torch-mice 0.4.0

Mixture of Convex Experts (MiCE) for PyTorch

MiCE (Mixture of Convex Experts)

What is MiCE?

MiCE is a lightweight PyTorch library for building convex mixture-of-experts models via a novel max-of-means atlas. Instead of softmax routing or hard top-k gating, MiCE fuses an ensemble of convex “petal” subnetworks with a fixed, irreducible atlas of affine charts:

1. Forward projection into each petal’s chart
2. Convex evaluation in chart space (Input-Convex Neural Nets)
3. (Optional) exact inversion back to the global frame
4. Overlapping mean-of-pairs + max fusion

This guarantees convexity, interpretability, and efficient compute without exponentials or discrete dispatch.

Why MiCE?

• Global convexity
  Each chart is convex; max-fusion preserves convexity in any dimension.

• Two operating modes

  • Forward-only (invert=False): fast chart tiling with learnable shifts
  • Atlas mode (invert=True): full chart atlas with exact reprojection, no shifts needed

• Efficiency
  No softmax, no log-sum-exp, no sparse dispatch. Fusion is just a handful of GEMMs, adds, and a single max per group. Compute scales ~2.6× a 2-layer MLP.

• Interpretability
  Clear regions of dominance — visualize arg-max and margins over petals in any 2-D slice.

---

Feature Comparison

Feature	MiCE (MoMx)	Softmax MoE	Hard MoE	Standard MLP
Routing	max(mean(…))	softmax(weights)	top-k mask	none
Convexity	✅ (vector-valued)	✅ (scalar only)	❌	❌
Atlas inversion	optional (invert)	—	—	—
Compute cost	~2.6× MLP	>10× (exp/log)	~k× experts	baseline
Params	~2.6× MLP	high	high	baseline
Gradient smoothness	high (piecewise convex)	smooth	sparse	smooth
Interpretability	high	medium	low	low

---

Installation

pip install torch-mice

import torch
from torch_mice import VectorHull

# Forward-only mode (default):
hull = VectorHull(in_dim=512, petals=8, out_dim=512, invert=False)
y_fwd = hull(x)

# Full atlas mode with exact inversion:
hull_atlas = VectorHull(in_dim=512, petals=8, out_dim=512, invert=True)
y_atlas = hull_atlas(x)

License

Licensed under the Gratis Public License © 2025 Joshuah Rainstar