🔍 Read the full analysis: A Closer Look At How AI Identifies A Handwritten 7 on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A November 8 explainer from ThorstenMeyerAI.com follows a handwritten digit through a neural network, showing how pixel values and learned weights combine to produce a prediction. It also uses a tiny XOR network to show how changing one bias can change behavior, while noting that a digit-reading example contains 109,386 parameters.
ThorstenMeyerAI.com published an explainer on November 8 showing how a neural network can classify a handwritten seven by passing image pixels through layers of calculations. The article illustrates the mechanics with a small, hand-specified network, then estimates that a sample digit reader has 109,386 parameters—numbers that training adjusts to shape the model’s behavior.
For a 28-by-28 image, the input can be represented as 784 pixel values. In the multilayer perceptron described by the article, the input layer holds those values and passes them forward; it does not itself calculate a prediction. Hidden layers transform the values, and an output layer returns scores for the possible digits. The highest score is treated as the predicted digit.
Each neuron in a dense layer multiplies its inputs by weights, adds a bias, and applies an activation function. The article summarizes the operation as z = W x + b, followed by an activation. A layer’s outputs then become the next layer’s inputs. These steps are arithmetic, the article says, even when the number of operations makes a trained network difficult to interpret by inspection.
To show how a parameter can affect a network, the explainer constructs a small network for XOR, a function that returns one when two input bits differ. Its hand-chosen network uses nine numbers. Changing the output bias from −2 to −1 makes the example return one for all four possible inputs, rather than only when the bits differ. This is a teaching example, not a reported experiment on a digit-recognition system.
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 Individual Weights Matter
The example makes a practical distinction between a network’s architecture and its learned parameters. Architecture describes features such as layer count and width; weights and biases are the values that training normally adjusts. The article says about 92% of the digit reader’s 109,386 parameters are in its first layer, where each of the 784 input pixels connects to each of 128 units.
That scale helps explain why directly changing a trained model is not usually as simple as changing one bias in the XOR demonstration. In a larger network, behavior may depend on patterns distributed across many parameters and layers. Still, understanding the arithmetic gives readers a way to think about what a model computes—and why a change to its internal values can alter its output.
The source also points to a separate claim about AI systems editing other networks’ weights. It says MIT CSAIL researchers reported that 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, so those figures should be understood as the report’s account rather than independently established findings here.
Products Worth Considering
Anker SOLIX C1000 Gen 2 Portable Power Station, 1,024Wh
Samsung 49-Inch Odyssey G9 G91F Curved Dual QHD Gaming Monitor, 144Hz
Sony WH-1000XM4 Wireless Noise-Canceling Headphones, Silver
As an affiliate, we earn on qualifying purchases.
From Pixels to Layer Outputs
The explainer focuses on a multilayer perceptron, a feedforward network with dense layers. It uses a flattened handwritten image as its example: instead of passing a 28-by-28 grid through the network as a grid, the example represents it as a list of 784 pixel values. The model processes those values in order through intermediate layers before producing digit scores.
The XOR demonstration is included to make parameter effects visible with only two input bits and a few calculations. It does not show how a digit reader learns to recognize handwriting. In a trained model, weights and biases are generally set through training on data; the XOR values in the article are assigned directly to produce a chosen behavior. The article also says that early image-network layers can respond to simple patterns such as edges, while later layers may respond to object parts, but the specific digit-reader architecture and training details are not provided.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com
Limits of the Demonstration
The source does not identify the training dataset, evaluation method, or specific architecture behind its 109,386-parameter digit reader. It also does not report a measured recognition accuracy for handwritten sevens, so the explainer should not be read as a new benchmark or study of digit-recognition performance.
The account of the MIT CSAIL weight-editing result is not accompanied here by the underlying logs, model details, or a peer-reviewed publication. The reported score change and edit size are therefore claims attributed to the source’s description; their evaluation conditions and reproducibility are not established in the supplied material. The XOR example, meanwhile, establishes only what happens in that small, explicitly specified network.
What the Explainer Leaves Open
The article’s next step is conceptual rather than a reported research milestone: it moves from a network small enough to calculate by hand toward the much larger parameter count of a digit reader, then raises the question of how one AI might edit another model’s weights. The supplied material gives no announced follow-up experiment or publication date.
To evaluate the reported weight-editing result, readers would need the linked public logs and fuller details about the model, task, scoring procedure, and comparison conditions. For the digit example, an accuracy measure and information about the data and training setup would be needed to assess how reliably the network recognizes handwritten sevens.
Key Questions
How does a neural network identify a handwritten 7?
In the article’s example, a 28-by-28 image is represented as 784 pixel values. Layers transform those values using weights, biases, and activation functions; the output layer provides scores for the possible digits, and the highest score is the prediction.
What are weights and biases?
Weights control how much inputs contribute to a neuron’s total. A bias shifts that total before an activation function determines the neuron’s output. Training normally learns these parameter values.
Does changing one weight always change a model’s behavior?
No. The explainer shows a single-bias change that alters a deliberately small XOR network. In a trained model, a behavior may depend on patterns across many parameters, so the effect of an individual edit can vary.
Is the 109,386-parameter figure a reported benchmark?
No benchmark is provided in the source material. 109,386 is the parameter count given for the explainer’s sample digit reader; its accuracy and training setup are not specified.
Has the AI weight-editing result been peer reviewed?
The source describes the MIT CSAIL material as a blog post with public logs, not a peer-reviewed paper. The supplied account does not establish the result’s evaluation details or reproducibility.
Source: ThorstenMeyerAI.com
