Brandon Smock

Research Scientist, Machine Learning

[Brandon Smock @deepsmock]

An LLM’s understanding of text, built through modeling word relationships, might approximate real text and real word relationships very closely. But it’s not because an LLM understands the meaning of individual words. *This* meaning is essentially impossible for it to infer.

[Brandon Smock @deepsmock]

Large language models, trained purely on text to predict the next word (token), are not taught the meaning of any individual word—nor do they learn it.

[Brandon Smock @deepsmock]

The consequence is that LLMs produce words without understanding any of their *individual* meanings.

[DailyPapers @HuggingPapers]

Kensho just released PubTables-v2 on Hugging Face

The first large-scale dataset for multi-page table extraction
with 135k+ cropped tables, 548k+ full-page tables, and 9k+ documents

[Brandon Smock @deepsmock]

More than 4 years after the release of PubTables-1M, PubTables-v2 is finally here! Among the highlights: this is the first dataset for document table extraction containing thousands of multi-page tables, some spanning up to 13 pages!

[Kensho @Kensho]

Announcing PubTables-v2: a first-of-its-kind large-scale dataset of 500k+ tables empowering end-to-end document AI. Annotates tables spanning multiple pages, unlocking deeper doc context for model training. Dive in and download:
https://t.co/kYNn03uXIe

#DocumentAI #VLM #OpenData

[Brandon Smock @deepsmock]

This is achieved via prompting in natural language, which provides a way to interpolate new tasks in between tasks the model has previously seen.

While still a shallow form of generalization, this has undoubtedly been a massive leap in the capabilities of these models.

[Brandon Smock @deepsmock]

In deep learning today generalization is powered by interpolation.

Before LLMs, generalization was constrained to handling novel input for a single narrowly-defined task.

The true breakthrough of LLMs in deep learning was to generalize beyond new inputs to entirely new *tasks*.

[Brandon Smock @deepsmock]

@tdietterich @ylecun @Michael_J_Black This is exactly how I remember it, as well. As someone who has always been a huge fan of Yann and his work, his consistent distortion and/or misunderstanding of the backlash to Galactica’s own overhyped claims has been painful to read.

[Brandon Smock @deepsmock]

Looking forward to #CVPR2024 in Seattle! If you’re attending, reach out and let’s grab ☕️.

My team at Kensho is hiring a new grad MLE for visual document understanding. We don’t have a booth this year but if you track me down I have some free Kensho swag to hand out. 🧢✍️🛍️

[Brandon Smock @deepsmock]

Today we released the weights for three new pre-trained Table Transformer (TATR) models, for recognizing and extracting data from tables in unstructured documents. Check them out here: https://t.co/mxhYFoVWPI #icdar2023

[Brandon Smock @deepsmock]

@rubenaros Gracias! I think you might be the first person to have discovered the updates coming soon to the main GitHub repository 🙂

[Brandon Smock @deepsmock]

@fchollet François, your clarity of thought and willingness to share in resistance to hype and groupthink is unmatched and immensely valuable. Thank you.

[Novo Nordisk @novonordisk]

#PRESS Novo Nordisk and Microsoft collaborate to accelerate drug discovery and development using big data and artificial intelligence.

[Brandon Smock @deepsmock]

@fchollet Mine just arrived today, I’ll need a few days to read it first 😉

[Brandon Smock @deepsmock]

We are excited to announce our work “PubTables-1M: Towards comprehensive table extraction from unstructured documents” has been accepted at #CVPR2022! @RohithPesala

Paper, code, and data are all publicly released: https://t.co/mxhYFpcZRI

[Papers with Datasets @paperswithdata]

PubTables-1M: a large-scale dataset consisting of ~1 million tables from PubMed Central Open Access scientific articles.

Can be used for training/evaluating machine learning models for the tasks of table detection, table structure recognition, and more.

https://t.co/4x0DJd4FxX