Datasets

Explore resources related to datasets to help implement AI solutions for your business.

Getting datasets to train AI to find a business opportunity and develop growth strategy

Listened to a podcast about training AI on competitive, consumer and market datasets to find the right segments for business opportunities. One example they used was feeding it case studies of companies that have become unicorns - building out their marketing and sales strategies so the AI can find 'rules'. I'm launching a telehealth clinic but I want to develop the optimal strategy for the right first customer segment, how we evolve as a platform (e.g. become more like gymshark or more health orientated) and what the marketing strategy should be. Therefore, I need a bunch of datasets such as consumer demographics and awareness, competitive offering and performance and growth lessons from benchmark companies and their strategies. How could I find such data sets? How could I put this together? Is there a scrappy version I could put together through scraping? I'm wary of feeding it shit data and getting shit back so would love some guidance.

#reddit #getting #datasets

[R] Apple’s Machine Learning Researchers Have Developed A No-Code AI Platform Called ‘Trinity’ For Complex Spatial Datasets

techsucker•This week

[R] Apple’s Machine Learning Researchers Have Developed A No-Code AI Platform Called ‘Trinity’ For Complex Spatial Datasets

Apple’s machine learning research team has developed a no-code Artificial Intelligence (AI) platform called Trinity. This AI is designed to enable machine learning researchers and non-technical geospatial domain experts alike to experiment with different signals or datasets in order to solve problems on their own, such as complex issues that arise from the world around us every day. The ability to solve diverse problems is made possible by transforming complex Spatio-temporal datasets so that they can be consumed and solved with a standard deep learning model, like Convolutional Neural Networks. This new way of looking at data has the potential to formulate disparate problems in one standardized form for easy consumption. Quick Read: https://www.marktechpost.com/2021/07/23/apples-machine-learning-researchers-have-developed-a-no-code-ai-platform-called-trinity-for-complex-spatial-datasets/ Paper: https://arxiv.org/pdf/2106.11756.pdf

#reddit #apple’s #machine

brad_wade07•This week

AI for Business Consultancy

Good morning folks, I am doing some research on the impact of AI on business consultancy. I want to initially try out existing models to first learn how to improve my datasets before diving into building my own model. I'm specifically looking for ideas about parameters or metrics that I should be covering for a small cafe business. Since I'm coming from a CS background, my knowledge of the restaurant/cafe industry is pretty limited, so I thought I'd reach out to this community for some guidance. What kind of parameters should I be considering besides the obvious ones like customer and sale data, Inventory information, etc. Pricing strategy analysis? I'm really open to any suggestions! I appreciate any help or direction you all could provide. Just trying to get my bearings in this space before going too deep down any particular path. Thanks in advance!

#reddit #business #consultancy

brad_wade07•This week

AI for Business Consultancy

#reddit #business #consultancy

My Manager Thinks ML Projects Takes 5 Minutes 🤦‍♀️

SaraSavvy24•This week

My Manager Thinks ML Projects Takes 5 Minutes 🤦‍♀️

Hey, everyone! I’ve got to vent a bit because work has been something else lately. I’m a BI analyst at a bank, and I’m pretty much the only one dealing with machine learning and AI stuff. The rest of my team handles SQL and reporting—no Python, no R, no ML knowledge AT ALL. You could say I’m the only one handling data science stuff So, after I did a Python project for retail, my boss suddenly decided I’m the go-to for all things ML. Since then, I’ve been getting all the ML projects dumped on me (yay?), but here’s the kicker: my manager, who knows nothing about ML, acts like he’s some kind of expert. He keeps making suggestions that make zero sense and setting unrealistic deadlines. I swear, it’s like he read one article and thinks he’s cracked the code. And the best part? Whenever I finish a project, he’s all “we completed this” and “we came up with these insights.” Ummm, excuse me? We? I must’ve missed all those late-night coding sessions you didn’t show up for. The higher-ups know it’s my work and give me credit, but my manager just can’t help himself. Last week, he set a ridiculous deadline of 10 days for a super complex ML project. TEN DAYS! Like, does he even know that data preprocessing alone can take weeks? I’m talking about cleaning up messy datasets, handling missing values, feature engineering, and then model tuning. And that’s before even thinking about building the model! The actual model development is like the tip of the iceberg. But I just nodded and smiled because I was too exhausted to argue. 🤷‍♀️ And then, this one time, they didn’t even invite me to a meeting where they were presenting my work! The assistant manager came to me last minute, like, “Hey, can you explain these evaluation metrics to me so I can present them to the heads?” I was like, excuse me, what? Why not just invite me to the meeting to present my own work? But nooo, they wanted to play charades on me So, I gave the most complicated explanation ever, threw in all the jargon just to mess with him. He came back 10 minutes later, all flustered, and was like, “Yeah, you should probably do the presentation.” I just smiled and said, “I know… data science isn’t for everyone.” Anyway, they called me in at the last minute, and of course, I nailed it because I know my stuff. But seriously, the nerve of not including me in the first place and expecting me to swoop in like some kind of superhero. I mean, at least give me a cape if I’m going to keep saving the day! 🤦‍♀️ Honestly, I don’t know how much longer I can keep this up. I love the work, but dealing with someone who thinks they’re an ML guru when they can barely spell Python is just draining. I have built like some sort of defense mechanism to hit them with all the jargon and watch their eyes glaze over How do you deal with a manager who takes credit for your work and sets impossible deadlines? Should I keep pushing back or just let it go and keep my head down? Any advice! TL;DR: My manager thinks ML projects are plug-and-play, takes credit for my work, and expects me to clean and process data, build models, and deliver results in 10 days. How do I deal with this without snapping? #WorkDrama

#reddit #manager #thinks

anishathalye•This week

MIT Introduction to Data-Centric AI

Announcing the first-ever course on Data-Centric AI. Learn how to train better ML models by improving the data. Course homepage | Lecture videos on YouTube | Lab Assignments The course covers: Data-Centric AI vs. Model-Centric AI Label Errors Dataset Creation and Curation Data-centric Evaluation of ML Models Class Imbalance, Outliers, and Distribution Shift Growing or Compressing Datasets Interpretability in Data-Centric ML Encoding Human Priors: Data Augmentation and Prompt Engineering Data Privacy and Security MIT, like most universities, has many courses on machine learning (6.036, 6.867, and many others). Those classes teach techniques to produce effective models for a given dataset, and the classes focus heavily on the mathematical details of models rather than practical applications. However, in real-world applications of ML, the dataset is not fixed, and focusing on improving the data often gives better results than improving the model. We’ve personally seen this time and time again in our applied ML work as well as our research. Data-Centric AI (DCAI) is an emerging science that studies techniques to improve datasets in a systematic/algorithmic way — given that this topic wasn’t covered in the standard curriculum, we (a group of PhD candidates and grads) thought that we should put together a new class! We taught this intensive 2-week course in January over MIT’s IAP term, and we’ve just published all the course material, including lecture videos, lecture notes, hands-on lab assignments, and lab solutions, in hopes that people outside the MIT community would find these resources useful. We’d be happy to answer any questions related to the class or DCAI in general, and we’d love to hear any feedback on how we can improve the course material. Introduction to Data-Centric AI is open-source opencourseware, so feel free to make improvements directly: https://github.com/dcai-course/dcai-course.

#reddit #introduction #data-centric

How do I get a job in this job market? How do I stand out from the crowd?

__proximity__•This week

How do I get a job in this job market? How do I stand out from the crowd?

About me - I am an international grad student graduating in Spring 2025. I have been applying for jobs and internships since September 2024 and so far I haven't even been able to land a single interview. I am not an absolute beginner in this field. Before coming to grad school I worked as an AI Software Engineer in a startup for more than a year. I have 2 publications one in the WACV workshop and another in ACM TALLIP. I have experience in computer vision and natural language processing, focusing on multimodal learning and real-world AI applications. My academic projects include building vision-language models, segmentation algorithms for medical imaging, and developing datasets with human attention annotations. I’ve also worked on challenging industry projects like automating AI pipelines and deploying real-time classifiers. How can I improve my chances in this competitive job market? Are there specific strategies for international students navigating U.S. tech job applications? How can I stand out, especially when competing with candidates from top schools and with more experience?

#reddit #market?#stand

Additional-Ad-7043•This week

How do byte-level language models work?

I've recently been trying to pre-train my own small language model on the tiny-series datasets on huggingface: https://huggingface.co/collections/nampdn-ai/tiny-series-6503910fd491144159519c70. I also wanted to use a model similar to MEGABYTE: https://arxiv.org/pdf/2305.07185.pdf, but I don't understand how using bytes would work. The only implementation I could find: https://github.com/lucidrains/MEGABYTE-pytorch used str(chr(max(32, token))) to decode any token (byte) to a character and put the embedding size as 256. Firstly, why 256 and not 256-32 as any values below 32 are ignored? Also, many byte-level models including this and ByteT5 mention that they can process any text sequence even in a multilingual setting, however how would that be true if we are only using one byte, would we have to move to 2 bytes or use an UNK token, and if we did use 2 bytes that would make our embedding size around 65000 which defeats sort of the point as one of the advantages mentioned is that we are able to use a small embedding matrix? Furthermore, most language models add special tokens like bos, eos, unk and even for llama they use beginning of instruction, end of instruction, and more for system instructions, response, context... Should I use something like this as my dataset has some structures where there is a context, instruction and response, and if i did how would I add these if I'm using byte-level encodings? Final questions: Firstly, for the datasets mentioned including code,stories,webtext,... would I tokenise all of these datasets then concatenate them to then randomly sample from, or should i train seperately on each as some like code and webtext are much larger than the others? Finally, for the webtext part of the dataset, there is a passage of text then a passage analysing the text (main ideas,purpose,...), how should I encode this, should I use an extra ANALYSE token or just concatenate? Thank you for reading this far, I am sort of a beginner so if I said something stupid please point it out. Also, if there were unclear parts in my question I'm sorry as I struggled how to word these questions. Any help would be appreciated!

