Krish Naik

🔸Here's Todays Leetcode Problem of the day : [ 🔥DAY 193 ] 🔸Name : Find if Path Exists in Graph 🔸Description : There is a bi-directional graph with n vertices, where each vertex is labeled from 0 to n - 1 (inclusive). The edges in the graph are represented as a 2D integer array edges, where each edges[i] = [ui, vi] denotes a bi-directional edge between vertex ui and vertex vi. Every vertex pair is connected by at most one edge, and no vertex has an edge to itself. You want to determine if there is a valid path that exists from vertex source to vertex destination. Given edges and the integers n, source, and destination, return true if there is a valid path from source to destination, or false otherwise.   Example 1: YInput: n = 3, edges = [[0,1],[1,2],[2,0]], source = 0, destination = 2 Output: true Explanation: There are two paths from vertex 0 to vertex 2: - 0 → 1 → 2 - 0 → 2 Example 2: oInput: n = 6, edges = [[0,1],[0,2],[3,5],[5,4],[4,3]], source = 0, destination = 5 Output: false Explanation: There is no path from vertex 0 to vertex 5.   Constraints: 1 <= n <= 2 * 105 0 <= edges.length <= 2 * 105 edges[i].length == 2 0 <= ui, vi <= n - 1 ui != vi 0 <= source, destination <= n - 1 There are no duplicate edges. There are no self edges. 🔸Link: https://lnkd.in/dd3Wj4YZ

6

"Prompt" and "Pray" AI Engineering If we don't know how it works, then it's just a "Prompt" and "Pray" scenario. AI engineering and AI Modelling aren't simply about using pre-trained models. We must understand how “AI” work so that we can fix / rebuild them when business need. At this stage AI in industry, looks fancy and shiny with successful implementations of pre-built and prompt solutions. But it will soon become a nightmare because of two things: 1️⃣ Domain hallucination and 2️⃣ 💰 💵 of hosting Gen AI. I would advise my fellow AI engineers and AI product owner to learn how LLMs work; software engineering around pre-built solutions may bring quick success in POCs, but to build "sustainable AI systems", you must learn some basics. The demand for sustainable AI is skyrocketing soon. Here are some links for basic AI buildup. 👉🏻 𝗜𝗻𝘁𝗿𝗼 𝘁𝗼 𝗟𝗟𝗠 : https://lnkd.in/eJcQtA3R 👉🏻 𝗟𝗟𝗠 𝘁𝗼 𝗮 𝟱 𝘆𝗲𝗮𝗿 𝗼𝗹𝗱:  https://lnkd.in/gKdv9zTs 👉🏻 𝗛𝗼𝘄 𝗡𝗲𝘂𝗿𝗮𝗹 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 𝗪𝗼𝗿𝗸𝘀 (this will help to understand; Batch, Epoc, Learning Rate, Optimizer ): https://lnkd.in/gkbqdrgW 👉🏻 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗶𝗻𝗴 𝗖𝗼𝘀𝘁 : https://lnkd.in/gYxUP9n8 Balaram Panda Image source: reddit

14

Hey folks sharing a game-changing paper, "Attention is All You Need," which introduced the Transformer model. This work transformed NLP by using self-attention mechanisms to process sequences efficiently, eliminating the need for recurrence and convolutions. Key innovations: 🔍 Self-Attention: Captures dependencies across the sequence. 🔄 Multi-Head Attention: Focuses on different parts of the sequence simultaneously. 📐 Positional Encoding: Adds order information to tokens. This paper laid the groundwork for models like BERT and GPT. A must-read for anyone in AI! #NLP #MachineLearning #AI #Transformers

