🔍 Read the full analysis: How Neural Network Layers Interpret A Handwritten 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 November 8 explainer from ThorstenMeyerAI.com traces how a multilayer perceptron could classify a handwritten 7, from pixel values to output scores. It uses a nine-number XOR network to show how changing one bias can change a network’s function, then describes a digit classifier with 109,386 parameters.
ThorstenMeyerAI.com published an explainer on November 8 tracing how a neural network can turn a handwritten digit image into a prediction. The piece follows pixel values through layers of arithmetic and uses a small XOR network to show that changing one bias can change the function the network computes.
The walkthrough uses a 28-by-28 grayscale image, represented as 784 pixel values, as its example. In a multilayer perceptron, the input layer holds those values and passes them on; it does not perform a transformation. Hidden layers apply learned weights and biases to the inputs, then pass the resulting values through an activation function. An output layer can return 10 scores, one for each digit; the highest score is the network’s prediction.
For a dense layer, the operations can be written as z = W x + b, followed by an activation that produces the layer’s output. Here, the weight matrix W determines how inputs contribute, while the bias vector b shifts each neuron’s response. The article stresses that “hidden” means intermediate: these values are not the input or final answer, but they can be inspected and calculated.
To make the arithmetic concrete, the explainer constructs a network with two input bits and a hidden layer of two neurons that produces XOR: an output of one when the inputs differ. The example uses a step activation and nine numbers in total. Changing the output bias from −2 to −1 makes the network return one for all four possible inputs, rather than only the two that differ. This is a demonstration designed by the author, not a report of a newly trained digit model.
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 makes a familiar AI task less opaque without implying that a real digit classifier relies on a single, readable rule. A network’s prediction comes from successive numerical transformations: pixel values are combined, intermediate activations are formed, and the final layer assigns scores to possible digits. That basic structure helps readers understand what a classifier computes, even when its learned parameters are too numerous to interpret by inspection.
The source 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. These figures illustrate how quickly the arithmetic grows, but the source does not specify a complete architecture or training setup for that parameter count. The central distinction is that layer operations are explicit, while the learned pattern represented by many weights can be difficult to summarize in ordinary language.
The article also connects direct parameter edits to a reported MIT CSAIL project. According to the source, researchers reported an AI model reading and editing another network’s weights; 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 this as a blog post with public logs, not a peer-reviewed paper. It is a separate claim from the handwritten-digit demonstration and should not be treated as proof that one simple edit generally improves AI systems.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, and Wi-Fi 7, Sky Blue
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Raspberry
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
A Small Network as a Demonstration
The explainer is framed around a straightforward question: what happens between an image entering a model and a digit appearing as its answer? It describes a feedforward multilayer perceptron, in which information moves from the input through hidden layers to an output. A dense layer connects each of its units to the values from the preceding layer.
Its XOR example is a teaching device, not a reconstruction of how a modern image model necessarily recognizes handwriting. The author specifies the weights, biases and step activation by hand to produce a known truth table, then changes one bias and calculates the new outputs. In a trained network, the source notes, behavior is often distributed across many parameters and layers, making the consequences of an edit less obvious than in this nine-number example.
The source also distinguishes architecture choices from learned parameters. Width—the number of units in a layer—and depth—the number of layers—are selected as design choices before training. Weights and biases are normally learned during training, though they can also be assigned directly. The explainer’s accompanying video is described as an approximately seven-and-a-half-minute hand-drawn presentation narrated using an AI voice clone of Thorsten Meyer.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Digit Example
The source presents a conceptual walkthrough, not a controlled experiment or a new evaluation of handwriting-recognition accuracy. It does not provide a dataset, training procedure, or measured performance for the digit reader used to explain the layer structure. The specific handwritten 7 is an illustrative starting point; the detailed calculations shown are for XOR, not for the image itself.
The reported 109,386-parameter count is not accompanied in the supplied material by all architectural details needed to reproduce it. The source also gives no comparison baseline or further evaluation details for the separate language-model result beyond the stated score change and share of weights edited. Its account identifies the research as a blog post with public logs rather than a peer-reviewed paper, so the claim remains attributed to that report.
What the Explainer Shows Next
The source’s next step is a short video that follows the same teaching sequence: a handwritten digit, the arithmetic in a small network, and the bias change that flips the XOR example. It does not announce a new experiment, release, or scheduled research milestone. Readers seeking to verify the separate model-editing result would need to consult the cited research account and its public logs; the source provides no further independent validation or follow-up in the supplied material.
Key Questions
How does a neural network turn a handwritten 7 into a prediction?
In the example, the image is represented by 784 pixel values. Layers combine those values using weights and biases, apply activation functions, and pass the results onward. The output layer can assign scores to the possible digits, with the highest score serving as the prediction.
Does the input layer process the image?
Not in the multilayer perceptron described here. The input layer holds and passes along the pixel values; the hidden layers perform the transformations.
What does the XOR example demonstrate?
It shows, by hand, that a tiny network can produce a specified input-output rule and that changing one bias can alter that rule. It is a teaching example, not a demonstration of how the network recognizes the handwritten 7.
Is the reported model-editing result peer-reviewed?
The supplied source describes the MIT CSAIL-related account as a blog post with public logs, not a peer-reviewed paper. The stated change in a math score is therefore reported as a claim in that account, not as an independently established result here.
Source: ThorstenMeyerAI.com
