Tal Perry

I am building an agent that measures rooms from photos.

Over the weekend I built an AI agent.

built a system to teach my kid to read, using a free program called Anki for the “planning” and AI to make content that would lure him in. t worked, he and I enjoyed it immensely, he reads fluently (though mechanically).

Triton Inference Server has become a popular choice for production model serving, and for good reason: it is fast, flexible, and powerful.

Before diving into the reasons behind my decision, it’s essential to know that I am a Jewish Israeli atheist living in Berlin.

LightTag has been acquired by Primer.ai.

Cross-entropy-based training is the standard in deep learning and NLP.

Snorkel promises “Data Programming” - the user writes noisy labeling functions, and Snorkel learns probabilistic labels we can use as training data.

At LightTag, we do database multi-tenancy with Django and Postgres.

This post comes with a repo Our previous post on aligning span annotations to Hugging Face’s tokenizer outputs discussed the various tradeoffs one needs to consider, and concluded that a windowing strategy over the tokenized text and labels is optimal for…

Practical NLP operates on long texts, and annotations for sequence-labeling tasks often come in offset format.

You’d think that getting a cursor position in text is a solved problem, but here are some surprising bugs: In Windows PowerShell pressing backspace doesn’t work as you’d expect Jupyter Notebook had misaligned cursor data until 2017 DraftJS has users occas…

Active Learning is a popular technique to reduce annotation costs by using AI to decide what to label next.

tl;dr Calculate Krippendorff’s Alpha on any DataFrame in two lines.

An Alternative To BERT Pre-trained language models are great until you get the GPU bill or have to work on data that the model wasn’t pre-trained on (or worse yet, both).

tl;dr I compared four open-source NLP models by putting their outputs into LightTag and then reviewing their outputs manually.

What has four letters, sometimes has nine letters but never has five letters.

Range types are one of our favorite features in Postgres, we use them for NLP to detect conflicting entity annotations .

tl;dr Textual search is the easiest way to apply domain expertise to a corpus of text You think you need a server to power your text, but you are wrong.

1.

We’re a natural language processing company and we usually blog about natural language processing.

Like any new technology, AI can be used to reduce costs by optimizing existing processes or unleash value by enabling new processes and products.

NLP systems, like life, are about choices and tradeoffs.

TensorFlow’s Estimator API makes the engineering and operational aspects of deep/machine learning simpler.

In this post we’ll discuss the recent paper TextRay: Mining Clinical Reports to Gain a Broad Understanding of Chest X-rays focusing on the best practices the paper exemplifies with regards to labeling text data for NLP.

Deep learning applied to NLP has allowed practitioners understand their data less, in exchange for more labeled data.

Labeled data has become paramount to the success of many business ventures and research projects.

On Thursday June 7th 2018 a new customer emailed us reporting that they were unable to submit the work they’d done on our platform.

tl;dr RNNS work great for text but convolutions can do it faster Any part of a sentence can influence the semantics of a word.

Update 15.03.2024 I wrote this more than seven years ago.