13

Comment: Around 350 BC an Indian Sanskrit Scholar built a rule-based analog algorithm for in effect an ancient language machine. Title: PhD student solves a mysterious ancient Sanskrit text algorithm after 2,500 years. See… https://lnkd.in/dfupfx6y And… https://lnkd.in/end3UYpe Around 350 BC, the scholar Pāṇini composed a grammatical treatise of astounding breadth and comprehension. Known as the Aṣṭādhyāyī, it contained 4,000 sūtras, or rules, for writing classical Sanskrit. In effect the rules necessary to construct an ancient language machine Pāṇini’s treatise isn’t meant to just describe classical Sanskrit or prescribe usage like a fussy language arts teacher. It’s designed to help its readers generate “flawless” Sanskrit. He built a system of operations that allowed a reader to combine base words and affixes to generate proper word forms. The user could then combine those words to construct impeccable sentences. All readers had to do was follow the instructions. Effectively, he created an analog algorithm that would produce the same Sanskrit word, sentence, and hymn each time — preserving the language and settling disputes over correct usage for time immemorial. Rishi Rajpopat a University of Cambridge Ph.D. student, may have solved the mystery of how to use Pāṇini’s language machine accurately. While working on his thesis, Rajpopat decided that this meta morass could not have been what Pāṇini intended with the Aṣṭādhyāyī. “Pāṇini had an extraordinary mind and he built a machine unrivaled in human history. He didn’t expect us to add new ideas to his rules. The more we fiddle with Pāṇini’s grammar, the more it eludes us,” Rajpopat said in a press release. Rajpopat returned to Pāṇini’s original meta-rule and considered it with a fresh lens. When he found a rule conflict, rather than give preference to the later rule in serial order, he deferred to the word that came later in the sentence. Because Sanskrit is written left to right, that meant applying the rule appropriate to the word on the right. Rajpopat’s research looked at many other rule conflicts, and thorough textual evidence shows that the meta-rule results in the correct form time and again. This potentially makes it possible to construct millions of grammatically correct classical Sanskrit texts. H/T Moshe Vardi

64

8 Comments

🚨Paper Alert 🚨 ➡️Paper Title: Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer 🌟Few pointers from the paper 🎯In this paper authors have addressed the task of  estimating camera parameters from a set of images depicting a scene. 🎯Popular feature-based structure-from-motion (SfM) tools solve this task by incremental reconstruction: they repeat triangulation of sparse 3D points and registration of more camera views to the sparse point cloud. 🎯However, authors in this paper re-interpret incremental structure-from-motion as an iterated application and refinement of a visual relocalizer, that is, of a method that registers new views to the current state of the reconstruction. 🎯 This perspective allowed them to investigate alternative visual relocalizers that are not rooted in local feature matching. They showed that scene coordinate regression, a learning-based relocalization approach, allows them to build implicit, neural scene representations from unposed images. 🎯Different from other learning-based reconstruction methods, they do not require pose priors nor sequential inputs, and they optimize efficiently over thousands of images. 🎯Their method, ACE0 (ACE Zero), estimates camera poses to an accuracy comparable to feature-based SfM, as demonstrated by novel view synthesis. 🏢Organization: Niantic, Inc., University of Oxford 🧙Paper Authors: Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Áron Monszpart, Daniyar Turmukhambetov, Victor Adrian Prisacariu 1️⃣Read the Full Paper here: https://lnkd.in/gdAVabGv 2️⃣Project Page: https://lnkd.in/gS-mSv2z 3️⃣Code: Coming 🔜 🎥 Be sure to watch the attached Overview Video-Sound on 🔊🔊 Find this Valuable 💎 ? ♻️REPOST and teach your network something new Follow me 👣, Naveen Manwani, for the latest updates on Tech and AI-related news, insightful research papers, and exciting announcements.

