Large Language Model (LLM) Talk

LLaMA-2 is a collection of large language models (LLMs), with pretrained and fine-tuned versions ranging from 7 billion to 70 billion parameters. The fine-tuned models, called Llama 2-Chat, are designed for dialogue and outperform open-source models on various benchmarks. The models were trained on 2 trillion tokens of publicly available data, and were optimized for both helpfulness and safety using techniques such as supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF). Llama 2 also includes a novel technique, Ghost Attention (GAtt), to maintain dialogue flow.

LLaMA-1 is a collection of large language models ranging from 7B to 65B parameters, trained on publicly available datasets. LLaMA models achieve competitive performance compared to other LLMs like GPT-3, Chinchilla, and PaLM, with the 13B model outperforming GPT-3 on most benchmarks, despite being much smaller, and the 65B model being competitive with the best large language models. The document also discusses the training approach, architecture, optimization, and evaluations of LLaMA on common sense reasoning, question answering, reading comprehension, mathematical reasoning, code generation, and massive multitask language understanding, as well as its biases and toxicity. The models are intended to democratize access and study of LLMs with some models being able to run on a single GPU, and to be a basis for further research.

The surveys of large language models (LLMs), covering their development, training, and applications. Key areas include data collection and preprocessing, which is crucial for model quality, and methods for adapting LLMs using instruction tuning or reinforcement learning with human feedback. The survey also discusses prompt engineering, which is important for task performance and involves designing clear instructions for the models. Additionally, the survey examines techniques like in-context learning and chain-of-thought prompting, and it addresses evaluation of LLMs in terms of factual accuracy and helpfulness. Finally, advanced topics such as long context modeling and retrieval-augmented generation are explored, along with techniques for improving efficiency.

Mixture of Experts (MoE) models use multiple sub-models, or experts, to handle different parts of the input space, orchestrated by a router or gating mechanism. MoEs are trained by dividing data, specializing experts, and using a router to direct inputs. Not all parameters are activated for each input, using sparse activation, and techniques such as load balancing and expert capacity are used to improve training. MoE models can be built through upcycling or sparse splitting. While MoEs offer faster pretraining and inference, they also present training challenges such as imbalanced routing and high resource requirements, which can be mitigated using techniques such as regularization and specialized algorithms.

Multi-task learning (MTL) is a machine learning approach where a model learns multiple tasks simultaneously, leveraging the shared information between related tasks to improve generalization. MTL can be motivated by human learning and is considered a form of inductive transfer. Two common methods for MTL in deep learning are hard and soft parameter sharing. Hard parameter sharing involves sharing hidden layers across tasks, while soft parameter sharing utilizes separate models for each task with regularized parameters. MTL works through mechanisms like implicit data augmentation, attention focusing, eavesdropping, representation bias, and regularization. In addition, auxiliary tasks can help improve the performance of the main task in MTL.

Gradient descent is a widely used optimization algorithm in machine learning and deep learning that iteratively adjusts model parameters to minimize a cost function. It operates by moving parameters in the opposite direction of the gradient. There are three main variants: batch gradient descent, which uses the whole training set; stochastic gradient descent (SGD), which uses individual training examples; and mini-batch gradient descent, which uses subsets of the training data. Challenges include choosing the learning rate and avoiding local minima or saddle points. Optimization algorithms like Momentum, Nesterov accelerated gradient, Adagrad, Adadelta, RMSprop, Adam, AdaMax, and Nadam address these issues. Additional techniques such as shuffling, curriculum learning, batch normalization, early stopping, and gradient noise can improve performance.

Generative Pre-trained Transformers (GPTs) are a family of large language models that use a transformer deep learning architecture. They are pre-trained on vast amounts of text data and then fine-tuned for specific tasks. GPT models can generate human-like text, translate languages, summarize content, analyze data, and write code. These models utilize self-attention mechanisms to process input and predict the most likely output, with a focus on long-range dependencies. GPT models have accelerated generative AI development and are used in various applications, including chatbots and content creation.

Linear Transformers address the computational limitations of standard Transformer models, which have a quadratic complexity, O(n^2), with respect to input sequence length. Linear Transformers aim for linear complexity, O(n), making them suitable for longer sequences. They achieve this through methods such as low-rank approximations, local attention, or kernelized attention. Examples include Linformer (low-rank matrices), Longformer (sliding window attention), and Performer (kernelized attention). Efficient attention, a type of linear attention, interprets keys as template attention maps and aggregates values into global context vectors, thus differing from dot-product attention which synthesizes pixel-wise attention maps. This approach allows more efficient resource usage in domains with large inputs or tight constraints.

BERT (Bidirectional Encoder Representations from Transformers) is a groundbreaking NLP model from Google that learns deep, bidirectional text representations using a transformer architecture. This allows for a richer contextual understanding than previous models that only processed text unidirectionally. BERT is pre-trained using a masked language model and a next sentence prediction task on large amounts of unlabeled text. The pre-trained model can be fine-tuned for various tasks such as question answering, language inference, and text classification. It has achieved state-of-the-art results on many NLP tasks.

Sora is an AI model from OpenAI that creates videos from text using a diffusion process, starting with noise and refining it. It employs a transformer architecture and handles videos as spacetime patches. Sora can extend existing footage, animate images, and blend videos. It has shown an ability to simulate elements of the real world, but has some shortcomings in depicting accurate physics and cause-and-effect relationships. The model is trained on large datasets of captioned videos and uses a "re-captioning" technique to enrich training data. Sora is not yet available to the public.