Jump to content

Machine Learning


Captainant

Recommended Posts

There's some wild new machine learning techniques out there these days - reinforcement learning lets robots teach themselves how to do a task without training data, unsupervised learning alerts heavy machinery operators know when something is fucky in a system, and NLP is unlocking the data within unstructured text and giving us a mechanism to rapidly crunch through thousands of pages of documents.

Newest to the fold for me, and what I've been getting my mind blown by: neural graphics primitives (short video of what that actually means lol). It's effectively a new mechanism to encode spatial and visual data so that rather than storing the polygons and geometry, it's training a neural net on how to generate that data on the fly based on a small number of input 2d images. The results are astounding. I'm working on getting it running on my local desktop computer but alas my puny 1080ti doesn't have tensor cores so it runs slow as hell.

The real innovation of this technique is that it's using NV's tensor cores to calculate the intersection of rays projected from your input images to create a fully 3D representation, and then training an ML model to infer the rest of the 3D space. It's some crazy higher order math - the originating paper for these concepts were published just a couple years ago, and they've already developed new techniques to operationalize and optimize the math. For context: these models fully train in two seconds using these new techniques and hardware. It's wild.

 

Any other ML nerds on surly? Also, does anyone have an RTX 3080 I can buy? lol

  • Hook 'Em 1
Link to comment
Share on other sites

  • 4 weeks later...
On 2/23/2022 at 8:37 AM, Captainant said:

There's some wild new machine learning techniques out there these days - reinforcement learning lets robots teach themselves how to do a task without training data, unsupervised learning alerts heavy machinery operators know when something is fucky in a system, and NLP is unlocking the data within unstructured text and giving us a mechanism to rapidly crunch through thousands of pages of documents.

Newest to the fold for me, and what I've been getting my mind blown by: neural graphics primitives (short video of what that actually means lol). It's effectively a new mechanism to encode spatial and visual data so that rather than storing the polygons and geometry, it's training a neural net on how to generate that data on the fly based on a small number of input 2d images. The results are astounding. I'm working on getting it running on my local desktop computer but alas my puny 1080ti doesn't have tensor cores so it runs slow as hell.

The real innovation of this technique is that it's using NV's tensor cores to calculate the intersection of rays projected from your input images to create a fully 3D representation, and then training an ML model to infer the rest of the 3D space. It's some crazy higher order math - the originating paper for these concepts were published just a couple years ago, and they've already developed new techniques to operationalize and optimize the math. For context: these models fully train in two seconds using these new techniques and hardware. It's wild.

 

Any other ML nerds on surly? Also, does anyone have an RTX 3080 I can buy? lol

But what are the enterprise use cases?

Seriously though, this is a cool thread and post-- wish it was in a different forum with more eyes and traffic instead of buried in nerdz because I guarantee this place has at least a half a dozen ML/AI/Data Scientist whizs.

Link to comment
Share on other sites

On 2/23/2022 at 8:37 AM, Captainant said:

There's some wild new machine learning techniques out there these days - reinforcement learning lets robots teach themselves how to do a task without training data, unsupervised learning alerts heavy machinery operators know when something is fucky in a system, and NLP is unlocking the data within unstructured text and giving us a mechanism to rapidly crunch through thousands of pages of documents.

Newest to the fold for me, and what I've been getting my mind blown by: neural graphics primitives (short video of what that actually means lol). It's effectively a new mechanism to encode spatial and visual data so that rather than storing the polygons and geometry, it's training a neural net on how to generate that data on the fly based on a small number of input 2d images. The results are astounding. I'm working on getting it running on my local desktop computer but alas my puny 1080ti doesn't have tensor cores so it runs slow as hell.

The real innovation of this technique is that it's using NV's tensor cores to calculate the intersection of rays projected from your input images to create a fully 3D representation, and then training an ML model to infer the rest of the 3D space. It's some crazy higher order math - the originating paper for these concepts were published just a couple years ago, and they've already developed new techniques to operationalize and optimize the math. For context: these models fully train in two seconds using these new techniques and hardware. It's wild.

