101_nomenclature_and_acronyms.txt

[[{PM.draft.intention_to_be]]
# DATA SCIENCE NOMENCLATURE (State: Draft/Planned)
 

- SEGMENTATION: pre-processing where objects of interest are "extracted" from background.

- FEATURE EXTRACTION: Process that takes-in a pattern and produces feature values.
  Number of features is virtually always chosen to be fewer than the total
  necessary to describe the complete taret of interest, and this leads to a loss
  in information.

  In acts of associate-memory, the system takes in a pattern and emits another
pattern which is representative of a general group of patterns. It thus reduces
the information somewhat, but rarely to the extent that pattern classification
does. In short, because of the crucial role of a decision in pattern recognition
information, IT IS FUNDAMENTALLY AN INFORMATION REDUCTION PROCESS.
THE CONCEPTUAL BOUNDARY BETWEEN FEATURE-EXTRACTION AND CLASSIFICATION IS ARBITRARY.

- SUBSET AND SUPERSET PROBLEM: Formally part of mereology, the study of part/whole
  relationships. It appears as though the best classifiers try to incorporate
  as much of the input into the categorization as "makes sense" but not too much.

- RISK: Total spected cost of making a wrong classification/Decision
  ```
  | (total spected cost = wrong-classification cost x probability of event)
  ```

- SFT: Supervised fine-tunning. [[training]]

- DPO: Direct Preference Optimization [[training]]

- SOTA: State of the art.

- ReLu, GeLu Activation function: Rectified Linear unit is defined as the 
  non-negative part of its argument.
  It was argued that ReLU approximates the biological relationship between 
  neural firing rates and input current, in addition to enabling recurrent neural 
  network dynamics to stabilise under weaker criteria.<br/>
  Prior to 2010, most activation functions used were the logistic sigmoid 
  (which is inspired by probability theory; see logistic regression) and its more 
  numerically efficient[14] counterpart, the hyperbolic tangent. Around 2010, the 
  use of ReLU became common again. <br/>
  - Ge-Lu Activation function: Gaussina-error linear unit, is a variant of the 
    TeLu function fixing some technical problems. 
  (further details: <https://en.wikipedia.org/wiki/Rectified_linear_unit>)

- Negative Log-Likehook (NLL) and Perplexity (PPL) are two commonly used metrics
  when evaluating (during training) large language models. [[training.metric,LLM]]
  - NNL: loss function aimed at maximizing the probability of correct predictions.
    Its value is contrained within a positive interval. A high value indicates
    that the networks has not yet learned patterns in the training set.
  - PPL: Exponentiation of NLL. It provides a more intuitive way to understand
    the model's performance. High values indicate poor learning performance.

- Foundation model: Also known as base model, is a general pursose LLM trained
  on huge data. It compares to the fine-tuned model, a further trained model
  that start its training with an initial foundation model and continues to
  train it for an specific task [[LLM.101]]

- Measuring Massive Multitask Language Understanding (MMLU) is a popular 
  benchmark for evaluating the capabilities of large language models. It inspired 
  several other versions and spin-offs, such as MMLU-Pro, MMMLU and MMLU-Redux.  [[LLM]]

- Reinforcment Learning from Human Feedback (RLHF): Trains the model using feedback from
  humans.  RLHF implementation requires three main components:
  - Supervised Fine-Tuning (SFT) - Initial model training on demonstration data
  - Reward Model Training - Learning to predict human preferences
  - Reinforcement Learning - Optimizing policy using the reward mode

- Text-to-Image Models: Models trained to create realistic images  from text.
  - Midjourney:
  - DALL-E 2:
  - Stable Diffusion:
  - :..
  Popular models use a text encoder to map input text into an embedding (latent?)
  space.  This embedding is fed into a series of conditional diffusion models, 
  which denoise and refine a latent image in successive stages. The final output
  is a high-resolution image aligned with the input text. To main architectures exists:
  - Generative Adversarial Networks (GANs): highly realistic but computationally 
    expensive and unstable images.
  - Diffusion models: General (not only for text-to-image) prommising models. 
    They reduce the computational power and the sequential error accu8mulation.
  - Stable Diffusion: (By CompVis group at LMU Munich - High-Resolution Image 
    Synthesis with Latent Diffusion Models by Blattmann et al., 2022). [[doc_has.bibliography]]
    - The model runs on a modest personal GPU. 

- Stable Diffussion Latent Space: lower-dimensional space representation (of a higher dimensional
  N-pixels-wide x M-pixels-hight) 2D "image".
  - A varational autoencoder (VAE) provides latent space compression (called 
    perceptual compression in the paper) while a U-Net (Noise-predictor CNN) performs 
    iterative denoising.

- U-Net: "Popular" type of convolutional Neural Network (NNT) with a symmetric encoder-decoder
  structure. 
  - It is commonly used for image segmentation taks.
  - In Stable Diffusion, used to introduce/remove noise. It takes a noisy image (seed)
    as input and processes it through a series of convolutional layers to extract 
    features and learn semantic representations.