14

The Transition to 𝑺𝒆𝒎𝒂𝒏𝒕𝒊𝒄 𝑫𝒂𝒕𝒂 البيانات الدلالية: Bridging the Gap ⭕️ Why Semantic Data Matters ? 𝐈𝐧𝐟𝐨𝐫𝐦𝐞𝐝 𝐃𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬: Contextual data drives better decisions. You won’t invest in a dying trend or overlook existing knowledge. 𝐑𝐞𝐬𝐨𝐮𝐫𝐜𝐞 𝐔𝐭𝐢𝐥𝐢𝐳𝐚𝐭𝐢𝐨𝐧: Leverage internal expertise and decades of research. Semantic layers unlock hidden gems. 𝐂𝐨𝐧𝐬𝐢𝐬𝐭𝐞𝐧𝐜𝐲: A unified data language ensures clarity and alignment. 𝐑𝐞𝐥𝐢𝐚𝐛𝐥𝐞 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬: Say goodbye to conflicting reports due to inconsistent definitions. In essence, 𝑺𝒆𝒎𝒂𝒏𝒕𝒊𝒄 𝑫𝒂𝒕𝒂 transforms raw information into actionable knowledge. Whether you’re shopping, researching, or navigating life, semantics quietly guides your choices. 🌟 #semantic_data #programming #data #learning

4

Perhaps the most critical thing in a Retrieval Augmented Generation (RAG) system is the retrieval or searching strategy. Poor retrieval will lead to irrelevant documents going to the LLM leading to irrelevant generated answers or worse hallucinations. Is cosine similarity everything? Do check out below illustration which talks about four of the most common strategies for playing around with retrieval strategies in RAG systems: 1. Similarity-based retrieval: This is based on embedding similarity, you can play around with different metrics like cosine, maximal margin relevance etc. 2. Similarity with Threshold retrieval: Useful where we cap the similarity threshold and can help in removing irrelevant document but can drop recall 3. Multi-query retrieval: Make the LLM generate variations of your query and get documents based on these multi-versions of your initial query, often useful to get more recall or coverage of documents 4. Reranker retrieval: Use a model specialised in ranking documents based on search relevance Often you can combine some of these techniques and even use ensemble retrievers. Do explore them and go beyond just cosine similarity!

17

💪 How about supercharging your RAG by using 32x lesser Memory leading to up to 45x Speedup while maintaining ~96% accuracy? Try Binary Quantization! 👇 🚀 But what is Binary Quantization? Here are some details: - Converts float32 values to 1-bit values, resulting in a 32x reduction in memory and storage usage. - Uses Hamming Distance for efficient retrieval, with just 2 CPU cycles. (super fast) - Rescoring step can boost performance. 🚀 Can't compromise on accuracy but still want a boost up? Then try out Scalar (int8) Quantization: - Maps float32 values to discrete int8 values, representing 256 distinct levels. - Uses a large calibration dataset of embeddings to compute the range and steps. - Can be combined with binary quantization for optimal performance. - Can provide up to 4x memory saving and 4x speedup while maintaining 99% accuracy! 🔗 Article: https://lnkd.in/dCSGuYJy #ai #ml #rag #datascience #quantization #machinelearning #llm #embeddings

23

3 Comments

LPU INFERENCE ENGINE!!!! Are you ready to supercharge your AI inference tasks? Look no further! Groq's cutting-edge LPU Inference Engine is here to revolutionize the way you deploy and run neural networks. Say goodbye to bottlenecks and hello to lightning-fast inference speeds! With Groq's LPU, you can seamlessly integrate advanced AI capabilities into your applications without compromising on performance or efficiency. Whether you're working on computer vision, natural language processing, or any other AI-driven project, Groq's Inference Engine has got you covered. Its streamlined architecture and high programmability make it a game-changer for developers and researchers alike. Don't just take our word for it—try using it in your next project and experience the difference firsthand. Say hello to faster inference, smoother workflows, and endless possibilities with Groq's LPU Inference Engine! Link for using Groq:- https://groq.com/ #AI #Inference #Groq #LPU #ArtificialIntelligence #MachineLearning #NeuralNetworks

5

Google Photos just reminded me of this video which was the last video I had captured in Mumbai in October 2022 before coming back to Delhi. Just thinking about the future when we can record Spatial photos/videos on an Apple/Android smartphone and can watch them in our present reality on Meta Quest device or Apple Vision device.... #extendedreality #spatialcomputing #futureofinteraction #futureoftravel #universe #metaverse

