🔍 Read the full analysis: How Does AI Know A Handwritten Digit Is A 7? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A neural network recognizes a handwritten digit by passing pixel values through layers of weighted calculations, then selecting the highest of ten output scores. A worked example explains the process and counts 109,386 parameters in a small digit reader; it is an educational explainer, not a newly reported research result.
A November 8 explainer from ThorstenMeyerAI.com traces how a neural network can identify a handwritten seven: it receives the image as pixel values, processes them through layers of calculations, and returns scores for possible digits. The article also works through a tiny network in which changing one bias changes the output, illustrating how a model’s parameters shape its behavior.
For its example, the explainer uses a 28-by-28 image, represented as a list of 784 pixel values. In a multilayer perceptron, the input layer holds and passes on those values. Hidden layers calculate intermediate results, and the output layer produces ten scores, one for each digit. The largest score determines the network’s prediction in this example; that is a classification result, not a human-like explanation of the handwriting.
Each neuron in a dense layer multiplies its inputs by weights, adds a bias, then applies an activation function. The weights influence how much each input contributes, while the bias shifts the neuron’s response. The layer’s outputs become the next layer’s inputs. These steps can be written as z = W x + b, followed by an activation. The explainer stresses that the arithmetic is visible in principle; the difficulty is following the many operations and parameters in a trained network.
To demonstrate how a parameter can affect behavior, the article constructs a small network for XOR, a function that returns one when two input bits differ. It gives the network nine numbers—weights and biases—and shows that changing the output bias from −2 to −1 makes the network return one for all four possible inputs. This is a deliberately simple example, not a description of how a real image classifier’s learned behavior is usually organized.
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 Digit Scores
The example gives readers a concrete way to understand what a digit-recognition prediction represents: a sequence of numerical transformations, ending in competing scores. That distinction matters because saying a model “recognizes” a seven can sound as though it applies a clear written rule. In the account provided, the network instead computes outputs from pixel values using parameters learned during training.
The explainer says its example digit reader has 109,386 parameters, with about 92% in the first layer because each pixel connects to each of 128 units. The count conveys how quickly even a modest dense network can contain many adjustable values. It does not mean all those values have a simple, individually interpretable role: the article notes that trained-model behavior can be distributed across parameters and layers.
The source also points to work it attributes to MIT CSAIL researchers, saying an AI model read and edited other networks’ weights. It reports that 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 claims in the source article, which describes the work as a blog post with public logs rather than a peer-reviewed paper. They are separate from the handwritten-digit demonstration and should not be read as independently verified findings here.
Popular right now
Breville Barista Express Espresso Machine with Built-In Grinder, Brushed Stainless Steel
Ninja Luxe Café Mini Espresso and Drip Coffee Machine, Stainless Steel, ES301
TCL 75-Inch QM6K QLED Mini LED 4K HDR Smart Google TV, 144Hz
As an affiliate, we earn on qualifying purchases.
A Small Network as a Demonstration
The article uses a multilayer perceptron, or MLP: a feedforward network with dense layers. “Dense” means each unit in a layer receives input from every unit in the preceding layer. The input layer does not perform the transformation; the hidden layers do. Their name refers to their position between the supplied input and the final output, not to a process that is inherently inaccessible.
The XOR example provides a hand-checkable demonstration before the article turns to image classification. Its hidden units test whether at least one input bit is on and whether at most one is on. Both conditions hold when the bits differ. A final unit combines those intermediate results. The example shows how a network can implement a specified behavior, while a trained digit model has parameters learned from data rather than written out by a person in this way.
The source distinguishes architecture choices from learned parameters. Width describes the number of units in a layer and depth the number of layers; these are choices made when designing a network. Weights and biases are parameters that training normally learns, although they can also be assigned directly in a constructed example.
““A neural network is arithmetic in layers.””
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Digit Example
The source does not specify the digit model’s training data, training procedure, test results, or accuracy. It explains how a network of this type can produce a digit prediction, but it does not establish how well a particular model recognizes handwritten sevens across different writers, image quality, or writing styles. A maximum score identifies the model’s selected class in the example; no score values or confidence calibration are supplied.
The reported MIT CSAIL weight-editing result also needs qualification: the source calls its basis a blog post with public logs, not a peer-reviewed paper. It does not give enough detail here to evaluate the math benchmark, its test setup, or whether the score change would hold under other conditions. The report should not be treated as proof that a single weight edit generally controls a complex model’s behavior.
What the Explainer Leaves Open
The explainer’s next step is conceptual rather than a scheduled research milestone: it moves from a nine-number XOR network to the much larger parameter count of a digit reader, then introduces the idea of directly editing model weights. Readers seeking to assess a specific handwritten-digit system would need its architecture, training and evaluation details, including results on images beyond the examples used to explain the mechanism.
For the separately cited weight-editing work, the source points readers to public logs, but gives no peer-reviewed publication date or follow-up results. Whether the reported score change can be reproduced or generalized remains open on the information provided. The article’s video is described as a roughly seven-and-a-half-minute hand-drawn explanation; it is narrated using an AI voice clone of Meyer, according to the source.
Key Questions
How does a neural network decide an image is a 7?
It processes the image’s pixel values through layers of weighted calculations. In the example, the output layer returns a score for each digit, and the largest score is the prediction.
Does the network use a written rule for the shape of a 7?
Not in the account described. The network uses weights and biases that training normally learns from data; the article does not describe a manually written rule defining a seven.
What does changing one bias show?
In the deliberately small XOR example, changing the output bias from −2 to −1 changes the network’s result for every input. It demonstrates the effect of a parameter in that toy network, not a general one-parameter explanation of digit recognition.
What does 109,386 refer to?
It is the parameter count the source gives for its small digit reader. The article says about 92% are in the first layer, where each pixel connects to each of 128 units.
Does this article report a new digit-recognition result?
No. It is an educational explainer published November 8, 2026, about how layered calculations can produce a digit prediction. The separate weight-editing result is attributed to a source-described blog post, not presented as a peer-reviewed paper.
Source: ThorstenMeyerAI.com
