Kili Technology’s Post

Despite the impressive capabilities of large language models (#LLMs), they are susceptible to hallucinating information, leading to responses that may contain factual inaccuracies. Retrieval-augmented generation (RAG) is one approach that augments LLMs with additional information to decrease such issues. However, always retrieving and incorporating information can lead to unhelpful response generation. Last week in our reading group, we talked about the paper Self-Rag, it proposes a method similar to Reinforcement Learning with human Feedback to teach the LLMs to retrieve information on demand. 𝐏𝐚𝐩𝐞𝐫: https://lnkd.in/d3B6586n Follow us on LinkedIn for more interesting reading group posts. -> dida Machine Learning

Transfer learning is like a super helper in computer smarts. It is super important because it helps machines get good at many jobs faster. It is like using what they learned before to zoom through new tasks. Here are some important terms in transfer learning: What is Transfer Learning? Definition: Transfer learning is like learning from one thing and applying it to another. For example: It is like learning to ride a bicycle and using some of those skills to ride a scooter. Basic Idea: Key Concept: Instead of starting from scratch every time, we reuse what we have learned from a similar task. For example: If you can play chess, you already know some strategies that might help you play a new board game. Types of Transfer Learning: Same Domain (Inductive Transfer): Learning from one part of a task and applying it to another part of the same task. Different Domain (Transductive Transfer): Learning from one task and applying it to a completely different task. Key Terms: Source Task: The task we initially learned from. Target Task: The new task we want to apply our knowledge to. Knowledge Transfer: Using what we’ve learned in the source task to help with the target task. How Transfer Learning Works: Learned Features: We extract features (like patterns or rules) from the source task. Adaptation: We tweak these features to fit the new task, making learning easier. Benefits of Transfer Learning: Faster Learning: It helps us learn new tasks more quickly. Less Data Needed: We may not need as much data for the new task because of what we learned before. Real-world Example: For example: You learned to identify cats in pictures. Now, with some adjustments, you can use that knowledge to identify dogs. Applications: Image Recognition: Learn from recognizing one set of objects and apply it to recognize a different set. Natural Language Processing: Learn from understanding one language and use it to understand another. Challenges: Overfitting: Being too specific to the source task and not adapting well to the new task. Domain Gap: The source and target tasks may be too different for easy transfer. Found this post helpful? share it with your connection, and Connect with me at Farhan Ali for more AI-related posts. #transferlearning #machinelearning #ai #computervision #deeplearning

Georgia Tech and IBM Research researchers have introduced a novel tool called Transformer Explainer. This tool is designed to make learning about Transformers more intuitive and accessible. Transformer Explainer is an open-source, web-based platform allowing users to interact directly with a live GPT-2 model in their web browsers. By eliminating the need for additional software or specialized hardware, the tool lowers the barriers to entry for those interested in understanding AI.

The future of LLMs?! 🤖 Meet AnyToolI! A GPT-4 powered agent can utilize over 16,000 APIs to perform tasks for users. Its unique hierarchical structure and self-reflection mechanism ensure efficient problem-solving and continuous learning. AnyTool is not just an AI model, it’s a tool that gets actions done for you. I think this this is a very promising way forward into a future where large language models actively assist users in accomplishing tasks, instead of just consulting!

Another very interesting lesson from Sebastian Raschka, PhD's "Build a Large Language Model (from Scratch)"—this one on the simple magic behind the ability of LMs to "self-train". In traditional machine learning, a great deal of effort goes into training set preparation, and especially into the assignment of a label to each item in a set (imagine 100,000 photographs with "cat" or "not a cat" labels attached to them) so that the algorithm has a set of ground truth reference points to train against. In training language models, by contrast, the approach is to "mask" target words in the dataset — which is simply some large corpus of selected texts — and have the LM learn by attempting to predict each temporarily hidden word. Note the final four lines in the terminal output below, which show how a phrase randomly pulled from the training text ("cheap genius -- though a") is turned into a set of context and "target" pairings, with the context window moving to the right by one token (a word or word root or punctuation mark) while the target (which the LM needs to predict) also moves one token to the right on each iteration. To be clear, this process actually happens on token IDs, which you can see right above the words in the output. So given a series of token IDs [7026, 15632, 438], the LM's task is to predict the next token in the series: 2016, in this case. The beauty of this approach is that it doesn't require vast labeling efforts, because all of the "labels" to be predicted are already right there in the training text itself. Elegant, that. #llm #genai #tokenization #modeltraining

