Rico贾若童的博客

[ML] Soft Rounding - Differentiable Quantization for Neural Networks

Soft rounding is a smooth approximation to normal rounding. Normal rounding maps a continuous value to a discrete level: \[1.2 \rightarrow 1\] \[1.8 \rightarrow 2.\] This is useful for modeling ...

Posted by Rico's Nerd Cluster on August 8, 2026

[ML] Loss For Distribution Stats

loss for standard devations

Matching Real Point-Cloud Statistics with a Differentiable Distribution Loss Sometimes we want a synthetic point-cloud generator to produce range and intensity values that have the same statistica...

Posted by Rico's Nerd Cluster on August 2, 2026

[ML] Flow Matching

Optimal Transport, Concentration of Measures

1. Problem Setup and Definitions Imagine we now have a bunch of particles in the 1D world. They are all moving randomly. Along x axis, its current distribution can be characterized as a source dis...

Posted by Rico's Nerd Cluster on August 2, 2026

[ML] Straight Through Bernulli Output

1. Goal We want to use a CycleGAN to model a 3D imaging sonar that comes with random dropouts. The sonar has two channels: intensity and range. The dropout value is (0,0). As part of the CycleGAN ...

Posted by Rico's Nerd Cluster on July 31, 2026

[ML] Graph Engineering

Graph Engineering for AI Agents: Increasing Certainty by Bounding Autonomy As models become more capable, the hard problem is no longer simply getting an agent to do something. It is getting the a...

Posted by Rico's Nerd Cluster on July 28, 2026

[ML] Attention Pooling

Attention Pooling Pooling converts a variable number of vectors into one fixed-size representation. For N point features, \[\mathbf{x}_1,\mathbf{x}_2,\ldots,\mathbf{x}_N,\] max pooling independe...

Posted by Rico's Nerd Cluster on July 25, 2026

[ML] Continual Learning

LoRA, AdaBN

## 1. What continual learning actually means Suppose a detector is trained on Dataset 1. Later, Dataset 2 arrives. We want to train on Dataset 2 without keeping all of Dataset 1 in every training ...

Posted by Rico's Nerd Cluster on July 22, 2026

[Robotics] Two Paths from PyTorch to Fast GPU Inference

TensorRT, Torch Compile

torch.compile and TensorRT A PyTorch model can reach an NVIDIA GPU through several optimization paths. Two important choices are: compile PyTorch execution with torch.compile and TorchInductor...

Posted by Rico's Nerd Cluster on July 19, 2026

[ML] From GPU Kernel to Machine Code

Triton, CUDA, PTX, and SASS

Overview When you write a GPU operation, several different languages and compilation stages may sit between your source code and the instructions executed by the GPU. The short version is: 1 2 3...

Posted by Rico's Nerd Cluster on July 16, 2026

[Robotics] 3D-shape-detection-from-point-clouds

BEV, Model Training Experience

BEV BEV is full 7 dof representation: (x,y,z,l, w, h, yaw). Its scene is a heatmap on a top down grid, and where it localizes its peak. Here, (x,y,z) is a BEV pillar, which is the 2.5D representa...

Posted by Rico's Nerd Cluster on July 13, 2026

Rico's Nerd Cluster