#reddit #byte-level #language

𝗠𝗮𝘀𝘁𝗲𝗿 𝗟𝗶𝗻𝗲𝗮𝗿 𝗥𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝗼𝗻: 𝗔 𝗠𝘂𝘀𝘁-𝗞𝗻𝗼𝘄 𝗳𝗼𝗿 𝗕𝗲𝗴𝗶𝗻𝗻𝗲𝗿𝘀

Ambitious-Fix-3376•This week

𝗠𝗮𝘀𝘁𝗲𝗿 𝗟𝗶𝗻𝗲𝗮𝗿 𝗥𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝗼𝗻: 𝗔 𝗠𝘂𝘀𝘁-𝗞𝗻𝗼𝘄 𝗳𝗼𝗿 𝗕𝗲𝗴𝗶𝗻𝗻𝗲𝗿𝘀

Scikit-learn vs Statsmodel Linear regression is often the first model introduced to those stepping into the world of data science and machine learning. A deep understanding of this fundamental concept is crucial for building a solid foundation. In this post, I explore two widely used approaches to linear regression, each with distinct purposes: 1️⃣ 𝗦𝗰𝗶𝗸𝗶𝘁-𝗟𝗲𝗮𝗿𝗻’𝘀 𝗟𝗶𝗻𝗲𝗮𝗿 𝗥𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝗼𝗻: Optimized for machine learning applications and large datasets, this model focuses on efficiency and scalability. 2️⃣ 𝗦𝘁𝗮𝘁𝘀𝗺𝗼𝗱𝗲𝗹𝘀’ 𝗢𝗿𝗱𝗶𝗻𝗮𝗿𝘆 𝗟𝗲𝗮𝘀𝘁 𝗦𝗾𝘂𝗮𝗿𝗲𝘀 (𝗢𝗟𝗦): Known for its comprehensive statistical insights, this approach provides a detailed report ideal for understanding relationships and diagnosing issues like multicollinearity. It’s essential to gain hands-on experience with both libraries to appreciate their unique strengths. To make this learning process more accessible, I’ve created detailed videos and example code to guide you through practical applications: 🎥 𝗛𝗮𝗻𝗱𝘀-𝗢𝗻 𝗧𝘂𝘁𝗼𝗿𝗶𝗮𝗹𝘀: 📌 Learn Linear Regression in Python with LLM Prompt Chaining : https://www.youtube.com/watch?v=KOEG4rs1SUU 📌 In-Depth Linear Regression: Statsmodels OLS, Multicollinearity, and VIF : https://www.youtube.com/watch?v=QQWKY30XzNA 💻 𝗖𝗼𝗱𝗲 𝗳𝗼𝗿 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗲: 🔗 Scikit-Learn Implementation: https://github.com/pritkudale/ML-for-Teachers/blob/main/Linear%20Regression/Linear\Regression.ipynb 🔗 Statsmodels Implementation: https://github.com/pritkudale/ML-for-Teachers/blob/main/Linear%20Regression/Linear\regression\using\stats\model.ipynb What makes these tutorials unique? I’ve incorporated LLM prompt chaining, enabling beginners to confidently write code without requiring extensive Python expertise. 📩 𝘚𝘵𝘢𝘺 𝘢𝘩𝘦𝘢𝘥 𝘪𝘯 𝘈𝘐 𝘢𝘥𝘷𝘢𝘯𝘤𝘦𝘮𝘦𝘯𝘵𝘴! 𝘚𝘶𝘣𝘴𝘤𝘳𝘪𝘣𝘦 𝘵𝘰 𝘰𝘶𝘳 𝘯𝘦𝘸𝘴𝘭𝘦𝘵𝘵𝘦𝘳 𝘧𝘰𝘳 𝘦𝘹𝘱𝘦𝘳𝘵 𝘪𝘯𝘴𝘪𝘨𝘩𝘵𝘴: 𝘝𝘪𝘻𝘶𝘢𝘳𝘢 𝘈𝘐 𝘕𝘦𝘸𝘴𝘭𝘦𝘵𝘵𝘦𝘳: https://vizuara.ai/email-newsletter/

#reddit #𝗠𝗮𝘀𝘁𝗲𝗿 #𝗟𝗶𝗻𝗲𝗮𝗿

[Help Needed] How to tackle AI for DNS Security in a Hackathon for Beginner.

Baby-Boss0506•This week

[Help Needed] How to tackle AI for DNS Security in a Hackathon for Beginner.

Hi everyone! I've been selected to participate in an AI and Cybersecurity Hackathon, and the group I'm in focuses on AI for DNS Security. Our goal is to implement AI algorithms to detect anomalies and enhance DNS security. Here’s the catch: I have no prior background in cybersecurity, and I’m also a beginner in applying AI to real-world security problems. I’d really appreciate some guidance from this amazing community on how to approach this challenge. A bit more about the project: Objective: Detect anomalies in DNS traffic (e.g., malicious requests, tunneling, etc.). AI tools: We’re free to choose algorithms, but I’m unsure where to start—supervised vs. unsupervised learning? My skillset: Decent grasp of Python (Pandas, Scikit-learn, etc.) and basic ML concepts. No practical experience in network security or analyzing DNS traffic. What I’m looking for: Datasets: Any recommendations for open-source DNS datasets or synthetic data creation methods? AI methods: Which models work best for anomaly detection in DNS logs? Are there any relevant GitHub projects? Learning resources: Beginner-friendly material on DNS security and the application of AI in this domain. Hackathon tips: How can I make the most of this opportunity and contribute effectively to my team? Bonus question: If you’ve participated in similar hackathons, what strategies helped you balance learning and execution within a short timeframe? Thank you so much in advance for any advice, resources, or personal experiences you can share! I’ll make sure to share our project results and lessons learned after the hackathon.

#reddit #[help #needed]

Ready_Seesaw8408•This week

Implementing AI in real world problems

Hello everyone, i participated in a hackathon this weekend and i wanted to implement a ML model to solve the given problem. I faced a problem while collecting datasets, i didnt really find any helpful datasets in Kaggle or Hugging Face so i dropped the idea from the MVP solution. I wanna know how can implement ai models to solve some specific real world problems in the future like the one they gave us and thank you.