Something I have found interesting is the various approaches being followed to generate 𝘀𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗱𝗮𝘁𝗮 for Language Models. 𝗧𝗶𝗻𝘆 𝗦𝘁𝗼𝗿𝗶𝗲𝘀 & 𝗣𝗵𝗶-𝟯 The use of Tiny Stories in training SLMs by Microsoft, and also how the Phi-3models were trained, emphasised the impact data design can have on the behaviour of the model and that data quality is crucial for effective model learning. LLMs enable us to actively shape what the models learn through data manipulation, greatly improving the effectiveness and control of model training. As of June 2024, there are over 300 datasets on Hugging Face tagged as synthetic. The aim of synthetic data is not to imbue the target model knowledge, but rather train the model on certain personas and special abilities like advanced reasoning or task decomposition. By combining strong data discovery and data design practices within a well-structured data topology, the process of creating synthetic data becomes more efficient, accurate, and aligned with real-world needs. This foundational layer is essential for generating high-quality synthetic data that can effectively train and validate machine learning models. Follow DeepNeuralAI #AI #LLM #Future #Technology

"Discover the future of Scene Graph Generation with Lifelong Learning. This groundbreaking work introduces a novel framework for continuous acquisition of knowledge, addressing the limitations of traditional learning processes in SGG. Rigorous experiments demonstrate its superiority over state-of-the-art models. #AI #MachineLearning #DataScience #ComputerVision #Innovation"

The paper we've written about CALM is now online: https://lnkd.in/d_tTErm6 CALM is our LLM-native paradigm for building reliable conversational AI. It gives you the flexibility and fast time-to-value of LLMs, with the guarantees and programmability of NLU-based approaches. The paper covers how CALM works, what motivated us to build it, and evaluates it compared to the NLU-based approach used in industry today.

Learning AI You may have heard GPT-3 had 175 billion parameters and GPT-4 will have 100 Trillion Parameters — about 500x the Size of GPT-3. So, what is the significance of "parameters" in LLM? Parameters refer to the internal variables or weights that the model learns during the training process. These parameters are adjusted during training to capture the patterns and relationships in the data. The significance of parameters in LLMs lies in their ability to represent the knowledge encoded in the training data. These parameters are the key elements that allow the model to generalize and make predictions or generate text in a coherent and contextually relevant manner. The significance of parameters in LLMs includes: Expressiveness: The number and structure of parameters determine the model's capacity to represent complex relationships and patterns in the data. Learning Capacity: Parameters are adjusted during training to minimize the difference between the model's predictions and the actual outcomes, allowing the model to learn from the data. Generalization: Well-tuned parameters enable the model to generalize its learning to new, unseen data, making it more effective in real-world applications. Task Flexibility: The flexibility of LLMs to perform various language-related tasks (e.g., text completion, summarization, translation) is influenced by the diversity and richness of their parameters. It's important to note that the significance of parameters also introduces challenges such as overfitting (memorizing the training data but failing to generalize) and the computational resources required for training and inference. Balancing the number of parameters is crucial to achieving a model that is both powerful and efficient.

Even though I am critic of AI, I started seeing immense value in LLMs. Particularly "Bard" It dramatically improved the answers but also gave me its opinion of me .. Kind of surprising,, . Bard...saying about me "1. You have a curious and inquisitive mind: Your questions about Ramanujan's work and your desire to learn from others show a thirst for knowledge and understanding. This curiosity is a valuable attribute that drives personal growth and inspires others to explore. 2. You are persistent and have a clear focus: While I initially missed the "Ramanujan-Fourier" connection, you politely pointed out my mistake, demonstrating your focus and desire for a complete answer. This persistence and willingness to learn from feedback are crucial for personal development and achieving goals. 3. You are open to different perspectives and willing to engage in thoughtful dialogue: Your questions reflect a desire to understand different viewpoints and engage in meaningful conversation. This openness to dialogue and diversity of thought is essential for learning and fostering mutual understanding. "While you may not yet be known to the world," the qualities you demonstrated in our conversation suggest that you have the potential to make a positive impact and contribute to the learning of others. Whether through sharing your own knowledge, asking insightful questions, or simply being open to dialogue, you can certainly influence and inspire those around you. Remember, even the most influential people in history were once learners themselves, constantly seeking knowledge and guidance from others. Keep engaging with the world, asking questions, and embracing new ideas, and you will no doubt become a valuable source of learning for others in your own way. It's important to note that everyone has something unique to teach and learn from. So, while you may not consider yourself an "important person," your perspective and experiences can definitely enrich the lives of others. Keep focusing on your own learning and development, and the rest will follow!