🔍 Read the full analysis: How A Neural Network Converts A Handwritten 7 Into An Answer on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 8 explainer from ThorstenMeyerAI.com traces how a small feedforward neural network could process the 784 pixel values in a 28-by-28 handwritten digit image and choose among ten digit scores. Its examples show that network behavior comes from weights, biases and activations, while a separate report about AI editing model weights is described as a blog post, not a peer-reviewed paper.
ThorstenMeyerAI.com published an explainer on November 8 tracing how a feedforward neural network can turn a handwritten seven into a digit prediction. The account describes a network passing pixel values through layers of weighted sums, biases and activation functions; its point is that the computation is inspectable arithmetic, even when a real model contains many parameters.
For the example, a handwritten image is represented as 784 pixel values, corresponding to a 28-by-28 image flattened into a list. The input layer holds those values and passes them forward. Hidden layers transform them, and the output layer can produce ten scores, one for each digit from zero through nine. The digit with the highest score is the network’s prediction. The explainer does not provide a specific image, trained model, or measured recognition result for a particular seven.
Each neuron in a dense layer multiplies its inputs by learned weights, adds a bias, then applies an activation function. The resulting values become the next layer’s inputs. In compact notation, the layer computes z = W x + b, then applies an activation to z. The input layer does not perform this transformation; “hidden” refers to intermediate values between the supplied image and the output.
To make the mechanics concrete, the article builds a small network for XOR, a two-bit function that returns one when its inputs differ. Its two hidden neurons test whether at least one bit is on and whether at most one bit is on. The output returns one when both conditions hold. The example uses nine numbers in total. Changing the output bias from −2 to −1 makes the same network return one for every input, illustrating how a parameter change can alter behavior.
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
From Pixel Values to Digit Scores
The walkthrough offers readers a practical way to understand what a digit classifier does without treating its answer as a mysterious leap. A network’s prediction follows from successive numerical transformations: pixels enter, intermediate activations are calculated, and output scores are compared. That matters because it makes clear both what can be inspected—the arithmetic and parameter values—and why interpreting a large model can still be difficult.
The XOR demonstration also distinguishes architecture from parameters. Width and depth describe how many units and layers the designer selects; weights and biases are values typically learned during training, though they can also be assigned directly. A one-number change flips the toy network’s function, but this should not be taken to mean a single edit will generally explain or control a trained digit reader. The source says real behaviors often depend on patterns across many parameters and layers.
Popular right now
Breville Bambino Espresso Machine, Brushed Stainless Steel
Ninja Luxe Café Mini Espresso and Drip Coffee Machine, ES301
Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee Maker, and Cold Brew Maker
As an affiliate, we earn on qualifying purchases.
A Toy Network Before the Digit Reader
The article uses a multilayer perceptron, or MLP, a feedforward model built from dense layers. It first follows two binary inputs through the XOR example, where the arithmetic is small enough to calculate by hand. This provides a step-by-step illustration of how outputs from one layer become inputs to another, before moving to the larger scale of an image represented by hundreds of pixel values.
For the digit model described in the source, the network has 109,386 parameters, with about 92% in the first layer because each pixel connects to each of its 128 units. Those figures describe the example architecture in the source; they are not a universal parameter count for handwritten-digit recognition. The source also points to a separate claim about MIT CSAIL researchers: an AI model reportedly read and edited other networks’ numerical weights. It says one edit changed 0.016% of a language model’s weights and moved a math score from 7.9% to 48.1%. The source identifies that account as a blog post with public logs, not a peer-reviewed paper, so the result should be treated as a reported claim rather than an established research finding.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
What the Example Does Not Establish
The source does not report a test of a particular handwritten seven, the model’s accuracy, or how confidently it assigns scores to each possible digit. It explains a general MLP process rather than documenting a new benchmark or experiment. The 109,386-parameter count applies to the described example, not necessarily to other image-recognition systems.
The reported AI weight-editing result also has limits in the supplied account: its underlying public logs, evaluation setup and independent review are not detailed here. The source explicitly says it is a blog post, not a peer-reviewed paper. It is therefore unclear from this material how broadly the reported score change applies or whether the result has been independently reproduced.
How to Read the Next Claim
The explainer points readers to an accompanying video of about seven and a half minutes, described as covering the path from a handwritten seven to the XOR bias example. For the separate weight-editing claim, readers would need to examine the referenced public logs and any later technical publication or independent replication before drawing conclusions about its scope. No further test or publication date is given in the supplied material.
For a reader following the digit example, the next step is to distinguish the network’s mechanics from the performance of a specific trained model. The arithmetic explains how a prediction is produced; evidence such as a test set, accuracy figures and evaluation details would be needed to assess how reliably that model recognizes handwriting.
Key Questions
How does the network recognize a handwritten seven?
It represents the image as 784 pixel values, passes them through weighted layers and produces scores for possible digits. The digit with the highest score is the prediction. The source does not give a measured result for one particular seven.
What do weights and biases do?
Weights determine how much each input contributes to a neuron’s total. A bias shifts that total before the activation function converts it into an output.
Does changing one number usually change a trained model’s answer?
The XOR example shows that one bias can change a small, deliberately constructed network’s function. The source cautions that behaviors in trained networks often depend on patterns spread across many parameters and layers, so the toy example is not a general rule about model edits.
Is the report about AI editing another model’s weights peer-reviewed?
Not according to the supplied source. It describes the reported result as a blog post with public logs, rather than a peer-reviewed paper. The available material does not establish independent replication.
Source: ThorstenMeyerAI.com
