🔍 Read the full analysis: How Does A Neural Network Decide An Image Shows A 7? 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 neural network does not rely on a hand-written rule for what a 7 looks like: it transforms image pixel values through layers of arithmetic and returns scores for possible digits. A small example shows how changing one bias can flip a network’s output, while a digit-reading network can contain 109,386 learned parameters.
A November 8 explainer from ThorstenMeyerAI.com shows how a neural network can classify a handwritten digit by passing image values through layers of calculations, rather than following a written rule for the shape of a 7. The article also uses a small logic example to show that changing one bias can change a network’s output, while noting that behavior in larger trained networks is generally spread across many parameters.
For a 28-by-28 image, the input can be represented as 784 pixel values. In the multilayer perceptron described in the explainer, the input layer holds and passes along those numbers. It does not itself calculate a prediction. Hidden layers then transform the values, and an output layer returns scores for the possible digits. The digit with the highest score becomes the network’s prediction.
Each neuron in a dense layer multiplies inputs by weights, adds a bias and applies an activation function. The weights control how much each input contributes; the bias shifts the total; and the activation determines the neuron’s output. Those outputs feed the next layer. The explainer writes the calculation as z = W x + b, followed by an activation that produces the next values.
To make the mechanics concrete, the article constructs a network for XOR, a function that returns one when two input bits differ. Its three neurons use a step activation, and its weights and biases amount to nine numbers. In the example, changing the output neuron’s bias from −2 to −1 makes the network return one for all four possible inputs, instead of only when the bits differ. This demonstrates a specific small network’s behavior; it is not evidence that one edit will have the same effect in a digit reader.
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 Predictions
The walkthrough makes a technical point relevant to anyone using image-recognition systems: a label such as “7” is an output of a calculation, not a human-readable rule stored in one place. The result depends on the values passed through the network and on its parameters. That helps explain why a model can produce a confident-looking classification without providing a simple, verbal account of how it recognized the handwriting.
The scale of the parameter set also matters. The source describes a small digit reader with 109,386 parameters, about 92% of them in its first layer because each pixel connects to each of 128 units. That concentration helps illustrate why inspecting or editing a trained network can be difficult: many numbers contribute to the computation, and a change may interact with other values. The source gives no comparison baseline for the parameter count, so it is best read as a count for the stated example.
The article connects this arithmetic to a separate claim about AI systems editing other models. It says MIT CSAIL researchers reported an AI model that directly read and edited network parameters, including an edit to 0.016% of a language model’s weights that raised a reported math score from 7.9% to 48.1%. Those figures and the description are attributed to the source’s account; the post itself says the material is a blog post with public logs, not a peer-reviewed paper.
Products Worth Considering
Dell UltraSharp U3225QE 31.5-Inch 4K UHD 120Hz IPS Black Thunderbolt Hub Monitor
FLEXISPOT ErgoX Ergonomic Office Chair with Footrest, Black
Dell 27 Plus 4K USB-C Monitor (S2725QC), 27-inch IPS, 120Hz
As an affiliate, we earn on qualifying purchases.
A Small Network as a Demonstration
The article focuses on a multilayer perceptron, or MLP: a feedforward network built from dense layers. “Dense” means each unit in one layer connects to the units in the next. The example is useful for showing the flow of numbers, but it is not a claim that every image-recognition system uses this exact architecture or works in the same way.
The XOR demonstration separates architecture from parameters. The number of layers and units—the network’s depth and width—are choices made when designing a model. Weights and biases are parameters that training normally learns, though they can also be assigned directly. In the hand-built XOR example, the author chooses values to produce a desired behavior, making it possible to trace the result by hand.
For larger image networks, the source says researchers have visualized early layers responding to simple patterns such as edges and later layers to object parts. These are descriptions of patterns found in some visualizations, not a guarantee that every layer has a single, easily named purpose. The explainer’s central contrast is that a tiny example can be followed step by step, while a trained model’s behavior may depend on patterns distributed across many parameters.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Reported Results
The explainer does not specify the training data, evaluation setup or model architecture behind the 109,386-parameter digit reader. The count and the estimate that about 92% of parameters sit in the first layer apply to the example as described, not necessarily to other digit classifiers.
The source also does not provide enough detail to independently assess the reported language-model edit. It gives the change in weights and math scores, but not the benchmark name, test conditions, number of runs or whether the result was replicated. The source identifies the account as a blog post with public logs, rather than a peer-reviewed paper. The reported score change should therefore remain attributed to that account, not treated as a general result about AI editing models.
It is also unclear from the article how much any individual parameter contributes to a trained digit classifier’s decision. The XOR example shows a one-number change in a deliberately constructed, very small network; it does not establish that a single parameter explains why a complex model identified a particular handwritten 7.
What Further Evidence Would Show
The explainer is a published educational account, not a report of a newly announced experiment or a scheduled research milestone. Readers seeking to verify the separate model-editing claim would need to consult the linked public logs and any underlying research materials, then check the benchmark, evaluation procedure and replication status. The source does not give a timetable for a paper or further results.
For the digit example, the next step in understanding a specific prediction would be to examine the model’s architecture, trained parameters, input image and output scores. The article offers the arithmetic framework for doing that, but it does not supply a particular image’s intermediate values or explain a deployed classifier’s individual decision. Until those details are available, the confirmed takeaway is narrower: layers transform pixel values into scores, and parameter changes can alter a network’s behavior, with the effect depending on the network and the change.
Key Questions
How does a neural network decide an image shows a 7?
In the example, the image is represented by 784 pixel values. Layers transform those values using weights, biases and activation functions. The output layer produces scores for digit choices, and the highest score supplies the prediction.
Does a neural network have a written rule for the shape of a 7?
Not in the account described here. The network applies learned numerical parameters to the pixel values; the explainer does not describe a manually written rule defining what a 7 looks like.
What does a bias do in the XOR example?
A bias shifts a neuron’s calculated total before its activation. In the source’s hand-built XOR network, changing the output bias from −2 to −1 changes the result for all four input combinations.
Does one changed parameter usually flip a large model’s answer?
The source does not establish that. Its one-number demonstration uses a deliberately small network. It says that in trained networks, behavior is often distributed across many parameters and layers.
What is known about the reported AI model-editing result?
ThorstenMeyerAI.com reports that an AI model changed 0.016% of a language model’s weights and that a math score rose from 7.9% to 48.1%. The source calls its account a blog post with public logs, not a peer-reviewed paper, and does not provide enough evaluation detail in the supplied material to independently assess the result.
Source: ThorstenMeyerAI.com