 

Any other ML nerds on surly? Also, does anyone have an RTX 3080 I can buy? lol

My first thought because I'm old and stupid was "this seems like something PTC would be trying to bring to market" but I clicked your link and it looks like a labs/sub of nVidia. Which should be no surprise as they have quietly been one of the greatest underrated American companies the past decade and have an amazing CEO.

Also, institutional investors must be savvy to the use case potential I was half-way joking about:

Quote

Markets: Investors were feeling risky yesterday, piling into tech stocks and semiconductors in particular. Nvidia, the seventh-largest S&P 500 company by market cap, soared on expectations that it’ll play a pivotal role in the development of AI.

https://www.cnbc.com/2022/03/24/nvidia-and-intel-lead-rally-in-semiconductor-stocks-.html

Edited by TurkeyChew
Link to comment
Share on other sites

10 minutes ago, TurkeyChew said:

But what are the enterprise use cases?

Seriously though, this is a cool thread and post-- wish it was in a different forum with more eyes and traffic instead of buried in nerdz because I guarantee this place has at least a half a dozen ML/AI/Data Scientist whizs.

For this specific technique, imagine being able to take a series of flat ultrasound images from different angles and reconstruct the actual geometry of that object - either a tumor, or an organ, or even a baby's face. All non-invasively and with already existing tools, we're just using the data more effectively.

Enterprise specific, I have a customer that runs a huge fleet of vessels for ocean floor mapping. Imagine if instead of sending down an ROV, they could simply aggregate enough sonar depth data from enough points of reference to get the same result.

Neural nets in general are pretty incredible technology. We understand the mechanics of how they work, but are still learning why they work. 3brown1blue has an excellent explainer on them that's 90 minutes long in total.

 

 

  • Hook 'Em 1
Link to comment
Share on other sites

9 minutes ago, thunderlounge said:

So take some 2D images, do math, make a 3D hologram?

 

Interesting. How long before Hooli buys that one too?

Sony? samsung?

I mean, the method mentioned in my OP runs on a desktop computer lol. A 3090 trains the model in minutes or even seconds. The thing to buy/sell is the data that you use to train models or a trained model, not the algorithm or framework. 

"Data is the new oil" has been repeated constantly for the last 5 years. Machine learning is how we refine it into a usable product. 

But yeah, Nvidia is the head and shoulders leader for this stuff. They're creating hardware acceleration for extremely complex higher order matrix operations, and they somewhat shape the academic focus as they create new acceleration techniques. All this latest stuff is the beneficiary of ray tracing and tensor processing cores for cleaning data and training models

  • Hook 'Em 1
Link to comment
Share on other sites

Sure, but my light-hearted reply was meant more for the tech’s future beyond just that. 
 

Look what happened in a couple short years, and give it 10 more. 
 

I don’t think it’s something that will have a broad use today, but give some nut job engineers some time and some smoke and they’ll get there. 
 

 

Link to comment
Share on other sites

On 3/25/2022 at 3:15 PM, Captainant said:

I mean, the method mentioned in my OP runs on a desktop computer lol. A 3090 trains the model in minutes or even seconds. The thing to buy/sell is the data that you use to train models or a trained model, not the algorithm or framework. 

"Data is the new oil" has been repeated constantly for the last 5 years. Machine learning is how we refine it into a usable product. 

But yeah, Nvidia is the head and shoulders leader for this stuff. They're creating hardware acceleration for extremely complex higher order matrix operations, and they somewhat shape the academic focus as they create new acceleration techniques. All this latest stuff is the beneficiary of ray tracing and tensor processing cores for cleaning data and training models

I just read this fascinating interview with nVidia CEO:

Quote

 

In no time in history have humans have the ability to produce the single most valuable commodity the world’s ever known, which is intelligence. We now have a structure of a model, a structure of a computer science program called a deep neural network, that has the ability to scale up quite tremendously. It’s doubling every six months, I mean, this is not your Moore’s Law where it’s doubling every two years, it’s doubling every six months. The rate of doubling is incredible, the compounded effect of that on computing is incredible.