#reddit #implementing #real

[D] Why I'm Lukewarm on Graph Neural Networks

0.6

VodkaHaze•This week

[D] Why I'm Lukewarm on Graph Neural Networks

TL;DR: GNNs can provide wins over simpler embedding methods, but we're at a point where other research directions matter more I also posted it on my blog here, has footnotes, a nicer layout with inlined images, etc. I'm only lukewarm on Graph Neural Networks (GNNs). There, I said it. It might sound crazy GNNs are one of the hottest fields in machine learning right now. [There][1] were at least [four][2] [review][3] [papers][4] just in the last few months. I think some progress can come of this research, but we're also focusing on some incorrect places. But first, let's take a step back and go over the basics. Models are about compression We say graphs are a "non-euclidean" data type, but that's not really true. A regular graph is just another way to think about a particular flavor of square matrix called the [adjacency matrix][5], like this. It's weird, we look at run-of-the-mill matrix full of real numbers and decide to call it "non-euclidean". This is for practical reasons. Most graphs are fairly sparse, so the matrix is full of zeros. At this point, where the non-zero numbers are matters most, which makes the problem closer to (computationally hard) discrete math rather than (easy) continuous, gradient-friendly math. If you had the full matrix, life would be easy If we step out of the pesky realm of physics for a minute, and assume carrying the full adjacency matrix around isn't a problem, we solve a bunch of problems. First, network node embeddings aren't a thing anymore. A node is a just row in the matrix, so it's already a vector of numbers. Second, all network prediction problems are solved. A powerful enough and well-tuned model will simply extract all information between the network and whichever target variable we're attaching to nodes. NLP is also just fancy matrix compression Let's take a tangent away from graphs to NLP. Most NLP we do can be [thought of in terms of graphs][6] as we'll see, so it's not a big digression. First, note that Ye Olde word embedding models like [Word2Vec][7] and [GloVe][8] are [just matrix factorization][9]. The GloVe algorithm works on a variation of the old [bag of words][10] matrix. It goes through the sentences and creates a (implicit) [co-occurence][11] graph where nodes are words and the edges are weighed by how often the words appear together in a sentence. Glove then does matrix factorization on the matrix representation of that co-occurence graph, Word2Vec is mathematically equivalent. You can read more on this in my [post on embeddings][12] and the one (with code) on [word embeddings][13]. Even language models are also just matrix compression Language models are all the rage. They dominate most of the [state of the art][14] in NLP. Let's take BERT as our main example. BERT predicts a word given the context of the rest of the sentence. This grows the matrix we're factoring from flat co-occurences on pairs of words to co-occurences conditional on the sentence's context, like this We're growing the "ideal matrix" we're factoring combinatorially. As noted by [Hanh & Futrell][15]: [...] human language—and language modelling—has infinite statistical complexity but that it can be approximated well at lower levels. This observation has two implications: 1) We can obtain good results with comparatively small models; and 2) there is a lot of potential for scaling up our models. Language models tackle such a large problem space that they probably approximate a compression of the entire language in the [Kolmogorov Complexity][16] sense. It's also possible that huge language models just [memorize a lot of it][17] rather than compress the information, for what it's worth. Can we upsample any graph like language models do? We're already doing it. Let's call a first-order embedding of a graph a method that works by directly factoring the graph's adjacency matrix or [Laplacian matrix][18]. If you embed a graph using [Laplacian Eigenmaps][19] or by taking the [principal components][20] of the Laplacian, that's first order. Similarly, GloVe is a first-order method on the graph of word co-occurences. One of my favorites first order methods for graphs is [ProNE][21], which works as well as most methods while being two orders of magnitude faster. A higher-order method embeds the original matrix plus connections of neighbours-of-neighbours (2nd degree) and deeper k-step connections. [GraRep][22], shows you can always generate higher-order representations from first order methods by augmenting the graph matrix. Higher order method are the "upsampling" we do on graphs. GNNs that sample on large neighborhoods and random-walk based methods like node2vec are doing higher-order embeddings. Where are the performance gain? Most GNN papers in the last 5 years present empirical numbers that are useless for practitioners to decide on what to use. As noted in the [OpenGraphsBenchmark][4] (OGB) paper, GNN papers do their empirical section on a handful of tiny graphs (Cora, CiteSeer, PubMed) with 2000-20,000 nodes. These datasets can't seriously differentiate between methods. Recent efforts are directly fixing this, but the reasons why researchers focused on tiny, useless datasets for so long are worth discussing. Performance matters by task One fact that surprises a lot of people is that even though language models have the best performance in a lot of NLP tasks, if all you're doing is cram sentence embeddings into a downstream model, there [isn't much gained][23] from language models embeddings over simple methods like summing the individual Word2Vec word embeddings (This makes sense, because the full context of the sentence is captured in the sentence co-occurence matrix that is generating the Word2Vec embeddings). Similarly, [I find][24] that for many graphs simple first-order methods perform just as well on graph clustering and node label prediction tasks than higher-order embedding methods. In fact higher-order methods are massively computationally wasteful for these usecases. Recommended first order embedding methods are ProNE and my [GGVec with order=1][25]. Higher order methods normally perform better on the link prediction tasks. I'm not the only one to find this. In the BioNEV paper, they find: "A large GraRep order value for link prediction tasks (e.g. 3, 4);a small value for node classification tasks (e.g.1, 2)" (p.9). Interestingly, the gap in link prediction performance is inexistant for artificially created graphs. This suggests higher order methods do learn some of the structure intrinsic to [real world graphs][26]. For visualization, first order methods are better. Visualizations of higher order methods tend to have artifacts of their sampling. For instance, Node2Vec visualizations tend to have elongated/filament-like structures which come from the embeddings coming from long single strand random walks. See the following visualizations by [Owen Cornec][27] created by first embedding the graph to 32-300 dimensions using a node embedding algorithm, then mapping this to 2d or 3d with the excellent UMAP algorithm, like this Lastly, sometimes simple methods soundly beat higher order methods (there's an instance of it in the OGB paper). The problem here is that we don't know when any method is better than another and we definitely don't know the reason. There's definitely a reason different graph types respond better/worse to being represented by various methods. This is currently an open question. A big part of why is that the research space is inundated under useless new algorithms because... Academic incentives work against progress Here's the cynic's view of how machine learning papers are made: Take an existing algorithm Add some new layer/hyperparameter, make a cute mathematical story for why it matters Gridsearch your hyperparameters until you beat baselines from the original paper you aped Absolutely don't gridsearch stuff you're comparing against in your results section Make a cute ACRONYM for your new method, put impossible to use python 2 code on github (Or no code at all!) and bask in the citations I'm [not][28] the [only one][29] with these views on the state reproducible research. At least it's gotten slightly better in the last 2 years. Sidebar: I hate Node2Vec A side project of mine is a [node embedding library][25] and the most popular method in it is by far Node2Vec. Don't use Node2Vec. [Node2Vec][30] with p=1; q=1 is the [Deepwalk][31] algorithm. Deepwalk is an actual innovation. The Node2Vec authors closely followed the steps 1-5 including bonus points on step 5 by getting word2vec name recognition. This is not academic fraud -- the hyperparameters [do help a tiny bit][32] if you gridsearch really hard. But it's the presentable-to-your-parents sister of where you make the ML community worse off to progress your academic career. And certainly Node2Vec doesn't deserve 7500 citations. Progress is all about practical issues We've known how to train neural networks for well over 40 years. Yet they only exploded in popularity with [AlexNet][33] in 2012. This is because implementations and hardware came to a point where deep learning was practical. Similarly, we've known about factoring word co-occurence matrices into Word embeddings for at least 20 years. But word embeddings only exploded in 2013 with Word2Vec. The breakthrough here was that the minibatch-based methods let you train a Wikipedia-scale embedding model on commodity hardware. It's hard for methods in a field to make progress if training on a small amount of data takes days or weeks. You're disincentivized to explore new methods. If you want progress, your stuff has to run in reasonable time on commodity hardware. Even Google's original search algorithm [initially ran on commodity hardware][34]. Efficiency is paramount to progress The reason deep learning research took off the way it did is because of improvements in [efficiency][35] as well as much better libraries and hardware support. Academic code is terrible Any amount of time you spend gridsearching Node2Vec on p and q is all put to better use gridsearching Deepwalk itself (on number of walks, length of walks, or word2vec hyperparameters). The problem is that people don't gridsearch over deepwalk because implementations are all terrible. I wrote the [Nodevectors library][36] to have a fast deepwalk implementation because it took 32 hours to embed a graph with a measly 150,000 nodes using the reference Node2Vec implementation (the same takes 3min with Nodevectors). It's no wonder people don't gridsearch on Deepwalk a gridsearch would take weeks with the terrible reference implementations. To give an example, in the original paper of [GraphSAGE][37] they their algorithm to DeepWalk with walk lengths of 5, which is horrid if you've ever hyperparameter tuned a deepwalk algorithm. From their paper: We did observe DeepWalk’s performance could improve with further training, and in some cases it could become competitive with the unsupervised GraphSAGE approaches (but not the supervised approaches) if we let it run for >1000× longer than the other approaches (in terms of wall clock time for prediction on the test set) I don't even think the GraphSAGE authors had bad intent -- deepwalk implementations are simply so awful that they're turned away from using it properly. It's like trying to do deep learning with 2002 deep learning libraries and hardware. Your architectures don't really matter One of the more important papers this year was [OpenAI's "Scaling laws"][38] paper, where the raw number of parameters in your model is the most predictive feature of overall performance. This was noted even in the original BERT paper and drives 2020's increase in absolutely massive language models. This is really just [Sutton' Bitter Lesson][39] in action: General methods that leverage computation are ultimately the most effective, and by a large margin Transformers might be [replacing convolution][40], too. As [Yannic Kilcher said][41], transformers are ruining everything. [They work on graphs][6], in fact it's one of the [recent approaches][42], and seems to be one of the more succesful [when benchmarked][1] Researchers seem to be putting so much effort into architecture, but it doesn't matter much in the end because you can approximate anything by stacking more layers. Efficiency wins are great -- but neural net architectures are just one way to achieve that, and by tremendously over-researching this area we're leaving a lot of huge gains elsewhere on the table. Current Graph Data Structure Implementations suck NetworkX is a bad library. I mean, it's good if you're working on tiny graphs for babies, but for anything serious it chokes and forces you to rewrite everything in... what library, really? At this point most people working on large graphs end up hand-rolling some data structure. This is tough because your computer's memory is a 1-dimensional array of 1's and 0's and a graph has no obvious 1-d mapping. This is even harder when we take updating the graph (adding/removing some nodes/edges) into account. Here's a few options: Disconnected networks of pointers NetworkX is the best example. Here, every node is an object with a list of pointers to other nodes (the node's edges). This layout is like a linked list. Linked lists are the [root of all performance evil][43]. Linked lists go completely against how modern computers are designed. Fetching things from memory is slow, and operating on memory is fast (by two orders of magnitude). Whenever you do anything in this layout, you make a roundtrip to RAM. It's slow by design, you can write this in Ruby or C or assembly and it'll be slow regardless, because memory fetches are slow in hardware. The main advantage of this layout is that adding a new node is O(1). So if you're maintaining a massive graph where adding and removing nodes happens as often as reading from the graph, it makes sense. Another advantage of this layout is that it "scales". Because everything is decoupled from each other you can put this data structure on a cluster. However, you're really creating a complex solution for a problem you created for yourself. Sparse Adjacency Matrix This layout great for read-only graphs. I use it as the backend in my [nodevectors][25] library, and many other library writers use the [Scipy CSR Matrix][44], you can see graph algorithms implemented on it [here][45]. The most popular layout for this use is the [CSR Format][46] where you have 3 arrays holding the graph. One for edge destinations, one for edge weights and an "index pointer" which says which edges come from which node. Because the CSR layout is simply 3 arrays, it scales on a single computer: a CSR matrix can be laid out on a disk instead of in-memory. You simply [memory map][47] the 3 arrays and use them on-disk from there. With modern NVMe drives random seeks aren't slow anymore, much faster than distributed network calls like you do when scaling the linked list-based graph. I haven't seen anyone actually implement this yet, but it's in the roadmap for my implementation at least. The problem with this representation is that adding a node or edge means rebuilding the whole data structure. Edgelist representations This representation is three arrays: one for the edge sources, one for the edge destinations, and one for edge weights. [DGL][48] uses this representation internally. This is a simple and compact layout which can be good for analysis. The problem compared to CSR Graphs is some seek operations are slower. Say you want all the edges for node #4243. You can't jump there without maintaining an index pointer array. So either you maintain sorted order and binary search your way there (O(log2n)) or unsorted order and linear search (O(n)). This data structure can also work on memory mapped disk array, and node append is fast on unsorted versions (it's slow in the sorted version). Global methods are a dead end Methods that work on the entire graph at once can't leverage computation, because they run out of RAM at a certain scale. So any method that want a chance of being the new standard need to be able to update piecemeal on parts of the graph. Sampling-based methods Sampling Efficiency will matter more in the future Edgewise local methods. The only algorithms I know of that do this are GloVe and GGVec, which they pass through an edge list and update embedding weights on each step. The problem with this approach is that it's hard to use them for higher-order methods. The advantage is that they easily scale even on one computer. Also, incrementally adding a new node is as simple as taking the existing embeddings, adding a new one, and doing another epoch over the data Random Walk sampling. This is used by deepwalk and its descendants, usually for node embeddings rather than GNN methods. This can be computationally expensive and make it hard to add new nodes. But this does scale, for instance [Instagram][49] use it to feed their recommendation system models Neighbourhood sampling. This is currently the most common one in GNNs, and can be low or higher order depending on the neighborhood size. It also scales well, though implementing efficiently can be challenging. It's currently used by [Pinterest][50]'s recommendation algorithms. Conclusion Here are a few interesting questions: What is the relation between graph types and methods? Consolidated benchmarking like OGB We're throwing random models at random benchmarks without understanding why or when they do better More fundamental research. Heree's one I'm curious about: can other representation types like [Poincarre Embeddings][51] effectively encode directed relationships? On the other hand, we should stop focusing on adding spicy new layers to test on the same tiny datasets. No one cares. [1]: https://arxiv.org/pdf/2003.00982.pdf [2]: https://arxiv.org/pdf/2002.11867.pdf [3]: https://arxiv.org/pdf/1812.08434.pdf [4]: https://arxiv.org/pdf/2005.00687.pdf [5]: https://en.wikipedia.org/wiki/Adjacency_matrix [6]: https://thegradient.pub/transformers-are-graph-neural-networks/ [7]: https://en.wikipedia.org/wiki/Word2vec [8]: https://nlp.stanford.edu/pubs/glove.pdf [9]: https://papers.nips.cc/paper/2014/file/feab05aa91085b7a8012516bc3533958-Paper.pdf [10]: https://en.wikipedia.org/wiki/Bag-of-words_model [11]: https://en.wikipedia.org/wiki/Co-occurrence [12]: https://www.singlelunch.com/2020/02/16/embeddings-from-the-ground-up/ [13]: https://www.singlelunch.com/2019/01/27/word-embeddings-from-the-ground-up/ [14]: https://nlpprogress.com/ [15]: http://socsci.uci.edu/~rfutrell/papers/hahn2019estimating.pdf [16]: https://en.wikipedia.org/wiki/Kolmogorov_complexity [17]: https://bair.berkeley.edu/blog/2020/12/20/lmmem/ [18]: https://en.wikipedia.org/wiki/Laplacian_matrix [19]: http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=1F03130B02DC485C78BF364266B6F0CA?doi=10.1.1.19.8100&rep=rep1&type=pdf [20]: https://en.wikipedia.org/wiki/Principalcomponentanalysis [21]: https://www.ijcai.org/Proceedings/2019/0594.pdf [22]: https://dl.acm.org/doi/10.1145/2806416.2806512 [23]: https://openreview.net/pdf?id=SyK00v5xx [24]: https://github.com/VHRanger/nodevectors/blob/master/examples/link%20prediction.ipynb [25]: https://github.com/VHRanger/nodevectors [26]: https://arxiv.org/pdf/1310.2636.pdf [27]: http://byowen.com/ [28]: https://arxiv.org/pdf/1807.03341.pdf [29]: https://www.youtube.com/watch?v=Kee4ch3miVA [30]: https://cs.stanford.edu/~jure/pubs/node2vec-kdd16.pdf [31]: https://arxiv.org/pdf/1403.6652.pdf [32]: https://arxiv.org/pdf/1911.11726.pdf [33]: https://en.wikipedia.org/wiki/AlexNet [34]: https://en.wikipedia.org/wiki/Googledatacenters#Original_hardware [35]: https://openai.com/blog/ai-and-efficiency/ [36]: https://www.singlelunch.com/2019/08/01/700x-faster-node2vec-models-fastest-random-walks-on-a-graph/ [37]: https://arxiv.org/pdf/1706.02216.pdf [38]: https://arxiv.org/pdf/2001.08361.pdf [39]: http://incompleteideas.net/IncIdeas/BitterLesson.html [40]: https://arxiv.org/abs/2010.11929 [41]: https://www.youtube.com/watch?v=TrdevFK_am4 [42]: https://arxiv.org/pdf/1710.10903.pdf [43]: https://www.youtube.com/watch?v=fHNmRkzxHWs [44]: https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csr_matrix.html [45]: https://docs.scipy.org/doc/scipy/reference/sparse.csgraph.html [46]: https://en.wikipedia.org/wiki/Sparsematrix#Compressedsparserow(CSR,CRSorYaleformat) [47]: https://en.wikipedia.org/wiki/Mmap [48]: https://github.com/dmlc/dgl [49]: https://ai.facebook.com/blog/powered-by-ai-instagrams-explore-recommender-system/ [50]: https://medium.com/pinterest-engineering/pinsage-a-new-graph-convolutional-neural-network-for-web-scale-recommender-systems-88795a107f48 [51]: https://arxiv.org/pdf/1705.08039.pdf

