Perceptron — Handwritten Digit Classifier

Deepak Nalla · Intro to AI

A perceptron trained from scratch in Python to recognize the digit 8 in ASCII art images — with zero external ML libraries.

---

Overview

Features	600 (30 columns × 20 rows per image)
Training samples	86 positive (digit 8) · 129 negative (others)
Epochs	15
Learning rate	0.10
Bias	0.10
Convergence	0 misclassifications by epoch 11

---

How It Works

A perceptron is the simplest possible neural network — one neuron that computes a weighted sum of its inputs and fires if the result crosses a threshold.

output = +1  if  Σ(w · x) + bias > 0   →  "It is a 8"
output = −1  otherwise                  →  "It is NOT a 8"

The learning rule updates weights after every misclassification:

w  ←  w + η × x × error

where η = 0.10 (learning rate), x is the pixel value (0 or 1), and error = label − prediction.

Training loop (per epoch):

1. Forward pass — compute weighted sum for each image
2. Compute error — label − output (0 if correct, ±2 if wrong)
3. Update weights — nudge each weight by η × x × error
4. Repeat for all 215 training images

---

Architecture — Three-Script Pipeline

data/               →   generate_features.py   →   features.csv
(215 ASCII images)      (30×20 grid → 600-dim       (215 rows × 601 cols
                         binary vector)               + LABEL column)

features.csv        →   perceptron_train.py    →   weights.csv
                        (15 epochs, lr=0.10)        (601 learned weights)

weights.csv         →   evaluate.py            →   "It is a 8"
+ test image                                        or "It is NOT a 8"

Feature extraction: Each ASCII image is read as a 20×30 grid. A # character becomes 1, everything else becomes 0. The grid is flattened into a 600-element binary vector — one vector per image.

---

File Structure

perceptron/
├── generate_features.py   # feature extraction from ASCII images
├── perceptron_train.py    # perceptron training loop
├── evaluate.py            # classify a new image
├── features.csv           # generated feature matrix
├── weights.csv            # learned weights (output of training)
├── requirements.txt
└── data/
    ├── train8/            # 86 positive training samples (digit 8)
    ├── trainOthers/       # 129 negative training samples
    └── test/              # evaluation images

---

Training Results

Misclassifications per epoch on the full training set (215 images):

Epoch	Misclassifications
0	89
1	44
2	29
3	25
4	17
5	15
6	18
7	11
8	7
9	6
10	2
11	0 ✓
12–14	0

---

Run It Yourself

Requirements: Python 3.10+

# Install dependencies
pip install -r requirements.txt

# 1. Generate feature matrix from training images
python3 generate_features.py

# 2. Train the perceptron (saves weights.csv)
python3 perceptron_train.py

# 3. Classify a test image
cp data/test/8.txt data/test/img.txt
python3 evaluate.py
# → It is a 8

---

Parameters

Parameter	Value	Notes
num_features	600	30 cols × 20 rows
LEARNING_RATE	0.10	step size per weight update
MAX_ITERATIONS	15	epochs
BIAS	0.10	constant offset to weighted sum