The results of the capabilities of these neural networks and the software, another way of saying it, the software that is being created by computers is expanding and growing and achieving spectacular things at incredible rates. Our company is building the computers necessary to continue to advance that journey. I think that what companies are going to come to realize is that what they’re really all about is producing intelligence, and that’s what Nvidia’s really all about is producing intelligence. Some part of every company will automate the production of their intelligence, they’ll codify the production of their intelligence, which is one of the reasons why I believe every company will be an AI company, every company will produce intelligence at some level, all of that AI will be augmenting humans with humans in the loop, and the rate of that progress is accelerating and compounding at 2x every six months.

what is intelligence? Intelligence is the ability to recognize patterns, recognize relationships, reason about it and make a prediction or plan an action. That’s what intelligence is. It has nothing to do with general intelligence, intelligence is just solving problems. We now have the ability to write software, we now have the ability to partner with computers to write software, that can solve many types of intelligence, make many types of predictions at scales and at levels that no humans can.

For example, we know that there are a trillion things on the Internet and the number things on the Internet is large and expanding incredibly fast and yet we have this little tiny personal computer called a phone, how do we possibly figure out of the trillion things in the internet what we want to see on our little tiny phone? Well, there needs to be a filter in between, what people call the personalized internet, but basically an AI, a recommender system. A recommender that figures out based on the nature of the content, the characteristics of the content, the features of the content, based on your implicit and your explicit and implicit preferences, find a way through all of that to predict what you would like to see. I mean, that’s a miracle! That’s really quite a miracle to be able to do that at scale for everything from movies and books and music and news and videos and you name it, products and things like that. To be able to predict what Ben would want to see, predict what you would want to click on, predict what is useful to you. I’m talking about things that are consumer oriented stuff, but in the future it’ll be predict what is the best financial strategy for you, predict what is the best medical therapy for you, predict what is the best health regimen for you, what’s the best vacation plan for you. All of these things are going to be possible with AI.

 

Edited by TurkeyChew
Link to comment
Share on other sites

  • 2 weeks later...

In general, if you want to get your mind blown I'd recommend checking out the youtube channel Two Minute Papers - he posts a new 5ish minute video every so often showing off a novel implementation of an ML model. This is a pretty neat one that combines reinforcement learning with adversarial networks

 

And on a personal note - I just scored an open box RTX 3080 with 12GB of VRAM from my local MicroCenter so I'm looking forward to finally getting to kick the tires on some of this stuff

Edited by Captainant
  • Hook 'Em 1
Link to comment
Share on other sites

Ok now I'm just pandering, but this is pretty fucking cool. I took this UT tower flyaround video (in spoiler)

Spoiler

Did the above process against the video file (720p 30fps source file), trained it for about 120 seconds, mapped a camera flight through the environment and rendered it back out to a short 1080p video as seen below @thunderlounge @TwiceHorn:

Edited by Captainant
  • Hook 'Em 1
  • Like 1
Link to comment
Share on other sites

On 4/13/2022 at 7:58 AM, Captainant said:

Ok now I'm just pandering, but this is pretty fucking cool. I took this UT tower flyaround video (in spoiler)

  Reveal hidden contents

Did the above process against the video file (720p 30fps source file), trained it for about 120 seconds, mapped a camera flight through the environment and rendered it back out to a short 1080p video as seen below @thunderlounge @TwiceHorn:

 

spacer.png

Link to comment
Share on other sites

  • 7 months later...

Bit of a necrobump, I found a nice project package to install Stable Diffusion to run locally! If you've got a RTX 3060 or better and ~20GB of storage to hold all the neural nets, weights, biases, and other supporting data, it's a hell of a thing to have running locally on your desktop computer. Has a super nice locally hosted web browser UI that makes it easy to use too once you're past the first time bootstrapping

I'm still tinkering and haven't gotten any amazing results yet, but it's shocking good at making a "good enough" visual based on pretty minimal input and parameters. 

Link to comment
Share on other sites

  • 1 month later...
Just now, HamsterHookah said:

I'd be curious to hear your experience and point of view on IDP (as it seems to be the most unlike what chatpGPT does as an ingestion tool to extract data from unstructured data sources) because I've been following it for a while and the entire market seems to be crazy crowded and I can't really make heads or tails of who the best-of-breed engines are (AWS textract? Google Tesseract? VC-funded private start-ups that are 2 years old? Legacy software like Abbyy or Kofax?) and it just feels like a chaotic landscape where even analysts can't figure it out. Maybe another thread.

