Denoised diffusion probabilistic model

DDPM claims to perform better than GANs in a very recent article. The idea of DDPM is that an image can progressively be added with noise resulting in a white noise image. And the neural networks can be trained to perform the reverse process, converting a white noise image to something looks natural.

Hindsight experience replay

A motivating example of hindsight experience replay is a bit flipping problem. Say start with a binary number as an initial state, an action can be flipping any arbitrary bit, and the goal is to reach a particular state. It is, of course, a toy problem that can be solved easily. However, say if we…

Experience on Mynt Eye

Tested Mynt Eye on my Thinkpad because it seems that it only supported up to 18.04. The sample example can run off the box as described here. However, it seems to need quite some work to get it run with OpenCV. Moreover, the depth estimation is not accurate at all. Not sure if further calibration…

Robust PCA

Came across of this video explaining robust PCA. It came with a book and it looks okay. As PCA can be considered as decomposition of data matrix with the highest singular value. So what PCA is doing is simply low-rank matrix approximation of the data matrix. So robust PCA is a simple idea that tries…

dm_control

Try to play with dm_control and request a trial license for mujoco. Since the trial license is for local machine, I have to set colab to run locally. Things mostly work okay. I still have some problems here and there (like ffmpeg can’t generate video). I have to install pyopengl and add the following to…

Machine teaching and inverse reinforcement learning

The two concepts machine teaching and inverse RL are closely related. But one is not a special case of another. For machine teaching, the goal is to decide the optimum data set for an algorithm to learn. In other words, given a constraint of teaching cost (that can be proportional to the size of training…

Domain randomization

Machine learning system fails when the training data distribution is significantly different from testing data. For a “simple” problem like image classification, we can avoid this problem by including sufficient diversity for the training data. But for more complicated real-world problems, such as robotic AI, there are simply too many different possibilities that one cannot…

counterfactual explanation

Counterfactual explanations are proposed and studied in recent years. In logic, counterfactual refers to the scenario when the condition of an if-statement is universally false. Note that the if-statement is universally true when the condition is universally false. So the conclusion is false even though the if-statement holds true always. Counterfactual example in ML refers…

Distributed representation

It sounds like a misnomer to me. I probably will just call it a “vector” representation. It doesn’t have the “distributed” meaning of scattering information into different places. For example, to recognize a cat with “distributed” representation, we may distribute features into like “does it has a tail?”, “does it have four legs?”, and “does…