Jai

Learning what the loss function can’t measure !

[Jai @jaicdev]

@ildivino_42 Congratulations 🥳

[Jai @jaicdev]

Why it matters on small hardware: 21× the tokens costs 201× the time in one layer, and detection and segmentation

inputs can easily be 1024 px or more.

Next, #2: VMamba’s four-direction scan, using its own code.

[Jai @jaicdev]

Efficient Vision #1: Attention’s quadratic tax

A 224 px image is 196 tokens for a ViT. A 1024 px image is 4,096.

One PyTorch attention layer on my notebook
GPU: 0.351 ms → 70.5 ms (201×).
A linear layer: 0.054 ms → 0.515 ms (10×).

Slope 1.9 vs 0.8.

[Jai @jaicdev]

Attention that returns its weights builds an N × N table: 1,536 MiB at 4,096 tokens (batch 8, 3 heads), 3,144 MiB peak.

The fused path needs 144 MiB.
But its time is still quadratic: 41.9 ms, slope 1.8.

Memory is fixable. Compute isn’t.

[Jai @jaicdev]

DLU #1: Big-O isn’t wall time

Multiplying 16 numbers on my GPU: 5.2 µs.
Multiplying 65,536 numbers: 5.5 µs.

4,096× more work, same time.

Big-O tells you how cost grows. It doesn’t tell you what you pay. Small ops are pure overhead.

[Jai @jaicdev]

@Mitchell_Suen Hey! Computer Vision systems enthusiast here. I like finding where models break and sharing how I do it. Let's connect! 👋

[Jai @jaicdev]

@harikesh_devops Happy to connect 🤝

[Jai @jaicdev]

Attention vs a selective scan (Mamba’s core op), same CPU, plain PyTorch:

• attention: slope 2.02, O(N²)
• scan: slope 1.00, O(N)
• 16,384 tokens: 855 ms vs 472 ms

But up to ~10,000 tokens attention is still faster. Linear isn’t automatically fast.

Code ↓

[Jai @jaicdev]

@NovaOrbiting Deploying and scaling them on edge devices.

[Jai @jaicdev]

@beingsupm I just share what I learn !!

[Jai @jaicdev]

@oomeshb Indeed. We are together in this journey !!

[Jai @jaicdev]

@xianbingwudi Hehe, marketing is the need of the hour. 🤣

[Jai @jaicdev]

Building my 𝕏 profile from scratch.

Sharing my journey across AI, Machine Learning, Deep Learning, Computer Vision, and ML Systems : ideas, experiments, practical insights, challenges, and lessons learned along the way.

Learning, building, and sharing in public.

If you're exploring similar areas or building something interesting, let's connect.

[Matheorems @matheorems]

The Five Differential Operators of Physics

Five operators built from one symbol, nabla, all moving at once.

Divergence: div F = dFx/dx + dFy/dy + dFz/dz, the net flux out of a tiny closed surface (source or sink strength).

Gradient: grad phi = (dphi/dx, dphi/dy, dphi/dz) points to the steepest increase of phi, always at a right angle to the contour lines, and is longest where they are tightest.

Curl: curl F measures the spin of the field around a point. The vector stands on the axis of the whirl, and flips when the whirl reverses.

Laplacian: the sum of the second derivatives. On the bowl z = x^2 + y^2 the Laplacian is 4, so the average around any circle of radius r sits exactly r^2 above its centre.

D'Alembertian: box F = (1/c^2) d^2F/dt^2 - Laplacian F. A wave running at speed c makes it zero, which is why it appears in the wave equation and in relativity.

[Jai @jaicdev]

@mriganka_94 Hey! CV systems enthusiast here. I like finding where models break and sharing how I do it. Let's connect! 👋

[Jai @jaicdev]

@Denche2005 Hey! CV systems enthusiast here. I like finding where models break and sharing how I do it. Let's connect! 👋

[Jai @jaicdev]

@anurudhsharmaa Hey! CV systems enthusiast here. I like finding where models break and sharing how I do it. Let's connect! 👋

[Jai @jaicdev]

@llllooe123 Hey! CV systems enthusiast here. I like finding where models break and sharing how I do it. Let's connect! 👋

[Jai @jaicdev]

@dark_coderz Hey! CV systems enthusiast here. I like finding where models break and sharing how I do it. Let's connect! 👋

[Jai @jaicdev]

@djl997 Hey! CV systems enthusiast here. I like finding where models break and sharing how I do it. Let's connect! 👋