← Manish Krishan Lal

Literature on Projection based learning

Gradient descent on a loss Six nested ellipses are level sets of a quadratic loss. A path of gradient steps starts at theta zero, zigzags across the narrow valley and settles at the minimizer theta star. θ0 θ*

(a) Minimization by gradient descent θk+1 = θk − η∇L(θk)

Alternating projections between two sets Two bands, C one and C two, cross in a thin rhombus, their intersection. A path starts at z zero, outside both bands, projects onto C one, then onto C two, and keeps alternating. Each step is shorter than the one before as the path funnels into the sharp corner of the rhombus, where it stops at z star, a point in both bands. C1 C2 z0 z*

(b) Feasibility by alternating projections zk+1 = P2(P1(zk))

Training as minimization and as feasibility. Gradient descent needs a loss and its gradient. Alternating projections need only P1 and P2, the nearest-point maps onto C1 and C2, and stop at z*, a point of C1 ∩ C2. Both paths are computed, not sketched: (a) with automatic differentiation on a quadratic loss, (b) in projection softwares.
The dominance of gradient-based training risks paradigm lock-in, leading the community to overlook alternatives that might prove useful given comparable research investment. This page documents contributions to the field of projection-based learning. We aim to keep this resource simple and focused.

Software extensions of LWL framework (AI assisted)

  1. Projection-based learning, What is it, and why does it matter? (2024).
  2. A projection-based framework for gradient-free and parallel learning (2025). ◦ Empirics on Shakespeare dataset.
  3. Narrowing the Gap Between Projection and Gradient–Based Learning (2026). ◦ Empirics on CIFAR 10 with cyclic projections.

◦ Unique contribution