#reddit #lukewarm #graph

anishathalye•This week

[P] MIT Introduction to Data-Centric AI

#reddit #introduction #data-centric

[D] What are some good advanced platforms?

SemperZero•This week

[D] What are some good advanced platforms?

Hey. I'm 27 and I think I got most of the basics for ML. I'm very good at math, I understand statistics and probability quite deep, worked on research projects by myself, for which I had to build models on my own. Not really complex, but still requiring creativity and a good understanding of basic concepts. I will soon start a data science job at a FAANG company and I want to further improve my skills and use their resources to the fullest, but I'm not really sure where to go from here in terms of learning. Could you help me with some more advanced materials/forums for ML research/place with good papers/place with good articles? I'd also like to study the very best and see the way they code and explain advanced concepts (like Andrej Karpathy) where can I find them?? is there a Twitch for challenger level AI researchers streaming live processes? Or videos showing the entire project flow (how they do data visualizations, mining, choosing models, tuning, etc) like top digital artists show the highlights or the entire speed-up of their painting processes? Here's a list all of my projects to get a general idea of my level and where I'm at: calculating the distance between hundreds of 42.000 feature objects (containing categorical, strings, numbers, hashes, booleans as variables) and then clustering. with some vector processing and a neural network implemented from scratch in C some models like ARIMA (together with linear regression) combining a FFT with a neural network for a 42d wave classification T-SNE to split dataset into 2d grids -> Kullback–Leibler on grids for distance -> DBSCAN/KMEANS for clustering genetic algorithms for hyperparameter optimizations and reinforcement learning (neuro evolution) DBSCAN -> Levenberg-Marquardt for polynomial coefficients-> neural network predicting the coefficients based on different parameters playing with instance segmentation and some algorithms to synchronize a color and a depth camera simulations/statistics/probabilities for video games a lot of visualizations and data mining for patterns As you can see there is no LLM/ Generative AI/ Computer Vision stuff, which I would like to get into. I'm also not 100% sure what else would be nice to learn in general. I know most of the basic procedures for training, balancing datasets, avoid overfit, computing error plots, comparing models, etc and I'm familiar with most of math (not insanely advanced) used in ML. I didn't read many papers, but holy ... most of them are so unreadable and filled with pompous nonsense that 99% of the effort is de-obfuscating the bs and reading for so long just to figure out how the input is encoded, what's the output, and what's the model. Where can I find good, readable, structured papers which are actually on point? I'm from Eastern Europe and most of my learning has been done by my self after high school, the education quality is close to zero in the universities here and I never had any mentors at the jobs I worked. There's no research in this country, and getting to work on these projects was insanely hard, some of them being done in my free time or for free just to get experience... Fortunately after a lot of hard work I got into FAANG, and I hope things will be better here. Most of what I've learned has been from very fragmented places on the internet, and now I'm looking for centralized places and communities of top quality content. TL;DR: sorry for the long rambling. had to order my thoughts and figure what i actually want: Looking for top tier AI researchers showcasing their work processes, places with clear papers/articles, tips for someone who's no longer a very beginner, and other communities like this.