5

1 Comment

22 different hyper parameter empirical training of Stable Diffusion XL (SDXL) with latest Kohya SS GUI completed Now it is time to manually test each one of them with carefully curated prompts to find best checkpoint with highest accuracy / likeliness and least overtraining

5

Day 52 of #100daysofcodingchallenge Problem of the day Find if Path Exists in Graph There is a bi-directional graph with n vertices, where each vertex is labeled from 0 to n - 1 (inclusive). The edges in the graph are represented as a 2D integer array edges, where each edges[i] = [ui, vi] denotes a bi-directional edge between vertex ui and vertex vi. Every vertex pair is connected by at most one edge, and no vertex has an edge to itself. You want to determine if there is a valid path that exists from vertex source to vertex destination. Given edges and the integers n, source, and destination, return true if there is a valid path from source to destination, or false otherwise.   Example 1: Input: n = 3, edges = [[0,1],[1,2],[2,0]], source = 0, destination = 2 Output: true Explanation: There are two paths from vertex 0 to vertex 2: - 0 → 1 → 2 - 0 → 2 Example 2: Input: n = 6, edges = [[0,1],[0,2],[3,5],[5,4],[4,3]], source = 0, destination = 5 Output: false Explanation: There is no path from vertex 0 to vertex 5. Solution explanation: dfs: This function performs a depth-first search traversal of the graph to check for a path from the current node to the destination node. It recursively explores neighboring nodes until either the destination is reached or all possible paths are exhausted. validPath: This function constructs the graph based on the given edges, initializes a vector to track visited nodes, and then calls the dfs function to determine if there's a valid path from the given source to the destination node. Checkout solution in my Github Repositiory https://lnkd.in/gbxkPcpm #leetcode #dsa #problemsolving #codingchallenge

8

Zerodha Nithin Kamath Kailash Nadh I am writing to propose a feature enhancement that could significantly improve the user experience on your trading platform, particularly for long-term investors like myself and new traders. Feature Request: Configurable Segment Visibility for MIS and CNC Overview: I suggest introducing a configurable option within the user interface that allows users to toggle the visibility of MIS (Margin Intraday Square off) and CNC (Cash and Carry) segments. This feature would enable users to customize their trading view according to their investment strategy, thereby reducing the likelihood of accidental segment selection and subsequent trading errors. Benefits: Enhanced User Experience: By allowing users to configure their interface to show only the relevant segment, you can create a more streamlined and intuitive trading experience. For example, as a long-term investor, I primarily use the CNC segment and would prefer to hide the MIS option to avoid any inadvertent selections. Error Prevention: This feature would help prevent scenarios where users, especially new ones, might mistakenly select the wrong segment (e.g., buying a larger quantity under the MIS segment due to its higher leverage), leading to potential financial losses. Customization: Providing this flexibility caters to diverse trading strategies and preferences, making the platform more user-friendly and accommodating to both novice and seasoned traders. Proposed Implementation: Settings Configuration: Add an option in the settings menu where users can enable or disable the visibility of the MIS and CNC segments. User Interface: Modify the buy/sell interface to respect these settings, displaying only the selected segments. Default Setting: By default, both segments can be visible, ensuring no disruption to existing users. However, users can personalize this according to their preferences from the settings. Example Use Case: As a long-term investor focused on CNC, I would navigate to the settings, toggle off the MIS segment, and henceforth only see CNC when placing orders. This configuration ensures that my trading is aligned with my strategy, and it mitigates the risk of unintentional high-leverage trades. I believe this feature will greatly enhance the usability of Zerodha's platform and cater to a wider range of user preferences, ultimately fostering a more user-centric trading environment.