Moving the convo to the nerdz thread, because yeah lol.

 

AWS Textract isn't IDP, it's high-octane OCR that can also do some nice key-value, forms, and table extraction. To get the full IDP going, you'd have to roll in another service like Comprehend to read and index all of the extracted text, identify documents, detect sentiment, flag PII, etc etc. You used to have to build the integration yourself, but they recently released a more turnkey solution (link) that's very lean in terms of architecture.

You just upload the docs/text to S3, provide some info like document classes and example documents for training, and then it trains your model and will host it for real-time inference (running cost whether or not you're using it) or batch inference (only charged for the exact amount of compute used), and then it stores all the results back to S3.

The real question you should be asking is what business outcome are you trying to service with IDP? Are you trying to crunch through TB's of loan docs or business paperwork? Are you trying to gain some sort of market intelligence based on social media posts? Your architectural, business, and computing considerations are very different depending on what type of data you're consuming, and the timescale in which you need results relative to the data getting created. Most analysts skip that whole question of what a business actually needs to do, and try to find a broad one-size-fits-all solution, in my experience

  • Hook 'Em 1
Link to comment
Share on other sites

6 minutes ago, Captainant said:

Moving the convo to the nerdz thread, because yeah lol.

 

AWS Textract isn't IDP, it's high-octane OCR that can also do some nice key-value, forms, and table extraction. To get the full IDP going, you'd have to roll in another service like Comprehend to read and index all of the extracted text, identify documents, detect sentiment, flag PII, etc etc. You used to have to build the integration yourself, but they recently released a more turnkey solution (link) that's very lean in terms of architecture.

You just upload the docs/text to S3, provide some info like document classes and example documents for training, and then it trains your model and will host it for real-time inference (running cost whether or not you're using it) or batch inference (only charged for the exact amount of compute used), and then it stores all the results back to S3.

The real question you should be asking is what business outcome are you trying to service with IDP? Are you trying to crunch through TB's of loan docs or business paperwork? Are you trying to gain some sort of market intelligence based on social media posts? Your architectural, business, and computing considerations are very different depending on what type of data you're consuming, and the timescale in which you need results relative to the data getting created. Most analysts skip that whole question of what a business actually needs to do, and try to find a broad one-size-fits-all solution, in my experience

Good point re: AWS and Google Tesseract being OCR that can kinda sorta maybe do semi-structured things. Since you are very well-versed on this topic, would you mind answering a few questions?

I agree that having the business outcome you want to achieve first will inform the best route, but generally speaking it is my pov that IDP as a software "category", let's call it, is still immature and illegitimate. In that there is no true market leader or winner who has capabilities and features/functions that cannot be beat or repeated. I don't know that to be sure though, what are your thoughts in your experience? Who have you used (assuming you are an operator/user and not just a category buyer)? To what ends?

I know people at a few of these firms, but as you can see, it's insanely crowded from legacy tools to consulting firms who have a software offering to start-ups focusing on it. And this doesn't even include the software firms that are very narrowly focusing IDP on a very specific business unit, like AP/AR only a la a fintech, etc.

 

Everest Group’s IDP Products PEAK Matrix Assessment

Link to comment
Share on other sites

3 minutes ago, HamsterHookah said:

I don't know that to be sure though, what are your thoughts in your experience? Who have you used (assuming you are an operator/user and not just a category buyer)? To what ends?

I'm a builder, not a buyer so I haven't really used any off the shelf software and I usually support native implementations into long-running live environments. I can tell you that Google Cloud generally has a better user experience for small-scale workloads and early discovery, and that AWS will give you way better scalability and capabilities as you cross into petabyte-scale datasets and want to host more complex architectures. I also haven't heard of most of those businesses in your chart lol, but I can probably tell you how they're doing what they do.

