Perceptron Lab

Rosenblatt's learning machine, 1958 — rebuilt so you can watch it think

All handwritten examples are invented and generated in this lab.

what was written the weights its ink lands on evidence piling up, cell by cell silent output unit no example loaded Each inked square looks up its own weight. Those weights are added one at a time; the line is the running total.
2/s
1
200
100%
space steps  ·  R runs
Examples seen
0
Mistakes made
0
Passes completed
0
Target digits caught
—
False alarms
—
Press step to load the first example.

What it has learned

The 49 weights, laid back onto the grid they came from. Nobody drew this.

ink here argues for ink here argues against

Adding up the evidence

Only cells with ink contribute. Everything else is skipped entirely.

The learning rule, with this example's numbers

This is the whole of it. There is no other mechanism anywhere in the machine.







  

    
Mistakes per pass through the set

    
When this reaches zero the machine stops changing, because nothing is left to correct.

    
  


  

    
Draw your own

    
Tap cells to draw. The machine scores it live using what it has learned so far.

    

      

      

        
Score

        
0

        
silent

        
        
      

    

  






  
The training set

  
Click any example to load it into the apparatus. Outlined thumbnails are the digit being detected.