notes

Log | Files | Refs

concepts.txt (2025B)

      1 ==========================================
      2 ONNX
      3 ==========================================
      4 
      5 Machine learning data are often stored in `.onnx` format
      6 
      7 [ONNX Documentation](https://onnx.ai/onnx/index.html)
      8 
      9 
     10 ==========================================
     11 Image processing kernal
     12 ==========================================
     13 
     14 They are basically matrices
     15 
     16 [Image processing kernel](https://en.wikipedia.org/wiki/Kernel_(image_processing))
     17 
     18 
     19 ==========================================
     20 VAE (Variational autoencoder)
     21 ==========================================
     22 
     23 https://en.wikipedia.org/wiki/Variational_autoencoder
     24 
     25 Input -> Encoder -> Latent Space -> Decoder -> Output
     26 
     27 
     28 ==========================================
     29 Vector database
     30 ==========================================
     31 
     32 The vector set of 3D models is the same as the vector database
     33 
     34 https://en.wikipedia.org/wiki/Vector_database
     35 
     36 if you convert a word, mesh or other object into a vector, you can use classic
     37 trig to calculate how alike two objects are:
     38 
     39 such as Cosine Similarity
     40 
     41 Sim(A,B) = cos(θ) = A ⋅ B / ||A|| ||B||
     42 
     43 so when we convert meshes to vectors it enables us to classify different meshes.
     44 
     45 It's a sequence of simple steps pipelined into each other.
     46 
     47 
     48 ==========================================
     49 Loss function
     50 ==========================================
     51 
     52 AI is essentially taking small steps to make a small loss function:
     53 
     54 https://en.wikipedia.org/wiki/Loss_function
     55 
     56 The loss function can be built around things like wind drag, so the cost
     57 function gets larger the more drag, so the AI takes steps to alter the
     58 parameters to reduce the drag, but we can also make other loss functions for
     59 anything, like architecture, etc.
     60 
     61 
     62 ==========================================
     63 AB-UPT released by Emmi AI
     64 ==========================================
     65 
     66 https://www.emmi.ai/news/ab-upt-scaling-neural-surrogates-100m-cfd-meshes
     67 
     68 
     69 ==========================================
     70 Neural Operators
     71 ==========================================
     72 
     73 https://en.wikipedia.org/wiki/Neural_operators