🔍 Read the full analysis: How A Neural Network Reads A Handwritten 7 In Stages on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A November 8 explainer from ThorstenMeyerAI.com traces how a multilayer perceptron can turn the 784 pixel values in a 28-by-28 image into scores for the digits 0 through 9. It uses a nine-number XOR example to show how weights and biases shape behavior, then describes a separate, attributed report about AI editing another model’s parameters.
ThorstenMeyerAI.com published an explainer on November 8 tracing how a neural network can classify a handwritten seven, from the image’s pixel values to the model’s output scores. The article describes a multilayer perceptron as a sequence of calculations using weights, biases and activation functions, and uses a small XOR network to show how changing one number can change a network’s behavior.
For the digit example, a 28-by-28 image is flattened into a list of 784 pixel values. The input layer holds and passes those values on; it does not perform the transformations described in the explainer. Hidden layers calculate intermediate results, and the output layer can return ten scores, one for each digit. The highest score is treated as the prediction.
In a dense layer, each output unit combines inputs by multiplying them by weights, adding a bias, and applying an activation function. The article summarizes this as z = W x + b, followed by an activation that produces the next layer’s values. Weights affect how strongly inputs count, while biases shift the point at which a unit responds.
To make the arithmetic easier to follow, the explainer builds a small network for XOR, a function that returns one when two input bits differ. Two hidden units detect whether at least one bit is on and whether at most one bit is on; together those conditions identify inputs with exactly one active bit. The complete example uses nine numbers. Changing its output bias from −2 to −1 makes it return one for all four possible inputs, according to the article.
Infographic 01 · One bias, two behaviors
A network you can run by hand
Two hidden neurons check "at least one bit is on" and "at most one bit is on". The output neuron needs both. Change its bias and it needs only one.
Move the output bias from −2 to −1 and the network stops computing XOR: it returns 1 for all four inputs.
The whole network
Hidden 1: step( x₁ + x₂ − 1 ) Hidden 2: step(−x₁ − x₂ + 1 ) Output: step( h₁ + h₂ − 2 )
step: 1 if the total is zero or greater, otherwise 0. Nine numbers in total: six in the hidden layer, three in the output neuron.
All four inputs
Hidden outputs and the final output with the original and the changed bias
| Input | At least one | At most one | Output, bias −2 | Output, bias −1 |
|---|---|---|---|---|
| 0, 0 | 0 | 1 | 0 | 1 |
| 0, 1 | 1 | 1 | 1 | 1 |
| 1, 0 | 1 | 1 | 1 | 1 |
| 1, 1 | 1 | 0 | 0 | 1 |
Sources: own arithmetic, checked with a script · the XOR example as in Goodfellow, Bengio, Courville: Deep Learning, ch. 6.
Powered by Thorsten Meyer AI
Why One Bias Can Change a Prediction
The examples make a practical distinction between a network’s architecture and its learned parameters. Width and depth describe how many units and layers it has; weights and biases determine how those units respond to inputs. A network can retain the same structure while a parameter change alters its output, as the XOR example demonstrates.
That distinction matters when people try to understand or modify trained models. In the toy network, the effect of one bias is easy to calculate because the network has only nine numbers and a simple activation. In a digit classifier, the source says, a compact model has 109,386 parameters, with about 92% in its first layer because each pixel connects to each of 128 units. The small example illustrates the principle, but does not establish that changing one parameter in a trained digit reader will have a similarly clear effect.
The source also points to a reported experiment in which an AI model edited another model’s weights. It says one edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. Those figures are attributed to the source’s account of MIT CSAIL researchers’ work; the article says the material is a blog post with public logs, not a peer-reviewed paper. They should not be read as evidence that the same approach will work across models or tasks.
Popular right now
De'Longhi La Specialista Opera Espresso Machine with Coffee Grinder
Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee Maker, and Cold Brew Maker
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
From Pixel Values to Layer Outputs
The article focuses on a feedforward network, also called a multilayer perceptron. Information moves from the input through hidden layers to an output. “Hidden” refers to intermediate values rather than something inaccessible: each layer’s calculations can be written out, even if the number of calculations in a large model is difficult for a person to inspect.
In the digit-reading example, the network does not receive a written rule defining the shape of a seven. Instead, it processes the image’s numeric pixel representation and produces scores for possible digits. The explainer does not provide a particular model’s training history, image, prediction confidence, or measured accuracy. Its focus is the general path through the layers, not a report of a newly tested classifier.
The source accompanies the article with a video described as a roughly seven-and-a-half-minute hand-drawn explainer. It says the narration uses an AI voice clone of Thorsten Meyer. The same piece connects its basic demonstration of direct parameter changes to a separate report about AI systems editing model weights, while distinguishing that blog-post account from peer-reviewed research.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
What the Examples Do Not Establish
The source’s digit example explains a general network structure, but it does not identify a specific trained classifier or show the model processing a particular image of a seven. It gives no test results, accuracy figures, or confidence scores for that example. The explanation therefore shows how a prediction can be computed, not how reliably a particular system reads handwriting.
The reported weight-editing result is also limited by the source’s description. It characterizes the material as a blog post with public logs, rather than a peer-reviewed paper, and does not provide enough detail here to independently assess the experiment’s setup, evaluation, or reproducibility. The reported score change and fraction of weights edited remain claims attributed to that account, not a general performance guarantee.
Further Evidence on Weight Editing
The immediate next step for readers is to follow the source’s linked video and inspect the public logs it cites for the separate weight-editing report. Those materials may provide more detail about which model was edited, how the math score was measured, and what the intervention involved. The explainer itself does not announce a new experiment or give a schedule for further results.
For the broader claim about AI models editing other networks, stronger evidence would include a full account of the methods and evaluation, independent replication, and results across more than one task or model. Until such information is available, the XOR demonstration is best understood as a clear illustration of parameter sensitivity, while the reported language-model score change remains a specific result attributed to the source.
Key Questions
What does a neural network receive when it reads a handwritten seven?
In the example, it receives 784 pixel values representing a 28-by-28 image. The input layer holds those numbers and passes them to the next layer.
How does the network turn pixels into a digit prediction?
Hidden layers combine their inputs using weights and biases, then apply activation functions. The output layer can give a score for each of the ten digits; the highest score is the predicted digit in the explainer.
Why does changing one number matter in the XOR example?
The small network’s output depends on its bias as well as its inputs. Changing the output bias from −2 to −1 makes the example return one for all four two-bit inputs, rather than only when the bits differ.
Does the article show that an AI can reliably edit other AI models?
No. It reports a weight-editing result attributed to a separate account and says that account is a blog post with public logs, not a peer-reviewed paper. The source does not establish that the result generalizes to other models or tasks.
Source: ThorstenMeyerAI.com