In terms of business usage, I honestly haven't seen any business teams/groups that I work with really nicely explain what they want to do with their documents and data. Because I can show them how to get insights from their customer service interaction transcripts to see what parts of the business are causing more friction (without requiring tiresome customer surveys and extra steps for CSR's), but it's still up to their business owners to actually turn those insights into value by doing something with them.

12 minutes ago, HamsterHookah said:

I know people at a few of these firms, but as you can see, it's insanely crowded from legacy tools to consulting firms who have a software offering to start-ups focusing on it. And this doesn't even include the software firms that are very narrowly focusing IDP on a very specific business unit, like AP/AR only a la a fintech, etc.

Honestly, I think the idea of it as a market segment is putting the cart before the horse a little bit. Any Tom, Dick, or Harry with a bachelor's degree worth of knowledge and a month's worth of pizza and beer could realistically build a scalable IDP solution on a cloud platform, and wholly own the data and models and everything. How big such a hypothetical business becomes is really just a function of how good their sales and exec teams are - not how good the tech is. They also may be handling integrations and additional business logic, with the IDP really just being a datasource like in the case with Kofax.

I'm a little biased, but I usually prefer to roll my own solution or at least own the model and data and have the flexibility on where I scale up. The out of the box solutions are great for solving a tactical problem now, and building your own gives you the flexibility and control of your own design space to (hopefully) better meet your desired outcomes in the long run.

  • Like 1
Link to comment
Share on other sites

Quote

Today, we are announcing the third phase of our long-term partnership with OpenAI through a multiyear, multibillion dollar investment to accelerate AI breakthroughs to ensure these benefits are broadly shared with the world.

This agreement follows our previous investments in 2019 and 2021. It extends our ongoing collaboration across AI supercomputing and research and enables each of us to independently commercialize the resulting advanced AI technologies.

  • Supercomputing at scale – Microsoft will increase our investments in the development and deployment of specialized supercomputing systems to accelerate OpenAI’s groundbreaking independent AI research. We will also continue to build out Azure’s leading AI infrastructure to help customers build and deploy their AI applications on a global scale.
  • New AI-powered experiences – Microsoft will deploy OpenAI’s models across our consumer and enterprise products and introduce new categories of digital experiences built on OpenAI’s technology. This includes Microsoft’s Azure OpenAI Service, which empowers developers to build cutting-edge AI applications through direct access to OpenAI models backed by Azure’s trusted, enterprise-grade capabilities and AI-optimized infrastructure and tools.
  • Exclusive cloud provider – As OpenAI’s exclusive cloud provider, Azure will power all OpenAI workloads across research, products and API services.

...

https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/

 

Quote

... we've partnered with Microsoft to deploy our technology through our API and the Azure OpenAI Service — enabling enterprise and developers to build on top of GPT, DALL·E, and Codex. We’ve also worked together to build OpenAI’s technology into apps like GitHub Copilot and Microsoft Designer. ...

https://openai.com/blog/openai-and-microsoft-extend-partnership/

Edited by bernorange
Link to comment
Share on other sites

  • 2 months later...

Putting this here, although I worry it'll go zero eyeballs.

Preprint paper on a group that created generative agents using ChatGPT 3.5. https://arxiv.org/abs/2304.03442 (related article on the preprint here (https://arstechnica.com/information-technology/2023/04/surprising-things-happen-when-you-put-25-ai-agents-together-in-an-rpg-town/)

From the abstract:

Quote

Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools. In this paper, we introduce generative agents--computational software agents that simulate believable human behavior. Generative agents wake up, cook breakfast, and head to work; artists paint, while authors write; they form opinions, notice each other, and initiate conversations; they remember and reflect on days past as they plan the next day. To enable generative agents, we describe an architecture that extends a large language model to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior. We instantiate generative agents to populate an interactive sandbox environment inspired by The Sims, where end users can interact with a small town of twenty five agents using natural language. In an evaluation, these generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentine's Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time. We demonstrate through ablation that the components of our agent architecture--observation, planning, and reflection--each contribute critically to the believability of agent behavior. By fusing large language models with computational, interactive agents, this work introduces architectural and interaction patterns for enabling believable simulations of human behavior.

