Transformer Models and RNN Limitations

A type of neural network that processes sequential data, like text. It has a hidden state that keeps track of the context, but can struggle with long sequences due to vanishing gradients.

Specialized RNN that uses gates to control the flow of information, addressing the vanishing gradient problem and improving long-term dependency.

A sequence-to-sequence model that allows for inputs and outputs of different lengths. It consists of an encoder and a decoder, enabling complex tasks like machine translation.

A powerful model that breaks the reliance on RNNs for sequence-to-sequence tasks. It encodes the entire input and uses attention to focus on relevant parts for decoding. It can handle longer sequences.

The process of converting an entire input sequence into a compact representation that captures the essence of the input. This representation is then used for further processing.

A type of unsupervised learning where a model learns to predict the next token in a sequence based on previous tokens.

A technique used to compress the vocabulary of a language model by merging common character pairs into new tokens. This reduces the vocabulary size and allows the model to learn more efficiently.

The number of parameters in a language model, representing the model's complexity and ability to learn patterns.

The amount of data used to train a language model. Larger datasets generally lead to better performance.

The computational resources used to train a language model. These resources include compute power, memory, and time.

A type of language model architecture where the model only processes the output sequence, not the input sequence. GPT models are decoder-only.

A technique for adapting a language model to a specific task without updating the model's weights. The task is described in the prompt.

Updating the weights of a language model on a task-specific dataset to improve performance on that task.

A metric that measures how uncertain a language model is about predicting the next word in a sequence. A lower perplexity indicates the model is more confident in its predictions, while a higher perplexity suggests the model is less certain.

A metric used to evaluate the quality of a generated sequence against a reference sequence. It measures the precision of matching words between the generated and reference sequences. A higher BLEU score indicates a better match between the output and the reference.

A type of language model that is trained on a massive amount of text data and can generate realistic and coherent text. LLMs are often used for tasks like text summarization, translation, and chatbot development.

A type of statistical measure that penalizes the presence of low values more than the arithmetic mean. This makes it more sensitive to outliers and can be used to evaluate the performance of language models.

A family of large language models developed by Google, known for their impressive performance on various language tasks. GPT-3, GPT-4, and GPT-Neo are examples of models in this family.

A natural language processing (NLP) model that uses a bidirectional encoder to represent language, trained on two self-supervised tasks: masked language modeling (MLM) and next sentence prediction. It can then be fine-tuned for specific downstream tasks like sentiment analysis, text classification, and question answering.

BERT's task of predicting masked tokens in a sentence. It is a self-supervised learning method where the model learns to predict the missing words in a sentence based on the context of the surrounding words.

BERT's task of predicting whether two sentences are related (e.g., next sentence in a document). It helps BERT understand the relationship between sentences, making it useful for tasks like question answering and summarization.

The ability of a language model to understand the meaning of a word based on its surrounding words. This is crucial for NLP tasks that require understanding the context of language.

A type of language model that uses only a decoder, receiving input as a sequence and generating output in an autoregressive manner.

A language model that is pretrained on a text generation task and can be fine-tuned for specific tasks. Unlike BERT, GPT uses a decoder-only architecture and generates text in an autoregressive way.

A sampling method used in language models where the most probable token is chosen at each step of text generation, leading to repetitive or predictable text. Not ideal for creative or diverse text.

The ability of an agent to store and retrieve information, enabling it to learn and adapt over time.

A temporary storage space used by an agent to hold information relevant to the current task, like search results or conversation context.

A long-term storage space used by an agent to retain information about user preferences, historical interactions, and learned experiences.

The process of an AI agent evaluating its past decisions and actions to identify areas for improvement, ensuring it learns from successes and failures.

The ability of an AI agent to analyze its own performance critically and suggest improvements, acting as an internal feedback mechanism.

A strategy for solving complex problems by breaking them down into smaller, more manageable ones, ensuring a step-by-step approach.

A system for enhancing decision-making by retrieving relevant information from an agent's memory, including environmental data, past interactions, and learned experiences.

The process of updating an agent's memory with new information and experiences, enabling it to learn and adapt to new environments.

BLEU measures the overlap of n-grams (sequences of words) between generated and reference text. It's a simple, widely used metric, but often overestimates performance when there's a grammatical match, even if sentences don't convey the same meaning.

BERT score measures semantic similarity by comparing word vectors from a pre-trained BERT model. It assesses how similar the meaning of the generated text is to the reference text.

Exact Match (EM) is a binary metric that indicates if the generated text perfectly matches the reference text. It's commonly used in question answering, especially for closed-book settings where the model needs to retrieve the exact answer.

Ranking in text evaluation determines each word's position in the output sequence relative to its correct position in the reference text. Higher ranks indicate better word order and alignment, helping to evaluate the output's fluency.

Human evaluation involves subjective assessments by people to evaluate aspects of language quality that machines struggle to capture, such as fluency, coherence, and creativity.

LAMBADA is a benchmark dataset of narrative passages used to test a model's ability to predict the next word in a sentence. It's a challenging natural language understanding task.

ROCStories, HellaSwag, and StoryCloze are story-completion benchmarks that test a model's ability to select the most appropriate ending for a given story. They're useful for understanding a model's ability to comprehend and reason about narratives.

Question answering (QA) tasks measure a model's ability to provide answers to questions. Closed-book QA assesses a model's knowledge base, while open book QA allows access to information sources during the task.