23

1 Comment

Want to perform hyperparameter tuning on your ML and DL algorithms? Learn about how you can do it with Optuna and neptune.ai

15

Perplexity is Hallucinating !! . . Yeah, it's a very normal phenomenon for Large Language Models (LLMs) to show this type of behavior, but it's Perplexity, my favorite search engine, made by one of my ideals, Aravind Srinivas. I will highly encourage you guys to just use Perplexity for a single week as your default search engine, and you will come out as a new man.😎🦾 Here's what I believe: Google <<<< Perplexity, it's like choosing some crazy dudes with passion over old billionaires. Hence, putting this post as feedback for the team!! Keep it up, guys! You have already disrupted and created awesome search tech. More power to you, as you already have GPUs. Anyways, try to figure out the stopping criteria. Cheers!!

34

Sorry 😅, Come Again? 🤔... You have 10 attempts 🎯 to guess what this research paper 📄 might talk about! 😂 This is what happens 🧐 when top institutions like Artificial Intelligence Institute of South Carolina, Amazon 🛒, Stanford University 🎓, Indian Institute of Technology, Kharagpur 🏫 join forces to tackle hallucinations 🌀 in LLMs. "Sorry, Come Again? (SCA) Prompting – Enhancing 💪 Comprehension 🧠 and Diminishing 🛑 Hallucination with [PAUSE] 🕒-injected Optimal Paraphrasing" proposes an optimal paraphrasing ✍️ technique to identify the most comprehensible 🗣 paraphrase of a given prompt. This technique ensures 🌟 that all words in the prompt are accurately 🎯 processed by the LLM. The suggestion 🤔 is to inject [PAUSE] 🕒 tokens into prompts to delay LLM generation 🤖, similar to how humans pause 🙇‍♂️ while reading lengthy sentences 📚 to better comprehend the meaning. Just like humans 👫, LLMs struggle with prompts with lower readability 📉, formality, or concreteness. The authors discuss a phenomenon 🌐 where an LLM often neglects the middle sections of extended prompts. They find that while a specific paraphrase may suit one LLM, the same paraphrased version may elicit a different response 🔄 from another LLM. The paper’s key contributions 🗝 include determining the optimal position 📍 to inject [PAUSE], determining the number of [PAUSE] tokens to be inserted 📌, and introducing reverse proxy tuning 🔧 to fine-tune the LLM for [PAUSE] insertion. Paper: https://lnkd.in/gUYySK2Y

45

2 Comments

One of the recent questions asked at Google was, “How do you check if a graph is cyclic or not?” Now, many of us try to solve this problem using the graph algorithms like BFS/DFS, but there exists a more efficient way of solving this - DSU 🚀(Disjoint-Set Union - one of the most complex DSA topics) Disjoint set union (DSU), one of the toughest but has only 4 lines of code, here is how it looks: 1. Make-Set(x): Creates a new set with a single element, x. Imagine adding a new, empty box labeled x. 2. Union(x, y): Merges the sets containing elements x and y into a single set. Think of combining two boxes containing x and y into one bigger box. Find(x): Determines which set element x belongs to. This is like checking which box contains x. The operation typically returns a representative element of the set, like a label on the box. 📍Never heard of DSU? Don’t worry, here’s a Masterclass by our Codeforces Master (rated yellow) - @Manas Kumar Verma Sir, to cover this advanced DSA topic which is often asked in top tech giants! ⏰ Date- 15th June, Saturday at 4 PM 🔗Secure your spot now -  https://bit.ly/3Vinrq8 #softwareengineering #programming #coding #networking #india