#reddit #some #good

looking for ML aficionado in London for great chats and maybe a startup

0.333

MLstartupLondon•This week

looking for ML aficionado in London for great chats and maybe a startup

TL;DR? Here's the gist: Me: 3 startups under my belt. Started as a programmer, then trainer, then entrepreneur, now CTO & Board member for a leading customer insight company part of large bank. Large system and infrastructure specialist. Extensive & practical experience in raising funds and successfully managing both startup and established businesses. Fascinated by the power of data. Can't imagine myself spending the rest of my life being a cog in the machine. You: Machine learning specialist, programmer, analyst, understands how to navigate and crunch large datasets, from BI to predictive analytics. Interested in implementing applications from fraud detection to margin improvements through better clustering regardless of industry. Fascinated by the power of data. Can't imagine himself spending the rest of his or her life being a cog in the machine. The startup: The core idea it to build platforms and systems around the progressively larger datasets held by various sized companies, helping them solve big issues - cost reduction, profitability and reducing risk. I’m an infrastructure and software specialist and have access to 1) systems, 2) datasets 3) extensive practical in certain industry segments, namely web-scale companies and tier 1 retailers. This project is in the very early planning stages. I'm looking forward to discuss the form it could take with like-minded individuals but with complementary skills sets, namely: predictive analytics & AI as it applies to machine learning on large datasets. Want more specifics ideas? I have plenty of these, but I’m sure you do to, so let’s meet face to face and discuss them. Ultimately the goal is to crystallize on a specific concept, develop together a minimum viable product and get the company bootstrapped or angel-funded (something I also have plenty of experience with), all via a lean startup model. My philosophy on startups: Startups built in one’s free time often fail because they drag on, ending up as little more than side projects you can’t quite get rid of (due to co-founder guilt, or perhaps the little money they bring in every month). The core idea for this project is based on lean, that is, to launch a minimum viable product as early as possible. Getting feedback. Measuring results (important!). Pivot if it’s not working. This helps tremendously in staying motivated, limits the dreaded paralyzing fear of failure, and more importantly, keep the time from inception to first client/funding to a minimum. If it sounds interesting please message me and we can exchange contact details! Worst that can happen is we have a great chat!

#reddit #looking #aficionado

[N] TheSequence Scope: When it comes to machine learning, size matters: Microsoft's DeepSpeed framework, which can train a model with up to a trillion parameters

Kseniase•This week

[N] TheSequence Scope: When it comes to machine learning, size matters: Microsoft's DeepSpeed framework, which can train a model with up to a trillion parameters

Hi there! Offering to your attention the latest edition of a weekly ML-newsletter that focusing on three things: impactful ML research papers, cool ML tech solutions, and ML use cases supported by investors. Please, see it below. Reddit is a new thing for me, and I've been struggling a bit with it, so please don't judge me too harsh for this promotion. This weekly digest is free and I hope you'd find the format convenient for you. Your feedback is very appreciated, and please feel free to sign up if you like it. 📝 Editorial The recent emergence of pre-trained language models and transformer architectures pushed the creation of larger and larger machine learning models. Google’s BERT presented attention mechanism and transformer architecture possibilities as the “next big thing” in ML, and the numbers seem surreal. OpenAI’s GPT-2 set a record by processing 1.5 billion parameters, followed by Microsoft’s Turing-NLG, which processed 17 billion parameters just to see the new GPT-3 processing an astonishing 175 billion parameters. To not feel complacent, just this week Microsoft announced a new release of its DeepSpeed framework (which powers Turing-NLG), which can train a model with up to a trillion parameters. That sounds insane but it really isn’t. What we are seeing is a consequence of several factors. First, computation power and parallelization techniques have evolved to a point where it is relatively easy to train machine learning models in large clusters of machines. Second and most importantly, in the current state of machine learning, larger models have regularly outperformed smaller and more specialized models. Knowledge reusability methods like transfer learning are still in very nascent stages. As a result, it’s really hard to build small models that can operate in uncertain environments. Furthermore, as models like GPT-3 and Turing-NLG have shown, there is some unexplainable magic that happens after models go past a certain size. Many of the immediate machine learning problems might be solved by scaling the current generation of neural network architectures. Plain and simple, when it comes to machine learning, size matters. We would love to hear your opinions about the debate between broader-larger vs. smaller and more specialized models. Leave a comment Now, to the most important developments in the AI industry this week 🔎 ML Research GPT-3 Falls Short in Machine Comprehension Proposed by researchers from a few major American universities, a 57-task test to measure models’ ability to reason poses challenges even for sophisticated models like GPT-3 ->read more in the original paper Better Text Summarization OpenAI published a paper showing a reinforcement learning with human feedback technique that can surpass supervised models ->read more on OpenAI blog Reinforcement Learning with Offline Datasets Researchers from the Berkeley AI Research (BAIR) Lab published a paper unveiling a method that uses offline datasets to improve reinforcement learning models->read more on BAIR blog 🤖 Cool AI Tech Releases New Version of DeepSpeed Microsoft open-sourced a new version of DeepSpeed, an open-source library for parallelizable training that can scale up to models with 1 trillion parameters->read more on Microsoft Research blog 💸 Money in AI AI-powered customer experience management platform Sprinklr has raised $200 million (kudos to our subscribers from Sprinklr 👏). Sprinklr's “AI listening processing” solution allows companies to get structured and meaningful sentiments and insights from unstructured customer data that comes from public conversations on different websites and social platforms. Xometry, an on-demand industrial parts marketplace, raises $75 million in Series E funding. The company provides a digital way of creating the right combination of buyers and manufacturers. Another example of AI implementation into matching two sides for a deal. Real estate tech company Orchard raises $69 million in its recent funding round. Orchard aims to digitize the whole real estate market, by developing a solution that combines machine learning and rapid human assistance to smooth the search, match the right deal, and simplify buying and selling relationships. Cybersecurity startup Pcysys raised $25 million in its funding round. Pcysys’ platform, which doesn’t require installation or network reconfiguration, uses algorithms to scan and “ethically” attack enterprise networks. Robotics farming company Iron Ox raised $20 million in a funding round. The system of farming robots is still semi-autonomous, the company’s goal is to become fully autonomous. Insurtech company Descartes Underwriting raised $18.5 million. The company applies AI and machine learning technologies to climate risk predicting and insurance underwriting. Legaltech startup ThoughtRiver raised $10 million in its Series A round. Its AI solution applied to contract pre-screening aims to boost operational efficiency. Medtech startup Skin Analytics raised $5.1 million in Series A funding. Skin Analytics has developed a clinically validated AI system that can identify not only the important skin cancers but also precancerous lesions that can be treated, as well as a range of lesions that are benign. Amazon, along with several government organizations and three other industry partners, helped fund the National Science Foundation, a high-priority AI research initiative. The amount of funding is not disclosed. The content of TheSequence is written by Jesus Rodriguez, one of the most-read contributors to KDNuggets and TDS. You can check his Medium here.

#reddit #thesequence #scope:

[D] Last Week in Medical AI: Top Research Papers/Models 🏅(September 21 - September 27, 2024)

aadityaura•This week

[D] Last Week in Medical AI: Top Research Papers/Models 🏅(September 21 - September 27, 2024)

Last Week in Medical AI: Top Research Papers\/Models 🏅$September 21 - September 27, 2024$ Medical AI Paper of the Week A Preliminary Study of o1 in Medicine: Are We Closer to an AI Doctor? This paper presents o1, a Large Language Model (LLM) evaluated across 37 medical datasets demonstrating superior performance in clinical understanding, reasoning, and multilinguality compared to GPT-4 and GPT-3.5. Medical LLM & Other Models: DREAMS: Python Framework for Medical LLMs A comprehensive deep learning framework for EEG data processing, model training, and report generation. SLaVA-CXR: A Small Language and Vision Assistant for Chest X-Ray Report Automation This paper introduces SLaVA-CXR, an innovative small-scale model designed for automating chest X-ray reports with high accuracy and efficiency. O1 in Medicine: AI Doctor Potential Genome Language Model : Opportunities & Challenge It highlights key gLM applications like functional constraint prediction, sequence design, and transfer learning, while discussing challenges in developing effective gLMs for complex genomes. Medical LLMs & Benchmarks: MEDICONFUSION: Probing Medical LLM Reliability This paper introduces MediConfusion, a challenging benchmark for probing the failure modes of multimodal large language models (MLLMs) in medical imaging. CHBench: Chinese LLM Health Evaluation This paper introduces CHBench, the first comprehensive Chinese health-related benchmark designed to evaluate large language models (LLMs) on their understanding of physical and mental health. LLMs for Mental Illness Evaluation PALLM: Evaluating Palliative Care LLMs Protein LMs: Scaling Necessity? Frameworks and Methodologies: Digital Twin for Oncology Operations Enhancing Guardrails for Healthcare AI InterMind: LLM-Powered Depression Assessment Conversational Health Agents: LLM Framework Medical LLM Applications: LLMs for Mental Health Severity Prediction Fine-tuning LLMs for Radiology Reports LLMs in Patient Education: Back Pain Boosting Healthcare LLMs with Retrieved Context Continuous Pretraining for Clinical LLMs AI in Healthcare Ethics: Confidence Intervals in Medical Imaging AI Generative AI Readiness for Clinical Use ... Check the full thread in detail: https://x.com/OpenlifesciAI/status/1840020394880667937 Thank you for reading! If you know of any interesting papers that were missed, feel free to share them in the comments. If you have insights or breakthroughs in Medical AI you'd like to share in next week's edition, connect with us on Twt/x: OpenlifesciAI

