🔍 Read the full analysis: What Happens Inside A Neural Network? Trace A Handwritten 7 To Its Answer on ThorstenMeyerAI.com
Get movie nights 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 feedforward neural network turns pixel values into digit scores, using a nine-number XOR network to show how weights and biases affect behavior. It also describes a reported MIT CSAIL experiment in which an AI edited another model’s weights; the account is a blog post with public logs, not a peer-reviewed paper.
ThorstenMeyerAI.com published an explainer on November 8 tracing how a neural network can turn the pixel values of a handwritten seven into a digit prediction. The article walks through the arithmetic in a small network, then changes one bias in a separate XOR example to show how a single parameter can alter the output; it also discusses a reported AI system that edited another model’s weights.
The explainer describes a multilayer perceptron, a feedforward network built from dense layers. A 28-by-28 image can be represented as 784 pixel values. The input layer holds and passes those values onward; it does not transform them. In hidden layers, neurons multiply inputs by weights, add a bias and apply an activation function. The resulting values become inputs to the next layer. An output layer can return ten scores, one for each digit, with the highest score used as the prediction.
To make those operations easier to inspect, the article constructs a network for XOR, a function that returns one when two input bits differ and zero when they match. Two hidden neurons test whether at least one bit is on and whether at most one bit is on. An output neuron combines those results. With step activations, the network produces the XOR answer for all four possible inputs. The full example uses nine weights and biases.
The article then changes just the output neuron’s bias, from −2 to −1. Under the same step rule, that neuron returns one when either hidden condition is met. At least one condition holds for each of the four input pairs, so the modified network returns one for every pair rather than computing XOR. The example demonstrates a specific network whose behavior can be tracked directly; it does not establish that one parameter generally controls a complex model’s behavior.
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
How a Single Bias Changes the Output
The walkthrough makes an abstract idea concrete: a neural network’s result comes from many connected arithmetic operations, not from a written rule saying what a seven looks like. Weights determine how strongly inputs contribute, biases shift when neurons respond, and activations transform the totals. Following those values helps readers understand what a model is doing at the level of computation, even when a large network is difficult to interpret as a whole.
The XOR example also shows why parameter changes matter. In the toy network, one bias changes the answer for every input. In a trained digit reader or language model, however, behavior can depend on patterns distributed across many parameters. The explainer cautions against treating the small example as a direct map of how large models work.
The article says a compact digit reader discussed in its account has 109,386 parameters, with about 92% in its first layer because each pixel connects to each of 128 units. That count illustrates how quickly the number of learned values grows even in a small image model. It also sets up the article’s discussion of weight-space editing: directly changing a model’s numbers rather than only prompting it with text.
Popular right now
Breville Barista Express Espresso Machine with Built-In Grinder, Brushed Stainless Steel
iRobot Roomba Plus 678 Robot Vacuum and Mop with AutoWash Dock
Dyson Supersonic Nural Hair Dryer
As an affiliate, we earn on qualifying purchases.
From Pixel Inputs to Model Weights
The article distinguishes architecture choices from learned values. Width refers to the number of units in a layer, and depth to the number of layers. These are hyperparameters selected when designing a network. Weights and biases are parameters that training normally learns, though they can also be assigned directly, as in the hand-built XOR example.
For the digit case, a 28-by-28 image is flattened into a list of 784 values. A dense layer applies a matrix of weights and a bias vector, then passes the activated results forward. The article gives the calculation as z = W x + b, followed by h = activation(z). Here, the input vector x is transformed by the weight matrix W and bias b; h is the layer’s output.
The source also refers to a report by MIT CSAIL researchers describing an AI model that read and edited other networks’ parameters. It says one edit altered 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. ThorstenMeyerAI.com characterizes the account as a blog post with public logs, not a peer-reviewed paper, so those figures should be understood as reported results rather than independent confirmation presented here.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com
Limits of the Reported Results
The explainer does not provide a full technical evaluation of the MIT CSAIL weight-editing report, and it explicitly says the account is not a peer-reviewed paper. The information supplied does not establish whether the reported score change was reproduced independently, how the math score was measured, or whether the edit affected other capabilities. Those details should not be inferred from the two percentages alone.
The handwritten-seven example is an illustration of network computation, not a report of a particular model’s measured accuracy. The source does not specify the digit reader’s complete architecture, training data, evaluation results or the precise image used in the video. It also does not establish that the 109,386-parameter count applies to every small digit-recognition network.
What the Explainer Leaves to Test
The source points readers to a video of about seven and a half minutes that follows the handwritten-digit example and the XOR bias change. It gives no scheduled follow-up, new experiment or publication date. Further confirmation of the reported weight-editing result would require access to the underlying methods, evaluation details and independent replication; the material provided does not say whether those steps are planned.
For readers following the mechanics, the immediate next step is to trace how each layer’s output becomes the next layer’s input, then compare a model’s scores with its predicted digit. For the larger claim about AI editing AI, the key open issue is whether the reported improvement holds under broader testing and can be reproduced beyond the public logs cited by the source.
Key Questions
How does a neural network recognize a handwritten 7?
A network can take the image as 784 pixel values, transform them through layers of weighted calculations and produce scores for possible digits. The digit with the highest score is the prediction. The source describes this process but does not supply a particular model’s accuracy.
What do weights and biases do?
Weights set how much inputs contribute to a neuron. A bias shifts the neuron’s total before its activation function is applied. Together with the activation, they determine the output passed to the next layer.
What does the XOR example show?
It shows that a small network can compute XOR and that changing one bias from −2 to −1 makes this example return one for all four input pairs. It demonstrates how a parameter affects this constructed network, not a general rule that one edit controls a large model.
Is the reported AI weight-editing result peer-reviewed?
According to ThorstenMeyerAI.com, the account is a blog post with public logs, not a peer-reviewed paper. The source reports a change in a math score but does not provide enough information here to establish independent replication or broader effects.
Source: ThorstenMeyerAI.com
