Strange that they do not compare it against TabFN, which is another foundation model for tabular data. (https://github.com/PriorLabs/TabPFN)
TabPFN is an amazing innovation. But there are some crucial differences in model capabilities that make it hard for a fair comparison.
TabPFN can only operate on a single small table. But real-world datasets are actually multi-table and to make accurate prediction you need to capture signal from multiple tables (for example, customers, products, purchases).
So, the comparison to TabPFN would be unfair as it would only use data from a single table and that would lead to bad performance of TabPFN.
Interesting timing, they have recently reached out to my $dayjob. We will be probably be running a workshop on our (massive) dataset with them. I'd like to evaluate the performance of a couple of analytical models we've manually built against whatever this model can do based on some prompts. Exciting times!
Jure Leskovec was my Professor at Stanford a few years back, cool to see he's behind this.
He seemed like a good guy and got the sense that he was destined to do something big
Vid is a good friend of mine and he's wicked smart and also a very solid guy I adore.
I'm also guessing at some point he will probably read this comment, so hey Vid! See you at the next VRSA meetup!
Wickedly smart team indeed!
interesting! Super cool idea to augment software built with traditional DBs
I had some thoughts [1] around a concept similar to this a while ago, although it was much less refined. My thinking was around whether or not we could have a neural net remember a relational database schema, and be able to be queried for facts it knows, and facts it might predict.
This seems like a much more sensical (and actualised) stab at this kinda concept.
[1]: dancrimp.nz/2024/11/01/semantic-db/
So suppose I've got a database of behavioral and neuroimaging data from a research study on autism. Is this something that can be used to predict diagnosis from the other data fields?
Yes, I think this would work. For example, you'd organize the data into 3 tables: patients, behaviors and images. The patients table would have a partially filled-out "diagnosis" column. The model would then predict diagnosis of not-yet-diagnosed patients based on the patterns in data fields of previously diagnosed patients.
I feel like this is the next big thing for AI, having the ability to interact with any sort of structured dataset out of the box. Very cool project!
I'll suspect it'll be more like the next little thing. Most of don't interact that much with structured data, so the applications will be very specific.
However, the algo-trading crowd, will likely be very interested in this. They deal with structured data all day and it would surprise me if most of them don't already have things like this working in their networks. They seem to be very secretive, though, so we're not gonna hear much.
We all interact with structured data models constantly, like literally thousands of times each day, just indirectly.
Every single credit card purchase gets classified by a model as fraud or ok. When you go to Netflix and see recommended movies, it's all predictions on structured data. Every single post in every social media feed is there because a model predicted you'd like it.
Realistically, it might be more like 10s of thousands or even hundreds of thousands of predictions that we engage with in a day.
If reality matches the benchmarks for this model, it can kick off a whole new category of models that can potentially be bigger than LLMs
Structured data = relational data
This has more applications than you might first think.
Does AI for relational data work the same way as token predictions does for LLM AI?
So can this be used to predict patterns for traffic, restaurant table availability, and your customers’ demand for things based on other customers?
Hey! I'm one of the engineers who worked on this project.
These are all problems that KumoRFM is able to solve given that you have the right relational data of course! So e.g. for predicting restaurant table availability you would need at least an occupancy table which records how many seats were available historically and you can predict its future entries.
But you can also add more relevant data without joining into a single table, so you can add a restaurants table, a holiday-calendar table, weather patterns, etc. and KumoRFM should take it all into account when predicting.
Strange that they do not compare it against TabFN, which is another foundation model for tabular data. (https://github.com/PriorLabs/TabPFN)
TabPFN is an amazing innovation. But there are some crucial differences in model capabilities that make it hard for a fair comparison.
TabPFN can only operate on a single small table. But real-world datasets are actually multi-table and to make accurate prediction you need to capture signal from multiple tables (for example, customers, products, purchases).
So, the comparison to TabPFN would be unfair as it would only use data from a single table and that would lead to bad performance of TabPFN.
Interesting timing, they have recently reached out to my $dayjob. We will be probably be running a workshop on our (massive) dataset with them. I'd like to evaluate the performance of a couple of analytical models we've manually built against whatever this model can do based on some prompts. Exciting times!
Jure Leskovec was my Professor at Stanford a few years back, cool to see he's behind this.
He seemed like a good guy and got the sense that he was destined to do something big
Vid is a good friend of mine and he's wicked smart and also a very solid guy I adore.
I'm also guessing at some point he will probably read this comment, so hey Vid! See you at the next VRSA meetup!
Wickedly smart team indeed!
interesting! Super cool idea to augment software built with traditional DBs
I had some thoughts [1] around a concept similar to this a while ago, although it was much less refined. My thinking was around whether or not we could have a neural net remember a relational database schema, and be able to be queried for facts it knows, and facts it might predict.
This seems like a much more sensical (and actualised) stab at this kinda concept.
[1]: dancrimp.nz/2024/11/01/semantic-db/
So suppose I've got a database of behavioral and neuroimaging data from a research study on autism. Is this something that can be used to predict diagnosis from the other data fields?
Yes, I think this would work. For example, you'd organize the data into 3 tables: patients, behaviors and images. The patients table would have a partially filled-out "diagnosis" column. The model would then predict diagnosis of not-yet-diagnosed patients based on the patterns in data fields of previously diagnosed patients.
I feel like this is the next big thing for AI, having the ability to interact with any sort of structured dataset out of the box. Very cool project!
I'll suspect it'll be more like the next little thing. Most of don't interact that much with structured data, so the applications will be very specific.
However, the algo-trading crowd, will likely be very interested in this. They deal with structured data all day and it would surprise me if most of them don't already have things like this working in their networks. They seem to be very secretive, though, so we're not gonna hear much.
We all interact with structured data models constantly, like literally thousands of times each day, just indirectly.
Every single credit card purchase gets classified by a model as fraud or ok. When you go to Netflix and see recommended movies, it's all predictions on structured data. Every single post in every social media feed is there because a model predicted you'd like it.
Realistically, it might be more like 10s of thousands or even hundreds of thousands of predictions that we engage with in a day.
If reality matches the benchmarks for this model, it can kick off a whole new category of models that can potentially be bigger than LLMs
Structured data = relational data
This has more applications than you might first think.
Does AI for relational data work the same way as token predictions does for LLM AI?
So can this be used to predict patterns for traffic, restaurant table availability, and your customers’ demand for things based on other customers?
Hey! I'm one of the engineers who worked on this project.
These are all problems that KumoRFM is able to solve given that you have the right relational data of course! So e.g. for predicting restaurant table availability you would need at least an occupancy table which records how many seats were available historically and you can predict its future entries.
But you can also add more relevant data without joining into a single table, so you can add a restaurants table, a holiday-calendar table, weather patterns, etc. and KumoRFM should take it all into account when predicting.
A real-time in-context label generator. Nice...