#reddit #last #week

[News] AAAI 2025 Workshop on AI for Music 🎶

Saysike_rightnow69•This week

[News] AAAI 2025 Workshop on AI for Music 🎶

Hi everyone! We’re hosting the first “AI for Music” workshop at AAAI on March 3, 2025. The workshop will explore how AI is transforming music creation, recognition, education, and more. Topics include AI-driven composition, sound design, legal and ethical challenges, and AI’s impact on musicians’ careers. Submissions (up to 6 pages) are welcome until November 22, 2024. Work in progress is encouraged! Workshop Summary This one-day workshop will explore the dynamic intersection of artificial intelligence and music. It explores how AI is transforming music creation, recognition, and education, ethical and legal implications, as well as business opportunities. We will investigate how AI is changing the music industry and education—from composition to performance, production, collaboration, and audience experience. Participants will gain insights into the technological challenges in music and how AI can enhance creativity, enabling musicians and producers to push the boundaries of their art. The workshop will cover topics such as AI-driven music composition, where algorithms generate melodies, harmonies, and even full orchestral arrangements. We will discuss how AI tools assist in sound design, remixing, and mastering, allowing for new sonic possibilities and efficiencies in music production. Additionally, we'll examine AI's impact on music education and the careers of musicians, exploring advanced learning tools and teaching methods. AI technologies are increasingly adopted in the music and entertainment industry. The workshop will also discuss the legal and ethical implications of AI in music, including questions of authorship, originality, and the evolving role of human artists in an increasingly automated world. This workshop is designed for AI researchers, musicians, producers, and educators interested in the current status and future of AI in music. Call for Papers Submissions should be a maximum of 6 pages. Work in progress is welcome. Authors are encouraged to include descriptions of their prototype implementations. Additionally, authors are encouraged to interact with workshop attendees by including posters or demonstrations at the end of the workshop. Conceptual designs without any evidence of practical implementation are discouraged. Topics of interest are (but not limited to) AI-Driven Music Composition and Generation AI in Music Practice and Performance AI-based Music Recognition and Transcription AI Applications in Sound Design AI-Generated Videos and Lyrics Based on Music Legal and Ethical Implications of AI in Music AI’s Impact on Musicians’ Careers and Education Business Opportunities of AI in Music Music Datasets and Data Analysis Important Dates Submission Deadline: November 22, 2024 Notification: December 9, 2024 Final Version Due: December 31, 2024 We hope to see you there! 🎶

#reddit #[news]#aaai

[P] I built an open SotA image tagging model to do what CLIP won't

fpgaminer•This week

[P] I built an open SotA image tagging model to do what CLIP won't

I'm a hobbyist ML researcher and finally, after a year of work, built a state of the art machine vision model from scratch. It's ViT-B/16 based, 448x448x3 input, 91M parameters, trained for 660M samples, with multi-label classification as the target task, on over 5000 unique tags. All the big foundation vision models today were trained on heavily filtered datasets, greatly limiting the concepts they can represent, in line with arbitrary sets of rules for what is deemed "wholesome" by leading tech companies. Everything from innocuous to spicy is on the chopping block of those filters. And because CLIP pervades the industry, from StableDiffusion to LLaVA, so does OpenAI's sensibilities. My goal was to build a vision model for tagging images, mainly for labelling images for SD finetunes, but which wasn't as heavily filtered and handicapped as CLIP/BLIP/LLaVA. Something more inclusive, diverse, and sex positive. Starting from the wonderful work of SmilingWolf (https://github.com/SmilingWolf/SW-CV-ModelZoo) and the Danbooru2021 dataset, I iterated for a year on the model, training, and manually labeling a thousand images to help the model generalize beyond the danbooru domain. I'm releasing the first version of this model, dubbed JoyTag, today: https://github.com/fpgaminer/joytag It achieves a mean F1 score of 0.578 across all of its over 5000 tags and across both the anime/manga styled images of the original danbooru dataset, but also photographs and other mediums thanks to the auxiliary training data I provided to it. It was quite the struggle getting to this point, and I probably spent more time and money than any sane person should have. I learned a lot about dealing with datasets as large as danbooru2021, training models at scale, and how to keep yourself awake all night so your 8xA100 rental doesn't crash and blow all your money. In my manual testing outside of even the validation set, the model has generalized well to unseen images, so I'm quite happy with the results thus far. There's plenty more work to do expanding its dataset to improve that F1 score further, and roundout its weak points. With inclusivity and diversity being a major goal of this project, I'm disappointed by some of its remaining limitations (as documented in the GitHub README). But I'm already busy manually tagging more images using my model-augmented workflow. I'm happy to answer questions about the project, the training procedure, anything. All the training parameters are documented on GitHub, but there are so many little details that were hard won over the year. Like that damned loss multiplier. Ugh. Github: https://github.com/fpgaminer/joytag Model download: https://huggingface.co/fancyfeast/joytag/tree/main Demo: https://huggingface.co/spaces/fancyfeast/joytag

#reddit #built #open

[Dataset][R] 19,762 Garbage Images for Building AI Recycling Solutions

Downtown_Bag8166•This week

[Dataset][R] 19,762 Garbage Images for Building AI Recycling Solutions

Hi ML community! I’m excited to share the Garbage Classification V2 Dataset, featuring 19,762 high-quality images of garbage categorized into 10 distinct classes (e.g., metal, plastic, clothes, and paper). Why this matters: Train AI models for automated waste sorting and recycling. Develop waste segregation apps or sustainability-focused tools. Create innovative computer vision projects for environmental impact. 🔗 Dataset Link: Garbage Classification V2 This dataset has been used in the research paper, "Managing Household Waste Through Transfer Learning," proving its utility in real-world applications. Looking forward to seeing how you can use it to promote sustainability!

#reddit #[dataset][r]#19,762

Saysike_rightnow69•This week

[News] AAAI 2025 Workshop on AI for Music 🎶

#reddit #[news]#aaai

How a Small Startup in Asia Secured a Contract with the US Department of Homeland Security

Royal_Rest8409•This week

How a Small Startup in Asia Secured a Contract with the US Department of Homeland Security