- prosody: The set of a given speaker's voice and the patterns of rhythm and sounds. [[cognitive,use_case.digital_twin,]]

- Model regularization: ML training technique used to prevent overfitting.<br/>
  Overfitting happens when a models learns too much from the training data, including
  noise and outliers, and performs pooly on unseen data.<br/>
  Regularization works by adding a penalty term to the loss function, which discourages
  the learning algorithm from assigning too much importance to any individual predictor
  (feature).<br/>
  There are different types of regularization techniques including:
  - L1 (Lasso regression)
  - L2 (Ridge regression)
  - L1 + L2 combinations (Elastic net)
  
  [[training]]

- Vapnik–Chervonenkis theory: developed during 1960–1990 by Vladimir Vapnik 
  and Alexey Chervonenkis. The theory is a form of computational learning theory, 
  which **attempts to explain the learning process from a statistical point of 
  view**. 
  * <https://en.wikipedia.org/wiki/Vapnik%E2%80%93Chervonenkis_theory>
- Vapnik–Chervonenkis dimension: In Vapnik–Chervonenkis theory, it is a measure of
  the size (capacity, complexity, expressive power, richness, or flexibility) of a 
  class of sets. 
    The notion can be extended to classes of binary functions. It is defined as the 
  cardinality of the largest set of points that the function class can 
  shatter—that is, for which all possible binary labelings can be realized by 
  some function in the class. It was originally defined by Vladimir Vapnik and 
  Alexey Chervonenkis.
  * <https://en.wikipedia.org/wiki/Vapnik%E2%80%93Chervonenkis_dimension>

- Word2vec: algorithm or technique for obtaining representations of words in 
  a vector space (of hundreds of dimensions) 
  * <https://en.wikipedia.org/wiki/Word2vec>

- LLM Conditioning vs aligning: while conditioning can include (training) fine-tuning
  and (inference) prompt tunning and is focused on influencing the model through various 
  techniques at different layers of interaction, (inference) alignment is 
  concerned with the fundamental and holistic calibration of the model’s 
  behavior to human ethics and safety standards.


- Exploratory Data Analysis (EDA)

[[{]]
## Parameter efficient fine-tuning (PEFT)
- Parameter-Efficient Fine-Tuning (PEFT) are a set of methods used to fine-tune Large 
  Language Models (LLMs) by updating a small subset of the model's parameter 
  while keeping the majority of the pre-trained weights frozen.<br/>
  The best well known example is LoRa.<br/>
  Libraries exists to fine-tune base models in Hugging Face 
  (<https://huggingface.co/docs/peft/tutorial/peft_model_config>), ...

[[}]]

[[{]]
- Low-Rank (LoRa) fine-tunning training (<https://arxiv.org/abs/2106.09685>)
  "full fine-tuning of base models becomes less feasible as the complexity 
  increases. (ex: deploying independent instances of fine-tuned models
  of GPT-3 175B is prohibitively expensive).<br/>
   LoRA freezes the pre-trained model weights and injects trainable rank
  decomposition matrices into each layer of the Transformer architecture, 
  greatly reducing the number of trainable parameters for downstream tasks.
   Compared to GPT-3 175B fine-tuned with Adam, LoRA can reduce the number
  of trainable parameters by 10,000 times and the GPU memory requirement by
  3 times. [[PM.price,performance.101,LLM,training.fine_tunning.LoRa]]
   LoRA performs on-par or better than fine-tuning in model quality on 
  RoBERTa, DeBERTa, GPT-2, and GPT-3, despite having fewer trainable parameters,
  a higher training throughput, and, unlike adapters, no additional inference
  latency. <br/>
  PyTorch implmentation: <https://github.com/microsoft/LoRA>
  * QLORA is an extension of LoRA, enabling efficient fine-tunning of large models
    by backpropagating gradinets through a frozen 4-bit quantized model into 
    learnable low-rank adapters, allowning to tune a 65B parameter model on a
    single GPU. It achieves 99% of ChatGPT performance on Vicuna.<br/>
    It also reduces memory requirements from >780GB to <48GB, without affecting
    runtime or predictive performance. [[performance.101,training.fine_tunning.QLORA]]
- Quantization: Technique to reduce numerical precision of weights and activations 
  in neural networks like LLMs. Ex: from Floating Point 32bytes to 8-bit integers.
  - LLMs like BERT and GPT-3 have been shown to work well with 4-8 bit quantization.
[[}]]

[[{]]
- GPT: Generative Pretrained transformer.
[[}]]

[[{]]
- Generative Adversarial Network (GAN): [[training.GAN]]
[[}]]
[[PM.draft.intention_to_be}]]

[[{]]
- Autoregressive Transformer Language Model (ATLM): Technical architecture behind Language Models from OpenAI GPT, Facebook Llama, ...
  - Autoregressive: The model generates text by predicting one token at a time,
    conditioning each prediction on the previously generated tokens.
  - Transformer: The underlying neural network architecture.
  - Language Model: A model trained to understand and generate human-like text.
  - Common Shorthands  include:
    - "Autoregressive LLM"
    - "Causal Language Model" (since it uses causal masking to predict the next token)
    - "Transformer-based LLM"
    - LLM: This is technically incorrect since there exists LLM not based on ATLM.
[[}]]