🔍 Read the full analysis: What Goes On Inside A Neural Network When It Sees A 7? on ThorstenMeyerAI.com
Get the latest gadgets 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 neural network’s digit prediction comes from arithmetic applied in layers: weights and biases transform pixel values into scores for possible answers. Thorsten Meyer’s Nov. 8 explainer illustrates the process with a handwritten seven, a nine-number XOR network, and a small digit classifier with 109,386 parameters.
In an explainer published November 8, 2026, Thorsten Meyer traces how a neural network can turn the pixel values of a handwritten 7 into a digit prediction. The account shows the calculations inside a small network and demonstrates that changing a single bias can change a simple network’s output, while a digit-reading model contains 109,386 parameters.
A 28-by-28 image can be represented as a list of 784 pixel values. In the multilayer perceptron example, the input layer holds those values and passes them forward. Hidden layers transform them; an output layer returns scores for possible digits, and the digit with the highest score becomes the prediction. Each neuron combines inputs using weights and a bias, then applies an activation function.
Meyer illustrates the mechanics with XOR, a function that returns one when two bits differ and zero when they match. The worked network has two input values, two hidden neurons and one output neuron. With a step activation, the neurons’ calculations implement XOR using nine numbers in total. Changing the output bias from −2 to −1 makes the same network return one for all four possible input pairs.
The explainer says a small digit reader has 109,386 parameters, with about 92% in its first layer because each pixel connects to each of 128 units. It also describes a separate MIT CSAIL report in which researchers said an AI model directly read and edited another network’s parameters. According to Meyer’s account, one edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. The source describes that work as a blog post with public logs, not a peer-reviewed paper.
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 Pixels to a Digit Score
The example makes a commonly opaque process more concrete: a digit prediction is the result of successive numerical transformations, not a hand-written rule saying what a seven looks like. This does not mean a real model’s decision is easy to explain. In the toy XOR network, the behavior can be followed through a handful of numbers; in a trained model, behavior may depend on patterns spread across many parameters and layers.
The distinction matters when researchers try to interpret or edit models. Meyer’s account presents direct parameter editing as a reported demonstration, not proof that a small change will reliably produce a desired result in other systems. The stated score change is tied to the specific language model and evaluation described in the source; it should not be treated as a general measure of performance or as independently established research.
Products Worth Considering
Amazon Kindle Paperwhite 16GB, 7-Inch Display, Graphite
Apple iPad Pro 12.9-inch (2017), 512GB Wi-Fi, Space Gray
Sony WH-1000XM4 Wireless Noise-Canceling Headphones, Silver
As an affiliate, we earn on qualifying purchases.
A Toy Network Before the Digit Reader
The explainer uses a multilayer perceptron, a feedforward network made from dense layers. In a dense layer, each output unit receives input from every unit in the preceding layer. The layer can be written as z = W x + b, followed by an activation: x is the input, W contains weights, b contains biases, and the activation produces the values passed to the next layer.
The XOR example separates architectural choices from learned values. Width and depth describe how many units and layers a network has; these are choices made when designing it. Weights and biases are parameters that training usually learns, though they can also be assigned directly. The example is intentionally small enough to calculate by hand and is not itself a model recognizing handwritten digits.
Meyer’s video is described as a hand-drawn explainer of about seven and a half minutes, narrated with an AI voice clone of the author. The source also points to the MIT CSAIL parameter-editing account, while noting its status as a blog post with public logs rather than a peer-reviewed paper.
“It computes nothing.”
— Thorsten Meyer, describing the input layer
Limits of the Demonstration
The source does not specify the digit dataset, the trained model’s architecture or training procedure, or the accuracy of the handwritten-seven prediction. The 109,386-parameter count and the claim that about 92% sit in the first layer are presented for a small digit reader, but further model details are not supplied in the provided material.
For the MIT CSAIL example, the source does not name the language model, give the math evaluation’s dataset or test conditions, or explain whether the score change held across repeated runs. It describes a 0.016% weight edit and a score change from 7.9% to 48.1%, but those figures alone do not establish broader reliability. The cited account’s status as a blog post means the source does not present it as a peer-reviewed finding.
The XOR demonstration is a teaching example rather than evidence about how a trained image classifier internally represents a seven. In larger networks, the source says behavior may be distributed across many parameters; it does not provide a general method for identifying which values control a particular prediction.
What Further Evidence Would Show
The next step for readers seeking to check the reported parameter-editing result is to consult the public logs and original MIT CSAIL account cited by Meyer, including its model, evaluation, and editing details. Those materials would help establish what was changed and under which test conditions. The source gives no schedule for a peer-reviewed paper or for additional results.
For the digit example, a fuller account would need to name the dataset and model, show the input image and output scores, and explain how performance was measured. Until those details are available, the explainer supports a clear account of the basic layer-by-layer arithmetic, but not a detailed claim about the reliability or internal interpretation of a particular digit-reading system.
Key Questions
How does a neural network recognize a handwritten 7?
In the example, the image is represented by 784 pixel values. Layers combine those values with learned weights and biases, apply activation functions, and pass the results forward. The output layer scores possible digits; the highest score is the prediction.
What does a bias do in a neural network?
A bias shifts a neuron’s calculation before its activation is applied. Meyer’s XOR example shows that changing one output bias from −2 to −1 changes the network’s output for every input pair.
Does the XOR example explain how a real image model works?
It illustrates the arithmetic in a very small network, not the full behavior of a trained image classifier. The source says real-model behavior can be spread across many parameters and layers.
What is the reported MIT CSAIL parameter-editing result?
As described by Meyer, a model edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. The source characterizes the account as a blog post with public logs, not a peer-reviewed paper, and does not provide the evaluation details in the supplied material.
Source: ThorstenMeyerAI.com
