🔍 Read the full analysis: From Image To Answer: How AI Recognizes A Handwritten 7 on ThorstenMeyerAI.com
Get smart everyday buys 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 neural network can classify a handwritten seven, from pixel values through layers of arithmetic to a digit prediction. It also describes a separate MIT CSAIL report about an AI editing model weights; the cited source says that work appeared in 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 pixels of a handwritten seven into a digit prediction. The article follows numbers through a small, dense network and says a sample digit reader has 109,386 parameters, illustrating that the answer comes from layered calculations rather than a hand-written rule for the shape of a seven.
In the example, a 28-by-28 image is represented as 784 pixel values. The input layer holds those values and passes them forward; it does not itself calculate a prediction. In each dense layer, neurons multiply inputs by learned weights, add a bias, and apply an activation function. The resulting values feed the next layer. At the end, the digit model produces 10 scores, one for each possible digit, and the highest score determines its prediction.
The explainer makes the arithmetic concrete with a separate, hand-built network for XOR, a function that returns one when two input bits differ. Its network uses two hidden neurons and one output neuron, with a step activation. The article gives the network’s weights and biases and works through an input of (1, 0), which produces an output of one. Changing one output bias from −2 to −1 makes the network return one for every possible input. That example shows how a parameter can affect behavior in a deliberately simple system; it does not establish that a typical trained image classifier can be understood through a single edit.
For the digit-reader example, the source says the network has 109,386 parameters, with about 92% in its first layer. That share follows from the described dense layout: each of the 784 input values connects to each of 128 units in that layer. The source also points to a separate report involving MIT CSAIL researchers, who it says had an AI model read and edit another network’s numbers. The reported intervention 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.
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 walkthrough helps explain what a digit-recognition answer represents: a sequence of calculations across layers, not a rule that explicitly says what a seven looks like. That distinction matters when readers encounter claims about AI “understanding” images. The network processes numerical input and returns scores; the explainer describes how those operations work, but does not claim that its example reveals every internal feature of a production system.
The parameter count also gives a sense of scale. A small digit reader can contain more than 100,000 learned numbers, and a behavior may depend on interactions among many of them. The XOR example is useful precisely because it is simple enough to inspect fully. The separate model-editing report points to a possible research direction—changing weights to affect behavior—but its stated result should be read as a reported experiment, not as settled evidence that large networks are generally easy to control.
Products Worth Considering
FLEXISPOT Adjustable Workbench with Hand Crank, 47.2 x 23.6 Rubberwood Top
FLEXISPOT ErgoX Ergonomic Office Chair with Footrest, Black
HP OfficeJet Pro 8139 Wireless All-in-One Color Inkjet Printer
As an affiliate, we earn on qualifying purchases.
A Small Network Before the Digit Model
The article uses a multilayer perceptron, or MLP, as its teaching example. An MLP is a feedforward network built from dense layers: information moves from input to hidden layers and then to an output. In this setup, width means the number of units in a layer and depth means the number of layers. The explainer distinguishes those architectural choices from weights and biases, which are parameters ordinarily learned during training.
Its XOR demonstration is a deliberately small case, not the same model as the handwritten-digit reader. The network has only nine numbers to specify, and its behavior can be followed for all four possible two-bit inputs. The digit model has a much larger parameter set. The article uses the contrast to show why direct reasoning about one edited number can work in a toy network while a real model’s behavior may be distributed across many parameters and layers.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
What the Weight-Editing Report Establishes
The source does not provide enough detail here to independently evaluate the reported language-model experiment, including the model, the math test, the evaluation setup, or whether the score change persisted across other tests. It says the account was published as a blog post with public logs and was not a peer-reviewed paper. The figures should therefore be attributed to that report rather than treated as a broadly established result.
The explainer also does not identify the exact trained digit model or report its accuracy on a test set. Its 109,386-parameter count describes the sample architecture in the article; it is not evidence that every handwritten-digit classifier has that size. More generally, the XOR demonstration shows what a single bias change does in one constructed network, not how a comparable change would affect an image classifier.
The Next Test Is Reproducibility
The article’s video is presented as a roughly seven-and-a-half-minute hand-drawn explainer, moving from a handwritten seven to the XOR example and the weight-editing claim. Readers seeking to check the separate research result would need to examine the linked public logs and the underlying evaluation details. The source does not give a further publication date or say that a peer-reviewed paper is forthcoming.
For the digit-recognition example, the immediate takeaway remains the computation described: pixels enter as numbers, layers transform them, and the output scores support a prediction. Whether the reported weight-editing approach generalizes beyond its cited experiment remains an open question in the source material.
Key Questions
How does a neural network recognize a handwritten 7?
It receives the image as pixel values, transforms those values through layers of weighted calculations, and produces scores for possible digits. The highest score is the model’s prediction.
Does the input layer identify the digit?
No. In the described MLP, the input layer holds and passes on the 784 values from a 28-by-28 image. Later layers perform the transformations used to produce the output.
What does changing one bias show?
In the article’s small XOR example, changing the output bias from −2 to −1 changes the network’s output for all four inputs. It demonstrates the effect of a parameter in that constructed network, not a general rule for editing image classifiers.
Is the reported AI weight-editing result peer-reviewed?
The source says the MIT CSAIL-related account appeared in a blog post with public logs, not a peer-reviewed paper. It reports a math score change, but the supplied material does not include enough evaluation details to assess how broadly the result applies.
Source: ThorstenMeyerAI.com
