README.md

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

# threshold-mux2

2:1 multiplexer. The fundamental building block for data selection, routing one of two inputs to the output based on a select signal.

## Circuit

```
    d0      d1
    │       │
    │       │       s (select)
    │       │       │
    ├───────┼───────┤
    │       │       │
    ▼       │       │
┌───────┐   │       │
│ sel0  │◄──┼───────┤
│d0∧¬s  │   │       │
└───────┘   │       │
    │       ▼       │
    │   ┌───────┐   │
    │   │ sel1  │◄──┘
    │   │ d1∧s  │
    │   └───────┘
    │       │
    └───┬───┘
        ▼
    ┌───────┐
    │  OR   │
    └───────┘
        │
        ▼
      output
```

## Truth Table

| s | d0 | d1 | out | Description |
|---|----|----|-----|-------------|
| 0 | 0 | 0 | 0 | Select d0 (0) |
| 0 | 0 | 1 | 0 | Select d0 (0) |
| 0 | 1 | 0 | 1 | Select d0 (1) |
| 0 | 1 | 1 | 1 | Select d0 (1) |
| 1 | 0 | 0 | 0 | Select d1 (0) |
| 1 | 0 | 1 | 1 | Select d1 (1) |
| 1 | 1 | 0 | 0 | Select d1 (0) |
| 1 | 1 | 1 | 1 | Select d1 (1) |

Formula: `out = (d0 AND NOT s) OR (d1 AND s)`

## Mechanism

**Layer 1 - Selection Gates:**

The circuit uses two AND-with-complement gates that fire only when their respective data input is selected:

| Gate | Weights [d0, d1, s] | Bias | Function |
|------|---------------------|------|----------|
| sel0 | [1, 0, -1] | -1 | d0 AND NOT(s) |
| sel1 | [0, 1, 1] | -2 | d1 AND s |

**sel0** fires when d0=1 AND s=0:
- The negative weight on s means s=1 inhibits firing
- Bias -1 requires d0=1 to reach threshold when s=0

**sel1** fires when d1=1 AND s=1:
- Both d1 and s must be 1 to overcome bias -2

**Layer 2 - Output:**

Simple OR gate combines the two selection paths.

## Architecture

| Component | Weights | Bias | Function |
|-----------|---------|------|----------|
| sel0 | [1, 0, -1] | -1 | d0 AND NOT(s) |
| sel1 | [0, 1, 1] | -2 | d1 AND s |
| or | [1, 1] | -1 | OR(sel0, sel1) |

## Parameters

| | |
|---|---|
| Inputs | 3 (d0, d1, s) |
| Outputs | 1 |
| Neurons | 3 |
| Layers | 2 |
| Parameters | 11 |
| Magnitude | 10 |

## Building Larger Multiplexers

The MUX2 is the primitive for constructing larger multiplexers:

```
MUX4  = MUX2(MUX2(d0,d1,s0), MUX2(d2,d3,s0), s1)
MUX8  = MUX2(MUX4(d0-d3), MUX4(d4-d7), s2)
MUX16 = MUX2(MUX8(d0-d7), MUX8(d8-d15), s3)
```

Each doubling adds one layer of MUX2 gates.

## Usage

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

w = load_file('model.safetensors')

def mux2(d0, d1, s):
    inp = torch.tensor([float(d0), float(d1), float(s)])

    # Layer 1: Selection
    sel0 = int((inp @ w['sel0.weight'].T + w['sel0.bias'] >= 0).item())
    sel1 = int((inp @ w['sel1.weight'].T + w['sel1.bias'] >= 0).item())

    # Layer 2: Combine
    l1 = torch.tensor([float(sel0), float(sel1)])
    return int((l1 @ w['or.weight'].T + w['or.bias'] >= 0).item())

# Examples
print(mux2(1, 0, 0))  # 1 (selects d0)
print(mux2(1, 0, 1))  # 0 (selects d1)
print(mux2(0, 1, 1))  # 1 (selects d1)
```

## Applications

- Data routing in ALUs
- Register file read ports
- Conditional assignment in hardware
- Building blocks for barrel shifters
- Crossbar switches

## Files

```
threshold-mux2/
├── model.safetensors
├── model.py
├── create_safetensors.py
├── config.json
└── README.md
```

## License

MIT