TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A neural network turns pixel values into a digit prediction through layers of weighted sums, biases and activations. The source uses a small XOR example to show how one bias can change a network’s output, then counts 109,386 parameters in a digit reader.
A November 8 explainer from Thorsten Meyer traces how a neural network can turn a picture of a handwritten seven into a digit prediction, showing the arithmetic from 784 pixel values through layers of calculations to an output score. The post also uses a nine-parameter XOR network to demonstrate how changing one bias can change the function the network computes.
The digit example starts with a 28-by-28 image, flattened into a list of 784 pixel values. The input layer holds those values and passes them onward; it performs no calculation. In each dense layer, neurons multiply inputs by weights, add a bias, and apply an activation function. The resulting values become the next layer’s inputs.
The final layer in the example returns 10 scores, one for each digit. The digit with the highest score is the network’s prediction. The explainer says a small digit reader with 128 units in its first hidden layer has 109,386 parameters, about 92% of them in that first layer because each pixel connects to every unit there.
To make parameter changes easier to follow, the post builds an XOR network with two input bits, two hidden neurons and one output neuron. Its hidden neurons check whether at least one bit is on and whether at most one bit is on. The output combines those conditions to return one when the bits differ. The example lists nine numbers across the weights and biases.
Follow The Data How a Network Reads a Handwritten 7
Pixel values pass through layers of weighted sums, biases and activations to produce digit scores. The explainer shows the arithmetic and where a small parameter change can alter a network’s output.
From pixels to prediction
Each dense layer passes calculated values forward; the highest final score determines the predicted digit.
A 28-by-28 image becomes 784 pixel values.
Neurons multiply inputs by weights, add a bias and apply an activation.
The output layer returns 10 scores; the largest is the prediction.
What the parameters do
The layer equation is z = W x + b, followed by an activation. Weights control input contributions, while biases shift the calculation.
- Two input bits, two hidden neurons and one output neuron.
- Nine weights and biases combine to return one when the bits differ.
- Most parameters sit in the first hidden layer, where each pixel connects to every unit.
- The stated count makes manual changes difficult to track.
A bias changes the result
In the XOR demonstration, one parameter changes the function the network computes.
| Output bias | Network behavior |
|---|---|
| −2 | Returns one when the two bits differ. |
| −1 | Returns one for all four input combinations. |
Why Parameter Changes Matter
The examples make a technical point visible: a network’s behavior comes from many numerical parameters arranged in layers, and changing a parameter can alter what the network outputs. In the XOR demonstration, changing the output bias from −2 to −1 makes the network return one for all four possible inputs instead of distinguishing inputs whose bits differ.
The digit-reader count also shows why manually editing a larger model is difficult. Most of its stated parameters are concentrated in the first layer, and the full prediction depends on calculations carried through the network. The post links this idea to a separate report from MIT CSAIL researchers about an AI model editing other networks’ parameters. That connection is presented as an example of “weight-space programming,” not as evidence that all trained networks can be edited simply by changing one number.
Products Worth Considering
Apple iPad 11-inch with A16 Chip, 512GB, Wi-Fi 6, and Apple Pencil (USB-C) — Blue
Bose QuietComfort Ultra Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Black
Apple iPad Pro 12.9-inch (2017), 512GB Wi-Fi, Space Gray
As an affiliate, we earn on qualifying purchases.
From Pixels to Layer Outputs
A multilayer perceptron, or MLP, is a feedforward network made of dense layers. “Dense” means each unit in a layer receives input from every unit in the preceding layer. A layer can be written as z = W x + b, followed by an activation that produces the layer’s output. Here, x is the input vector, W holds weights, and b holds biases.
The explainer calls intermediate layers “hidden” because they sit between the provided input and the answer being read. Their values are still calculable: they are the intermediate results of the network’s arithmetic. In the handwritten-digit example, the process begins with pixel values and ends with a set of class scores. The source does not provide a specific image’s pixel values or the corresponding intermediate scores for a handwritten seven.
The post also refers to a separate MIT CSAIL report. It says an AI model read and edited other networks’ numbers, and that one edit changed 0.016% of a language model’s weights while raising 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.
“People like to call this a black box. It isn’t one in the sense that nothing can be seen: every step is arithmetic you could do on paper. The hard part is that there is a lot of it.”
— Thorsten Meyer, in the explainer
Limits of the Demonstration
The source does not name the specific digit-recognition model, provide its training data, or show the handwritten seven’s actual pixel values and layer-by-layer outputs. It gives the parameter count for a small digit reader, but does not specify the full architecture or how its 109,386 parameters were calculated.
The MIT CSAIL example is summarized rather than documented in detail. The source does not identify the language model, the math benchmark, the evaluation conditions, or whether the score change held across other tests. It characterizes the material as a blog post with public logs; the figures should be read as claims reported there, not as peer-reviewed findings.
Following the Calculations
The explainer’s next step is to scale the hand-worked example from nine numbers to the 109,386 parameters in the digit reader, showing how a network’s layers turn pixel inputs into scores. The source provides no release schedule or follow-up date for that material. Readers can use the stated layer equation and XOR example to trace how inputs, weights, biases and activations combine, while recognizing that the specific seven’s intermediate values are not included.
Key Questions
How does the network recognize a handwritten seven?
In the described example, a 28-by-28 image becomes 784 pixel values. Layers process those values through weighted sums, biases and activations, and the output layer assigns scores to the possible digits. The source does not give the scores for a particular seven.
What is a neural network weight?
A weight is a number that controls how much an input contributes to a neuron’s calculation. A neuron multiplies its inputs by weights, adds a bias, then applies an activation function.
What does changing the XOR network’s bias do?
Changing its output bias from −2 to −1 changes the example network’s output: it returns one for all four input combinations instead of returning one only when the two bits differ.
What is confirmed about the MIT CSAIL example?
The explainer reports that an AI model edited other networks’ parameters and cites a math score change from 7.9% to 48.1% after an edit to 0.016% of a language model’s weights. It describes the source as a blog post with public logs, not a peer-reviewed paper; the model and benchmark details are not supplied in this material.
Source: Thorsten Meyer AI