In addition to the nifty way they created a memory structure and loop for the agents, I found this part of the preprint interesting/terrifying:

Quote

Another application area is in the human-centered design process, similar to the intended applications of cognitive models such as GOMS [50] and the Keystroke Level Model [22]. Consider a generative agent that models Sal, the protagonist in Mark Weiser’s famous vignettes [101], based on her life patterns and interactions with technology. In this scenario, the agent acts as a proxy for Sal and learns plausible sets of behaviors and reflections that Sal may exhibit based on her life. The agent can encode information such as when Sal wakes up, when she needs her first cup of coffee, and what her typical day looks like. Using this information, the agent can automatically brew coffee, help get the kids ready for school, and adjust the ambient music and lighting to match Sal’s mood after a hard day at work. By utilizing generative agents as proxies for users, we can develop a deeper understanding of their needs and preferences, resulting in more personalized and effective technological experiences.

Imagine a company creating a simulation of you and letting that simulation go through its life in order for the company how best to target your future needs and wants. I'm not sure such a simulation would be any better at allowing a company to target users than current algorithms, but it sounds creep as hell. There is so much information at there on you, that it does seem plausible that they could utilize the method described in the preprint to incorporate it into an agent. It could even replicate family and friends and work. As the token limits increase for these models, you could eventually incorporate a stupid amount of information into such a simulation. 

They also put together a nifty little "demo," which appears to just be a recorded session of the simulation. You can watch as the agents move about their town and peer into their though processes, memories, and goals it detail. 

https://reverie.herokuapp.com/arXiv_Demo/

 

  • Hook 'Em 1
Link to comment
Share on other sites

This shit is just straight fucked up. 
 

On one hand, you have the truly fascinating work going on with these. I can respect that.

On the other you have the knowledge that no matter what controls are in place, there will be some form of great harm to come from it as it becomes exploited for nefarious purposes.

Quite the conundrum. 

Link to comment
Share on other sites

On 4/12/2023 at 4:44 PM, Dahobbs said:

Putting this here, although I worry it'll go zero eyeballs.

Preprint paper on a group that created generative agents using ChatGPT 3.5. https://arxiv.org/abs/2304.03442 (related article on the preprint here (https://arstechnica.com/information-technology/2023/04/surprising-things-happen-when-you-put-25-ai-agents-together-in-an-rpg-town/)

From the abstract:

In addition to the nifty way they created a memory structure and loop for the agents, I found this part of the preprint interesting/terrifying:

Imagine a company creating a simulation of you and letting that simulation go through its life in order for the company how best to target your future needs and wants. I'm not sure such a simulation would be any better at allowing a company to target users than current algorithms, but it sounds creep as hell. There is so much information at there on you, that it does seem plausible that they could utilize the method described in the preprint to incorporate it into an agent. It could even replicate family and friends and work. As the token limits increase for these models, you could eventually incorporate a stupid amount of information into such a simulation. 

They also put together a nifty little "demo," which appears to just be a recorded session of the simulation. You can watch as the agents move about their town and peer into their though processes, memories, and goals it detail. 

https://reverie.herokuapp.com/arXiv_Demo/

 

 

On 4/12/2023 at 6:50 PM, thunderlounge said:

This shit is just straight fucked up. 
 

On one hand, you have the truly fascinating work going on with these. I can respect that.

On the other you have the knowledge that no matter what controls are in place, there will be some form of great harm to come from it as it becomes exploited for nefarious purposes.

Quite the conundrum. 

Yeah this is definitely an emerging issue as we continue to find new uses for neural networks in problem solving. More generally speaking id say that Hobb's article is talking about synthetic data generation via simulated agent - the way it's written it sounds like there's a whole person happening in there, but I'd bet it's more akin to a nice UI that visualizes the state of a large connected graph of JSON entity definitions. 

 

To say that with less techno-jargon, they are not mimicking genuine thought and individuality, but rather the artifacts of those natural processes. But to borrow a phrase from Westworld, "if you can't tell the difference, does it really matter?"

 

Man, as bad as the last couple seasons were, they really swung for the fences with their ideation of negative externalities of AI. May be worth queueing up for a rewatch...

Link to comment
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.



×
×
  • Create New...