README.md

---
license: mit
tags:
- pytorch
- safetensors
- threshold-logic
- neuromorphic
- decoder
---

# threshold-3to8decoder

3-to-8 binary decoder. Converts 3-bit binary input to one-hot 8-bit output. A single-layer threshold circuit.

## Circuit

```
       a₂      a₁      a₀
        │       │       │
        ├───────┼───────┤
        │       │       │
    ┌───┴───┬───┴───┬───┴───┐
    │       │       │       │
    ▼       ▼       ▼       ▼
┌──────┐┌──────┐┌──────┐┌──────┐
│  y₀  ││  y₁  ││  y₂  ││ ...  │
│-1,-1,-1││-1,-1,+1││-1,+1,-1││      │
│b:  0 ││b: -1 ││b: -1 ││      │
└──────┘└──────┘└──────┘└──────┘
    │       │       │       │
    ▼       ▼       ▼       ▼
   y₀      y₁      y₂  ... y₇
```

## One-Hot Encoding

Each input value activates exactly one output:

| Input | a₂a₁a₀ | Output y₀y₁y₂y₃y₄y₅y₆y₇ |
|-------|--------|--------------------------|
| 0 | 000 | 10000000 |
| 1 | 001 | 01000000 |
| 2 | 010 | 00100000 |
| 3 | 011 | 00010000 |
| 4 | 100 | 00001000 |
| 5 | 101 | 00000100 |
| 6 | 110 | 00000010 |
| 7 | 111 | 00000001 |

## Mechanism

Each output yᵢ acts as a "pattern matcher" for input = i:

- **Weight +1** for bit positions that should be 1
- **Weight -1** for bit positions that should be 0
- **Bias** = -(number of 1 bits in i)

Example for y₅ (binary 101):
```
weights: [+1, -1, +1]  (match 1, reject 0, match 1)
bias: -2              (need 2 matches to fire)
```

When input = 101: sum = 1·1 + (-1)·0 + 1·1 = 2, fires
When input = 111: sum = 1·1 + (-1)·1 + 1·1 = 1, doesn't fire

## The Matching Principle

The circuit computes "how well does input match pattern i?"

- Perfect match: score = (number of 1s in i)
- One bit wrong: score = (number of 1s in i) - 1

The bias ensures only perfect matches pass.

## Weight Patterns

| Output | Binary | Weights | Bias |
|--------|--------|---------|------|
| y₀ | 000 | [-1, -1, -1] | 0 |
| y₁ | 001 | [-1, -1, +1] | -1 |
| y₂ | 010 | [-1, +1, -1] | -1 |
| y₃ | 011 | [-1, +1, +1] | -2 |
| y₄ | 100 | [+1, -1, -1] | -1 |
| y₅ | 101 | [+1, -1, +1] | -2 |
| y₆ | 110 | [+1, +1, -1] | -2 |
| y₇ | 111 | [+1, +1, +1] | -3 |

## Single-Layer Elegance

Unlike traditional logic (which uses AND, OR, NOT combinations), threshold logic can decode in one layer. Each output neuron directly computes "does input match me?"

## Architecture

**8 neurons, 32 parameters, 1 layer**

All neurons run in parallel - no dependencies.

## Usage

```python
from safetensors.torch import load_file
import torch

w = load_file('model.safetensors')

def decode(a2, a1, a0):
    inp = torch.tensor([float(a2), float(a1), float(a0)])
    return [int((inp * w[f'y{i}.weight']).sum() + w[f'y{i}.bias'] >= 0)
            for i in range(8)]

# Input 5 -> output 5 is hot
outputs = decode(1, 0, 1)
print(outputs)  # [0, 0, 0, 0, 0, 1, 0, 0]
```

## Files

```
threshold-3to8decoder/
├── model.safetensors
├── model.py
├── config.json
└── README.md
```

## License

MIT