Uzair Javaid, a Ph.D. with a passion for data privacy, co-founded Betterdata to tackle one of AI's most pressing challenges: protecting privacy while enabling innovation. Recently, Betterdata secured a lucrative contract with the US Department of Homeland Security, 1 of only 4 companies worldwide to do so and the only one in Asia. Here's how he did it: The Story So what's your story? I grew up in Peshawar, Pakistan, excelling in coding despite studying electrical engineering. Inspired by my professors, I set my sights on studying abroad and eventually earned a Ph.D. scholarship at NUS Singapore, specializing in data security and privacy. During my research, I ethically hacked Ethereum and published 15 papers—three times the requirement. While wrapping up my Ph.D., I explored startup ideas and joined Entrepreneur First, where I met Kevin Yee. With his expertise in generative models and mine in privacy, we founded Betterdata. Now, nearly three years in, we’ve secured a major contract with the U.S. Department of Homeland Security—one of only four companies globally and the only one from Asia. The Startup In a nutshell, what does your startup do? Betterdata is a startup that uses AI and synthetic data generation to address two major challenges: data privacy and the scarcity of high-quality data for training AI models. By leveraging generative models and privacy-enhancing technologies, Betterdata enables businesses, such as banks, to use customer data without breaching privacy regulations. The platform trains AI on real data, learns its patterns, and generates synthetic data that mimics the real thing without containing any personal or sensitive information. This allows companies to innovate and develop AI solutions safely and ethically, all while tackling the growing need for diverse, high-quality data in AI development. How did you conduct ideation and validation for your startup? The initial idea for Betterdata came from personal experience. During my Ph.D., I ethically hacked Ethereum’s blockchain, exposing flaws in encryption-based data sharing. This led me to explore AI-driven deep synthesis technology—similar to deepfakes but for structured data privacy. With GDPR impacting 28M+ businesses, I saw a massive opportunity to help enterprises securely share data while staying compliant. To validate the idea, I spoke to 50 potential customers—a number that strikes the right balance. Some say 100, but that’s impractical for early-stage founders. At 50, patterns emerge: if 3 out of 10 mention the same problem, and this repeats across 50, you have 10–15 strong signals, making it a solid foundation for an MVP. Instead of outbound sales, which I dislike, we used three key methods: Account-Based Marketing (ABM)—targeting technically savvy users with solutions for niche problems, like scaling synthetic data for banks. Targeted Content Marketing—regular customer conversations shaped our thought leadership and outreach. Raising Awareness Through Partnerships—collaborating with NUS, Singapore’s PDPC, and Plug and Play to build credibility and educate the market. These strategies attracted serious customers willing to pay, guiding Betterdata’s product development and market fit. How did you approach the initial building and ongoing product development? In the early stages, we built synthetic data generation algorithms and a basic UI for proof-of-concept, using open-source datasets to engage with banks. We quickly learned that banks wouldn't share actual customer data due to privacy concerns, so we had to conduct on-site installations and gather feedback to refine our MVP. Through continuous consultation with customers, we discovered real enterprise data posed challenges, such as missing values, which led us to adapt our prototype accordingly. This iterative approach of listening to customer feedback and observing their usage allowed us to improve our product, enhance UX, and address unmet needs while building trust and loyalty. Working closely with our customers also gives us a data advantage. Our solution’s effectiveness depends on customer data, which we can't fully access, but bridging this knowledge gap gives us a competitive edge. The more customers we test on, the more our algorithms adapt to diverse use cases, making it harder for competitors to replicate our insights. My approach to iteration is simple: focus solely on customer feedback and ignore external noise like trends or advice. The key question for the team is: which customer is asking for this feature or solution? As long as there's a clear answer, we move forward. External influences, such as AI hype, often bring more confusion than clarity. True long-term success comes from solving real customer problems, not chasing trends. Customers may not always know exactly what they want, but they understand their problems. Our job is to identify these problems and solve them in innovative ways. While customers may suggest specific features, we stay focused on solving the core issue rather than just fulfilling their exact requests. The idea aligns with the quote often attributed to Henry Ford: "If I asked people what they wanted, they would have said faster horses." The key is understanding their problems, not just taking requests at face value. How do you assess product-market fit? To assess product-market fit, we track two key metrics: Customers' Willingness to Pay: We measure both the quantity and quality of meetings with potential customers. A high number of meetings with key decision-makers signals genuine interest. At Betterdata, we focused on getting meetings with people in banks and large enterprises to gauge our product's resonance with the target market. How Much Customers Are Willing to Pay: We monitor the price customers are willing to pay, especially in the early stages. For us, large enterprises, like banks, were willing to pay a premium for our synthetic data platform due to the growing need for privacy tech. This feedback guided our product refinement and scaling strategy. By focusing on these metrics, we refined our product and positioned it for scaling. What is your business model? We employ a structured, phase-driven approach for out business model, as a B2B startup. I initially struggled with focusing on the core value proposition in sales, often becoming overly educational. Eventually, we developed a product roadmap with models that allowed us to match customer needs to specific offerings and justify our pricing. Our pricing structure includes project-based pilots and annual contracts for successful deployments. At Betterdata, our customer engagement unfolds across three phases: Phase 1: Trial and Benchmarking \- We start with outreach and use open-source datasets to showcase results, offering customers a trial period to evaluate the solution. Phase 2: Pilot or PoC \- After positive trial results, we conduct a PoC or pilot using the customer’s private data, with the understanding that successful pilots lead to an annual contract. Phase 3: Multi-Year Contracts \- Following a successful pilot, we transition to long-term commercial contracts, focusing on multi-year agreements to ensure stability and ongoing partnerships. How do you do marketing for your brand? We take a non-conventional approach to marketing, focusing on answering one key question: Which customers are willing to pay, and how much? This drives our messaging to show how our solution meets their needs. Our strategy centers around two main components: Building a network of lead magnets \- These are influential figures like senior advisors, thought leaders, and strategic partners. Engaging with institutions like IMDA, SUTD, and investors like Plug and Play helps us gain access to the right people and foster warm introductions, which shorten our sales cycle and ensure we’re reaching the right audience. Thought leadership \- We build our brand through customer traction, technology evidence, and regulatory guidelines. This helps us establish credibility in the market and position ourselves as trusted leaders in our field. This holistic approach has enabled us to navigate diverse market conditions in Asia and grow our B2B relationships. By focusing on these areas, we drive business growth and establish strong trust with stakeholders. What's your advice for fundraising? Here are my key takeaways for other founders when it comes to fundraising: Fundraise When You Don’t Need To We closed our seed round in April 2023, a time when we weren't actively raising. Founders should always be in fundraising mode, even when they're not immediately in need of capital. Don’t wait until you have only a few months of runway left. Keep the pipeline open and build relationships. When the timing is right, execution becomes much easier. For us, our investment came through a combination of referrals and inbound interest. Even our lead investor initially rejected us, but after re-engaging, things eventually fell into place. It’s crucial to stay humble, treat everyone with respect, and maintain those relationships for when the time is right. Be Mindful of How You Present Information When fundraising, how you present information matters a lot. We created a comprehensive, easily digestible investment memo, hosted on Notion, which included everything an investor might need—problem, solution, market, team, risks, opportunities, and data. The goal was for investors to be able to get the full picture within 30 minutes without chasing down extra details. We also focused on making our financial model clear and meaningful, even though a 5-year forecast might be overkill at the seed stage. The key was clarity and conciseness, and making it as easy as possible for investors to understand the opportunity. I learned that brevity and simplicity are often the best ways to make a memorable impact. For the pitch itself, keep it simple and focus on 4 things: problem, solution, team, and market. If you can summarize each of these clearly and concisely, you’ll have a compelling pitch. Later on, you can expand into market segments, traction, and other metrics, but for seed-stage, focus on those four areas, and make sure you’re strong in at least three of them. If you do, you'll have a compelling case. How do you run things day-to-day? i.e what's your operational workflow and team structure? Here's an overview of our team structure and process: Internally: Our team is divided into two main areas: backend (internal team) and frontend (market-facing team). There's no formal hierarchy within the backend team. We all operate as equals, defining our goals based on what needs to be developed, assigning tasks, and meeting weekly to share updates and review progress. The focus is on full ownership of tasks and accountability for getting things done. I also contribute to product development, identifying challenges and clearing obstacles to help the team move forward. Backend Team: We approach tasks based on the scope defined by customers, with no blame or hierarchy. It's like a sports team—sometimes someone excels, and other times they struggle, but we support each other and move forward together. Everyone has the creative freedom to work in the way that suits them best, but we establish regular meetings and check-ins to ensure alignment and progress. Frontend Team: For the market-facing side, we implement a hierarchy because the market expects this structure. If I present myself as "CEO," it signals authority and credibility. This distinction affects how we communicate with the market and how we build our brand. The frontend team is split into four main areas: Business Product (Software Engineering) Machine Learning Engineering R&D The C-suite sits at the top, followed by team leads, and then the executors. We distill market expectations into actionable tasks, ensuring that everyone is clear on their role and responsibilities. Process: We start by receiving market expectations and defining tasks based on them. Tasks are assigned to relevant teams, and execution happens with no communication barriers between team members. This ensures seamless collaboration and focused execution. The main goal is always effectiveness—getting things done efficiently while maintaining flexibility in how individuals approach their work. In both teams, there's an emphasis on accountability, collaboration, and clear communication, but the structure varies according to the nature of the work and external expectations.

#reddit #small #startup

Royal_Rest8409•This week

How a Small Startup in Asia Secured a Contract with the US Department of Homeland Security

#reddit #small #startup

AI SaaS: A website to fine-tune LLM model according to your requirements

Dangerous_Ferret3362•This week

AI SaaS: A website to fine-tune LLM model according to your requirements

Hey fellow entrepreneurs and AI enthusiasts! I'm exploring a business idea and would love your thoughts and feedback. The concept is a SaaS platform that allows users to easily fine-tune large language models (LLMs) on their own datasets without needing deep technical expertise. Here's the gist: The Problem: Many businesses and researchers want to leverage LLMs for specific use cases, but fine-tuning these models requires significant technical knowledge and resources. The Solution: A user-friendly web platform where users can: Choose from popular LLM architectures Upload their own dataset or input text Configure fine-tuning parameters through an intuitive interface Automatically fine-tune the model on our GPU infrastructure Download the fine-tuned model or use it via API Key Features: No coding required Scalable cloud infrastructure Support for various fine-tuning techniques (prompt tuning, adapter tuning, full fine-tuning) Job monitoring and results visualization API access for integrated use in applications Target Market: Researchers without extensive ML engineering resources Startups building AI-powered products Enterprises looking to customize LLMs for internal use Monetization: Tiered subscription model based on usage (compute time, model size, etc.) + potential enterprise contracts for high-volume users. I'd really appreciate your thoughts on: Is this solving a real pain point? Would you use a service like this? Why or why not? What features would make this a must-have for you? Any foreseeable obstacles or considerations I'm missing? Suggestions for go-to-market strategy? Thank you!

