🔍 Read the full analysis: Trace The Data: How A Neural Network Recognizes A 7 on ThorstenMeyerAI.com
Get hardware and tech essentials 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, 2026, explainer from ThorstenMeyerAI.com traces how a dense neural network turns the 784 pixel values in a 28-by-28 image into scores for handwritten digits. It uses a nine-number XOR network to show how one bias can change behavior, then describes a separate MIT CSAIL report about AI editing model weights.
ThorstenMeyerAI.com published an explainer on November 8 tracing how a neural network can classify an image of a handwritten seven, from its pixel values through layers of arithmetic to a digit prediction. The article uses a small XOR example to show how changing one bias can flip a network’s output, then describes a larger digit-recognition model with 109,386 parameters.
For its example, the article treats a handwritten digit as a 28-by-28 image containing 784 pixel values. In a multilayer perceptron, or MLP, the input layer holds and passes along those values; it does not transform them. Dense hidden layers multiply inputs by weights, add biases and apply activation functions. The output layer returns scores for the possible digits, and the highest score is the prediction.
To make the arithmetic easier to follow, the explainer builds a toy network for XOR, a function that returns one when two input bits differ and zero when they match. Its network has two input values, two hidden neurons and one output neuron, with nine weights and biases. The hidden neurons test whether at least one bit is on and whether at most one is on; both tests are true when exactly one bit is on.
The article then changes the output neuron’s bias from −2 to −1. With the other numbers unchanged, the network returns one for all four possible input pairs instead of computing XOR. The explainer says a digit reader with 784 inputs and a first hidden layer of 128 units has 109,386 parameters, about 92% of them in that first layer. The article also points to a separate MIT CSAIL account of AI editing neural-network weights; those reported results are not presented as part of the handwritten-digit demonstration.
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 Scores to Model Edits
The explainer makes the path from an image to a classification more concrete: a prediction emerges from successive numerical transformations, not from a hand-written rule stating what a seven looks like. That distinction matters for readers trying to understand both what neural networks compute and why their behavior can be difficult to explain. The individual operations are accessible, while the number of parameters and interactions can make the full system difficult to interpret.
The XOR example also clarifies why parameter changes matter. In the toy network, one bias has a clear, visible effect because the model was designed around a simple function. A trained digit classifier is more complicated: the article cautions that its behavior may depend on patterns distributed across many parameters and layers. Changing a single value in a larger model does not necessarily produce an equally simple or predictable result.
The source further connects this idea to what it calls weight-space programming: directly changing model parameters rather than only prompting a model or retraining it in the usual way. It cites an MIT CSAIL report in which an AI model edited another network’s numbers. That could matter to research into how models are inspected or modified, but the source describes the material as a blog post with public logs, not a peer-reviewed paper. The reported result should be read with that qualification.
Products Worth Considering
Ledger Nano X Bluetooth Cryptocurrency Wallet
Trezor Safe 5 Crypto Hardware Wallet with Color Touchscreen
LG UltraGear 45GX900A-B 45-Inch Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
The Arithmetic Behind Each Layer
The article focuses on a feedforward multilayer perceptron, where information moves from input to hidden layers and then to output. For a dense layer, the calculation can be written as z = W x + b, followed by h = activation(z). Here, x is the incoming vector, W contains the weights, b contains the biases, and the activation function converts intermediate totals into the values passed onward.
Weights determine how strongly inputs contribute to a neuron’s total; biases shift the point at which it responds. The article calls intermediate layers “hidden” because they are neither the supplied input nor the final answer—not because their arithmetic is inaccessible. Its hand-built XOR example uses a step activation that returns one when a total is zero or greater and zero otherwise.
The article also distinguishes architecture choices from learned values. Width describes the number of units in a layer, and depth the number of layers; these are chosen as architectural hyperparameters. Weights and biases are model parameters that training normally learns, although they can also be assigned directly in a deliberately simple example.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Demonstration
The XOR network is a hand-designed teaching example, not a demonstration that a trained digit classifier can be understood by changing one parameter. The source does not provide the trained digit model’s full architecture, its evaluation results, or a test showing how any particular weight change affects its recognition accuracy. Its count of 109,386 parameters is given for the described small digit reader, not as a count for every digit-recognition system.
The MIT CSAIL figures also need qualification. The source reports that an edit affecting 0.016% of a language model’s weights coincided with a math score rising from 7.9% to 48.1%, and says public logs are available. It does not provide enough detail here to assess the benchmark, comparison conditions, repeatability or whether the result has been peer reviewed. Those points remain unclear from the supplied material.
Further Tests of Weight Editing
The source does not announce a scheduled follow-up, release date or next research milestone. Readers seeking more detail on the reported model-editing result would need to consult the MIT CSAIL account and its public logs for the underlying methods and evaluation conditions. The explainer itself also links to a roughly seven-and-a-half-minute video covering the pixel-to-prediction process and the XOR bias change.
The practical next step for judging the broader claim is independent scrutiny of the reported editing experiment: what model and benchmark were used, how the score was measured, and whether the change holds across repeated tests. Until those details are available in the material provided, the numerical result remains a reported example rather than a settled finding.
Key Questions
How does the network identify a handwritten seven?
In the explainer’s example, the network receives 784 pixel values from a 28-by-28 image. Hidden layers transform those values using weights, biases and activation functions, and the output layer returns scores for possible digits. The highest score is the predicted digit.
Does the input layer perform calculations?
Not in the MLP described. The input layer holds and passes on the pixel values; the hidden layers perform the transformations.
What does changing the XOR network’s bias show?
Changing the output bias from −2 to −1 makes the toy network return one for every input pair rather than computing XOR. It illustrates how a parameter can affect behavior in a network designed to make that effect easy to trace.
Is the reported MIT CSAIL result peer reviewed?
The source describes it as a blog post with public logs, not a peer-reviewed paper. The supplied material does not establish whether the reported score change has been independently replicated.
Source: ThorstenMeyerAI.com