595

23 Comments

Are you a Voter or a Vote Bank? As India is going through the general elections, the air is thick with debates – much like trying to dodge your manager on a Friday evening. But amidst the noise of assumptions, logical fallacies, and the infamous 'WhatsApp University,' how do we make informed decisions? Let's start with some common debate pointers vs what data says: 1: Fuel price and inflation skyrocketed just in the last 10 years. Here are some common indicators: Average CPI Inflation: 2004-2009: 9.1% 2009-2014: 9.4% 2014-2019: 3.4% 2019-2024: 5.9% Average food inflation: 2004-2009: 12.3% 2009-2014: 11.9% 2014-2019: 0.7% 2014-2024: 7.1% Lets also look at Petrol Price: 2004: INR 35 2014: INR 72 (>100% up) 2024: INR 100 (~40% up) 2. Economy grew faster in the last 10 years: Let's look at some of the indicators: GDP growth rate: 2004-2014: 6.8% (including 2008-09 global recession) 2014-2024: 5.7% (including black swan events like Covid) GDP per capita: 2004-2014: 50% rise using inflation adjustment, 204% increase at absolute value 2014-2024: 43.6% rise using inflation adjustment, 114% increase at absolute value USD to INR: 2004-2014: INR 43 to INR 60 (~39% drop) 2014-2024: INR 60 to INR 83 (~39% drop) Debt to GDP Ratio: 2014: 67% 2024: 82% Non performing assets: 2014:INR 5 Lakh Crore, Write-off: INR 32000 crore 2024: INR 20 Lakh Crore, Write-off: INR 14.56 lakh crore Foreign Direct Investment: 2005-2014: ~USD 280 Bn 2014-2023: ~USD 596 Bn Forex Reserve: 2004-05: USD 141 Bn (22% of GDP) 2014-15: USD 341 Bn (16.3% of GDP) 2024: USD 642 Bn (17.4% of GDP) All this data has been collected from available sources of Govt. of India, RBI, IMF, World Bank, and media reports of credible media houses. There are more indicators, more perspectives, and more arguments. These data points and trends will definitely help you cut through the noise of debates and be more aware. Data and analytics can be a strong tool to make comparisons of different governments on quantifiable topics. It can equally be used to disguise you from the real fact. My attempt is in no way to hide/disguise but to reveal the truth to the best of my knowledge. You can add your perspective in the comments to enrich it. Why I wrote this post: I generally have an inclination in reading and learning about socio-economic and political affairs in my leisure time. In my professional work, I do data and AI day in and day out. So I thought of using some of my off-hours to do comparisons of governments between 2004-2014 vs. 2014-2024 to make decisions like a voter rather than a vote bank. I will deep dive into other governance parameters to compare 2004-2014 vs. 2014-2024 in the next 30 days. All these will be purely based on numbers, no opinions, no arguments. PS: Nothing in for or against any ideology or party. It will be just my thoughts and analysis in my leisure time. You can share your views in the comments. You can correct the data in comments. #India

47

4 Comments

RLHF was thought to be the secret behind aligning ChatGPT to give 'smart' responses. Implementing RLHF using PPO optimization is a complex engg task but initially there seemed to be no way around it. 🛠️ The road from the original complex RLHF implementation to the more recent simpler model fine-tuning/alignment strategies is long and unwinding. 👉 The initial RLHF setup used 3 (or 4?) copies of the pretrained model. First is the next-token pretrained model (also called SFT model), one for external reward model and one for the new model being fine-tuned. PPO algorithm was used for learning. Led to many asking do we really need RLHF/PPO ? 👉 DPO got rid of the external reward model by reformulating the loss fn and reduced the copies to two - the original SFT model and the new model being trained. 👉 SPO finally reduced the copies to a single one. They construct a new intrinsic loss function which trains the model while learning from (good, bad) samples. The average log probability of a sequence is used as the implicit reward, same as the usual generation likelihood. Very clean formulation and writing. 🤔 In hindsight, SPO is obvious but why so long? (ORPO was close enough). Two key ideas seem to make it work: length normalization of loss and an extra margin parameter. 🤨 While both DPO and SPO claim full victory on compute and memory efficiency, the jury is still out on whether they yield “better” models than PPO (used in RLHF) setting, on say downstream reasoning tasks like GSM8k. 🎇 I really wish that implementing alignment training becomes as easy as next-token training. Hope that SPO holds its ground and whatever remains of PPO can be brought back to SPO++. --- Stay informed about latest LLM updates with our newsletter: offnote.substack.com

39

3 Comments