🔍 Read the full analysis: The Step-by-Step Math Of AI Recognizing A Handwritten 7 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 neural network does not use a hand-written rule for what a seven looks like. It transforms pixel values through layers of arithmetic, using learned weights and biases to produce scores for possible digits. A recent AI weight-editing demonstration offers a related example of changing a network by modifying its parameters, though the cited account is a blog post, not a peer-reviewed paper.
A neural network can classify an image of a handwritten 7 by passing its pixel values through layers of calculations, then selecting the highest of 10 digit scores. A November 8 explainer from ThorstenMeyerAI.com walks through that process, using a small hand-built network to show how weights and biases shape a result—and how changing one number can change a network’s behavior.
For a 28-by-28 image, the input can be represented as a list of 784 pixel values. The input layer holds and passes along those values; it does not transform them. In a dense, or fully connected, layer, each neuron multiplies incoming values by weights, adds a bias, and applies an activation function. The resulting values become the next layer’s inputs.
The explainer illustrates the arithmetic with a network for XOR, a function that returns one when two bits differ. Its two hidden neurons detect whether at least one bit is on and whether at most one is on. An output neuron combines those results. The hand-designed network uses nine numbers; changing its output bias from −2 to −1 makes it return one for all four possible inputs instead of only the two XOR cases.
The article then gives a parameter count for a small digit-reading network: 109,386 weights and biases, with about 92% in the first layer. That concentration follows from connecting every pixel input to each of 128 units in that layer. The count describes the example architecture, not a universal requirement for digit recognition.
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
Why Network Weights Matter
The walkthrough makes a practical point about AI systems: their outputs come from many numerical operations, not from a single rule that explicitly says what a seven is. For a small example, people can trace each calculation. In a larger trained model, a prediction can depend on patterns distributed across many weights and layers, making the same kind of direct explanation harder.
That distinction also matters for attempts to edit models. The source says MIT CSAIL researchers reported an AI model reading and editing another network’s parameters. According to the account, an edit affecting 0.016% of a language model’s weights coincided with a math score rising from 7.9% to 48.1%. Those figures are reported claims from the cited work, not independent confirmation of broad or reliable performance. They illustrate why understanding parameter changes may matter, while leaving open how well such methods generalize.
Products Worth Considering
Amazon Kindle Paperwhite 16GB, 7-Inch Display, Graphite
Apple iPad Air 11-inch (M4), 128GB, Wi-Fi 7, Space Gray
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Black
As an affiliate, we earn on qualifying purchases.
From XOR to Digit Images
The XOR example is a teaching device, not the method used to recognize a handwritten digit. It has two binary inputs and a step activation that outputs either zero or one. A digit image has hundreds of pixel values, and a classifier typically produces a score for each candidate digit; the predicted class is the one with the highest score.
The source describes a multilayer perceptron, or MLP: a feedforward network with dense layers. It distinguishes architecture choices from learned values. Width—the number of units in a layer—and depth—the number of layers—are selected as design choices. Weights and biases are parameters that training learns, though a person or program can also assign them directly. The source also notes that image-network visualizations have found early layers responding to simpler patterns, such as edges, and later layers to more complex features. That is background, not a claim that every network uses the same internal sequence.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Reported Results
The source does not specify the digit classifier’s training data, accuracy, evaluation conditions, or the exact image used in its example. The 109,386-parameter count applies to the stated small architecture and should not be read as a standard size for all digit-recognition systems.
Details about the reported language-model edit are also limited in the supplied material. It does not identify the model version, math benchmark, test setup, or whether the score change was replicated. The source describes the account as a blog post with public logs, not peer-reviewed research, so the cited numbers should be treated as reported results rather than settled evidence about model editing in general.
What Further Evidence Would Show
The immediate next step for readers is to inspect the underlying demonstration and its public logs, then check whether the model, task, scoring method, and evaluation conditions are documented. For the digit example, a fuller account would show the image values, the network’s actual weights and biases, and the resulting scores for all ten classes.
The source provides no announced publication date or follow-up experiment for the reported model-editing work. Until further details or independent evaluations are available, the walkthrough supports a basic explanation of neural-network arithmetic, while the broader significance of the weight-editing result remains open.
Key Questions
How does a neural network recognize a handwritten 7?
It represents the image as pixel values, transforms them through layers using weights, biases and activation functions, and produces scores for candidate digits. The highest score is the network’s prediction.
Does the input layer process the image?
In the MLP described by the source, the input layer holds and passes along the pixel values. The later layers perform the transformations.
What happens when one network number changes?
It can change the output. In the source’s XOR example, changing the output bias from −2 to −1 makes the network return one for every input combination.
Is the reported AI weight-editing result peer-reviewed?
The source describes it as a blog post with public logs, not a peer-reviewed paper. It reports a score change, but the supplied material does not provide enough evaluation details to judge how broadly the result applies.
Source: ThorstenMeyerAI.com