#reddit #saas:#website

Interview with founder of ReadyPlayerMe (raised $70M+ from a16z)

Due_Cryptographer461•This week

Interview with founder of ReadyPlayerMe (raised $70M+ from a16z)

Thanks to everyone who replied to my previous post with the questions you had for Rainer, I added some of them into this interview. I’m Nikita of Databas3 , and that’s my first interview in a series where I’m learning more about the journey of the best tech and web3 founders. Would appreciate your feedback and suggestions for the next guest! Nikita: Let’s begin with a brief introduction. Can you share a bit about yourself and how the business started? Rainer: I’m Rainer, the CTO of ReadyPlayerMe. Our journey began in 2013 with four co-founders. Over the years, our focus has shifted mainly around our product’s evolution, but our core idea always revolved around virtual actors or virtual people. Our initial venture was into hardware. We created the first full-body scanner in the Nordics, a significant step in photogrammetry. This led us to develop the Luna Scanner, a three-meter tall structure designed to capture facial features and likenesses. When Facebook acquired Oculus in 2014, we foresaw the potential of VR and virtual worlds, especially in social experiences. Nikita: Interesting. How did you move on from there? Rainer: Recognizing the limitations of hardware, we transitioned into software. Our early scanner designs had limitations in scalability. For example, our three-meter tall scanner wasn’t a feasible solution for scanning millions of people. So, we leveraged the datasets from our initial projects and designed a mobile version, making facial scanning as easy as using your phone. Around 2015, this was a new territory, as facial scanning wasn’t a mainstream application. Nikita: What were the early applications of these scanned models? Rainer: In the beginning, we focused on 3D printed figurines from full-body scans. However, as we shifted to facial scanning, we licensed our technology to gaming companies, collaborating with giants like Wargaming and Tencent. We even ventured into virtual fittings with H&M. Each collaboration was custom-tailored, blending our technology with their systems. This model made us cash flow positive. Nikita: So this was the beginning of your foray into the gaming industry? Rainer: Precisely. The demand from gaming companies was substantial. As we built custom solutions for these enterprises, we saw a bigger potential. While our cash flow was positive, we realized the challenge of scaling through exclusive enterprise deals. We envisioned our avatar creation tech reaching indie games and beyond. Nikita: And that led to the birth of ReadyPlayerMe? Rainer: Exactly. Once we understood our market direction, we quickly developed the first iteration of ReadyPlayerMe as a web-based experience, emphasizing easy integration for game developers. The initial version was a character builder, allowing users to personalize their avatars, which many adopted for their social media profiles. Our goal was to create avatars that users could connect with and use across various platforms. Instead of licensing our technology, we offered it for free to everyone. As ReadyPlayerMe gained traction, especially in VR applications, we secured funding to further our mission. Nikita: Your growth seems swift and organic. Were there any challenges? Rainer: Our focus on easy integration significantly fueled our adoption. Pairing that with personalized avatars resonated well with our audience. But like any venture, we’ve faced our share of challenges and have always aimed to evolve and better our offerings. The rapid growth in Web3 projects and virtual worlds made personalization and customization more important. With the NFT boom, you could add utility by allowing access to selected collections. This played into web-based games and metaverse applications. The shift towards Web3 and personalization provided a significant tailwind for us. Many used our characters as profile pictures on social media. Nikita: I’ve heard from other founders that a16z really values viral marketing. Was this one reason they wanted to invest in your project? How was the process with them? Rainer: When a16z reached out, it felt like a natural fit. We wanted investors who understood the gaming space. Our main market is Web3, but we’re exploring the top games market. Their expertise in gaming was invaluable. They’ve been very supportive throughout. We were fortunate to be on their radar. Nikita: So your early growth and organic traction played a role in attracting investors? Rainer: Definitely. Early product growth and the potential future trajectory were essential in our discussions. Nikita: As the CTO, you must have faced challenges. Can you speak about the tech side and its evolution? Rainer: The early version of our platform was built by in-house engineers. As we grew, we had to adapt to increasing complexities and ensure we had the right team to execute our vision. My role often shifted between product management and tech, depending on the need. Nikita: It sounds like the startup environment remains strong within your company. Rainer: Absolutely. We’re all committed, hands-on, and working towards building the best product. Nikita: You mentioned the team earlier. How many people are in your team now? Rainer: We have 70 people, with about half in product and engineering. Nikita: And did you hire the tech team? Rainer: We brought on a head of engineering at the beginning of this year. He’s been instrumental in scaling the engineering organization, from increasing the headcount to refining engineering processes. We’ve recently reorganized into domain-specific teams. As the team grows, regular reorganization ensures we focus on delivering specific customer value. Every stage requires attention to the team’s composition to ensure efficient delivery. Nikita: Any advice for founders just starting with their first startup? Rainer: Focus on customer value, no matter how niche it might seem initially. Begin with a specific problem and solution, then expand from there. You don’t need a massive project right away. Begin small, prove the concept, and scale from there. Nikita: You’ve mentioned your love for books and podcasts. Any recommendations? Rainer: For startups, “High Growth Handbook” and “Lean Startup” are must-reads. “Working Backwards” offers insights into Amazon’s customer-centric approach. For podcasts, I listen to “Rework,” “Lenny’s Podcast,” and “Huberman Lab.” Nikita: All of us have some side project ideas from time to time. How do you handle these when managing a big project? Rainer: Over the years, I’ve built various side projects. Some are small applications to solve immediate problems, like a menu bar app for AirPods which made it to No. 1 on Product Hunt, and was nominated for Golden Kitty Award. I sometimes delve into 3D and AI, merging them for technical demos. I keep a list of ideas and pick from them as the urge arises. Nikita: Any final thoughts or advice? Rainer: As you scale, do so with clarity. Avoid scaling just for external appeal. Always hire when there’s genuine need, not just for the sake of expansion. It helps in staying lean and focused.

#reddit #interview #founder

Dangerous_Ferret3362•This week

AI SaaS: A website to fine-tune LLM model according to your requirements

#reddit #saas:#website

Dangerous_Ferret3362•This week

AI SaaS: A website to fine-tune LLM model according to your requirements

#reddit #saas:#website

I built an OCR powered by Mistral AI that extracts text, tables, formulas from docs (20+ languages & JSON output!)

hhe_kkm•This week

I built an OCR powered by Mistral AI that extracts text, tables, formulas from docs (20+ languages & JSON output!)

Hi everyone 👋 Most OCR tools struggle with complex documents—crumbling tables, garbled formulas, or unstructured text. Need clean data for RAG or apps? Good luck. So I built Mistral OCR (https://www.mistralocr.app/) using Mistral AI’s document understanding models. It doesn’t just scan—it understands the document’s structure, and extracts: ✅ Text (plain/formatted) ✅ Tables (pixel-perfect JSON with headers 🧮) ✅ Math formulas (LaTeX-ready via Mistral’s ML pipeline) ✅ Images (preserved or extracted) Why Mistral AI? Their models nail context-aware parsing—unlike rigid OCRs, Mistral’s tech handles: Cursed PDFs(scanned/watermarked/warped text) Mixed layouts (research papers with tables + formulas) 20+ languages (English, Japanese, Mandarin, Spanish...) Structured JSON output (directly feeds into RAG/APIs) See examples → https://www.mistralocr.app/ Why build this? I needed an OCR that could extract RAG-ready data without regex nightmares. Mistral AI’s models finally made this possible—they preserve relationships between text, tables, and formulas, something traditional OCRs butcher. Who’s using it? Devs automating document workflows Researchers digitizing datasets from papers Teams processing multilingual forms/contracts Anyone frustrated by copying tables from PDFs Challenge me: Send your worst documents (scanned receipts? handwritten tables?) and I’ll run them through Mistral OCR live. Try it here → https://www.mistralocr.app/ Let me know what you think! 🙏 Let me know if bugs🐛！🙏

Datasets

Getting datasets to train AI to find a business opportunity and develop growth strategy

[R] Apple’s Machine Learning Researchers Have Developed A No-Code AI Platform Called ‘Trinity’ For Complex Spatial Datasets

AI for Business Consultancy

AI for Business Consultancy

My Manager Thinks ML Projects Takes 5 Minutes 🤦‍♀️

MIT Introduction to Data-Centric AI

How do I get a job in this job market? How do I stand out from the crowd?

How do byte-level language models work?

𝗠𝗮𝘀𝘁𝗲𝗿 𝗟𝗶𝗻𝗲𝗮𝗿 𝗥𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝗼𝗻: 𝗔 𝗠𝘂𝘀𝘁-𝗞𝗻𝗼𝘄 𝗳𝗼𝗿 𝗕𝗲𝗴𝗶𝗻𝗻𝗲𝗿𝘀

[Help Needed] How to tackle AI for DNS Security in a Hackathon for Beginner.

Implementing AI in real world problems

[D] Why I'm Lukewarm on Graph Neural Networks

[P] MIT Introduction to Data-Centric AI

[D] What are some good advanced platforms?

looking for ML aficionado in London for great chats and maybe a startup

[N] TheSequence Scope: When it comes to machine learning, size matters: Microsoft's DeepSpeed framework, which can train a model with up to a trillion parameters

[D] Last Week in Medical AI: Top Research Papers/Models 🏅(September 21 - September 27, 2024)

[News] AAAI 2025 Workshop on AI for Music 🎶

[P] I built an open SotA image tagging model to do what CLIP won't

[Dataset][R] 19,762 Garbage Images for Building AI Recycling Solutions

[News] AAAI 2025 Workshop on AI for Music 🎶

How a Small Startup in Asia Secured a Contract with the US Department of Homeland Security

How a Small Startup in Asia Secured a Contract with the US Department of Homeland Security

AI SaaS: A website to fine-tune LLM model according to your requirements

Interview with founder of ReadyPlayerMe (raised $70M+ from a16z)

AI SaaS: A website to fine-tune LLM model according to your requirements

AI SaaS: A website to fine-tune LLM model according to your requirements

I built an OCR powered by Mistral AI that extracts text, tables, formulas from docs (20+ languages & JSON output!)

GenAI_Agents

AITreasureBox

h2o-llmstudio

aima-python

RD-Agent

prompt-injection-defenses

awesome-ai-in-finance

OpenAI-CLIP

How-to-learn-Deep-Learning

business-document-processing

airoboros

dcai-lab

Awesome-Ai-Tools

ai-learning-roadmap

coursera-practical-data-science-specialization

Best Programming Language For AI in 2024 | Intellipaat #Shorts #AI #Python

What is generative AI and how does it work? – The Turing Lectures with Mirella Lapata

AI-basics