiment

• intend: Select any template (e.g. MLP or RNN) and make sure it has been copied locally.
• workflow::
  1. Open parameter area: Find the cell areas labeled "Weights" and "Biases".
  2. modified value: Change a weight from the default value (e.g., 0.3) to another value (e.g., 1.2), or adjust the bias.
  3. running calculation: Press Enter or refresh the table and observe how the output changes.
  4. Comparative results: Record differences in output with different parameters to understand the effect of parameters on the model.
• Featured FunctionsThis "trial and error" approach allows users to visualize the sensitivity of the AI model, which is suitable for teaching or experimentation.
• tip: If the result is more than expected, you can use Excel's "Undo" function (Ctrl+Z) to restore the original value.

4. Getting more help

• video tutorial: Visit AI by Hand's YouTube channel (e.g. DeepSeek Lecture) to see Tom Yeh or an assistant in action.
• Community Interaction: Leave a comment under the Substack article with a question or share your template improvements and the author will usually reply.
• Resources for Advancement: Explore other pages of the site (e.g.https://www.byhand.aihomepage) for more model templates (e.g., AlphaFold, LSTM).

5. Cautions

• Equipment Requirements: Ensure that Excel or Google Sheets supports complex formulas and chart rendering, which may not display properly in lower versions.
• Learning Advice: Beginners can start with simple MLP templates and progressively challenge themselves with complex Transformer or computer vision models.
• Saving Progress: Save the file periodically to avoid data loss due to misuse.

Through the above steps, users can quickly get started with AI by Hand's Excel templates and master the core principles of AI modeling in practice. The biggest highlight of the website is that no programming foundation is required, and AI learning can be accomplished only through familiar Excel, which greatly reduces the threshold of entry.