Unlock the full potential of AI with Building LLMs for Production—our 470+ page guide to mastering LLMs with practical projects and expert insights!


The NLP Cypher | 01.10.21
Latest   Machine Learning   Newsletter

The NLP Cypher | 01.10.21

Last Updated on July 24, 2023 by Editorial Team

Author(s): Ricky Costa

Originally published on Towards AI.

Epiphany U+007C Ernst


The NLP Cypher U+007C 01.10.21

Melting Clocks

Once in a while you discover a goodie in the dregs of research. A cipher cracking paper emerged recently on the topic of using seq2seq models to crack 1:1 substitution ciphers. U+1F649

(1:1 substitution is when ciphertext represents a fixed character in the target plaintext. Read more here if you prefer to live dangerously.

Several deciphering methods used today make a big assumption. That we know the target language of the cipher we need to crack. But when diving into encrypted historical texts where the target language is unknown, well, you tend to get a big headache when the language origin is ambiguous.

When one begins to attack encrypted text. The state of the cipher can be in various conditions: alphanumeric (numbers/letters) or it can even be symbolic or it can be a mix of both (like the Zodiac Killer’s ciphers U+1F447).

However, IF we know ahead of time that the cipher’s plaintext language is… say English (and not Latin, or any other language), well, we are off to a good start and with a healthy advantage. Why? Because we can leverage the unique features of the English language that don’t occur in other languages. I.e. the letter “e” is the most frequent letter in English, so it’s possible the most frequent letter in the ciphertext could be the letter “e” , and by using these heuristics, letter by letter you slowly turn into Tom Hanks from the Da Vinci Code.

Letter Frequencies in the English Language

The third column represents proportions, taking the least common letter (q) as equal to 1. The letter E is over 56…


What’s really interesting about this paper is that the authors wanted to test if a multi-lingual seq2seq transformer would be able to crack ciphers WITHOUT knowing the origin of the language of the plaintext. They formulated the decipherment as a sequence-to-sequence translation problem. The model was trained on the character level.

What’s cool is that they tested the model on historical ciphers (that have been previously cracked) such as the Borg cipher and it was able to crack the first 256 characters with very low error. According to the authors, this is the first application of sequence-to-sequence neural models for decipherment!

NSA be like…


If you enjoy this read, please give it a U+1F44FU+1F44F and share with your friends! It really helps us out!

Don’t Worry There’s a Stack Exchange for Crypto Nerds

Cryptography Stack Exchange

Cryptography Stack Exchange is a question and answer site for software developers, mathematicians and others interested…


OpenAI Dropping Jewels

You probably have already heard of OpenAI’s model drops from this week so I’ll save you the recap. Added their two blogs in case you want to catch up. This week I added the Colab notebook for CLIP on LinkedIn and it got a good reception, will also append it here if you are interested:

Colab of the Week U+007C CLIP

Google Colaboratory

Edit description



DALL·E: Creating Images from Text

DALL·E is a 12-billion parameter version of GPT-3 trained to generate images from text descriptions, using a dataset of…



CLIP: Connecting Text and Images

We’re introducing a neural network called CLIP which efficiently learns visual concepts from natural language…


DALL-E Replication Already on GitHub

Surprise! Someone already replicated DALL-E on PyTorch U+1F601. U+1F525U+1F525

pip install dalle-pytorch


Implementation / replication of DALL-E, OpenAI’s Text to Image Transformer, in Pytorch — lucidrains/DALLE-pytorch


Object Storage Search Engine

Thank your local hacker

Hey you know how when you setup your S3 bucket or another object storage and you have the option to choose between public or private setting. Well have you ever wondered what it would look like if someone could harvest all the public bucket URLs for you to openly search them: U+1F447

Welcome to the Matrix

Inside the Rabbit Hole

The Ecco library allows one to visualize why language models bust moves the way they do. The library is mostly focused on autoregressive models (e.g. GPT-2/3 models). They currently have 2 notebooks to visualize neuron activation and input saliency.

It is built on top of PyTorch and Transformers.

Look Inside Language Models

Ecco is a python library that creates interactive visualizations allowing you to explore what your NLP Language Model…


Interfaces for Explaining Transformer Language Models

Interfaces for exploring transformer language models by looking at input saliency and neuron activation. Explorable #1…


Text-to-Speech with Swag

15.ai came on the scene in 2019 with its awesome text-to-speech demo and it’s been refining its models’ capabilities ever since. You can type in text and get deep learning generated speech conditioned on various characters ranging from HAL 9000 from 2001: Space Odyssey to Doctor Who.

15.ai: Natural TTS with minimal data

15.ai: Natural high-quality faster-than-real-time text-to-speech synthesis with minimal data


ML Metadata

Google came out with Machine Learning Metadata (MLMD). A library to keep track of your entire ML workflow. Allows you to version your models and datasets so you know why things go wrong when they do.

ML Metadata: Version Control for ML

January 08, 2021 – Posted by Ben Mathes and Neoklis Polyzotis, on behalf of the TFX Team When you write code, you need…


El GitHub:


ML Metadata (MLMD) is a library for recording and retrieving metadata associated with ML developer and data scientist…



mlmd.metadata_store.MetadataStore U+007C TFX U+007C TensorFlow

A store for the artifact metadata. mlmd.metadata_store.MetadataStore( config…


NNs for iOS with Wolfram Language

Wolfram out of left field, and he brought a smartphone. In a recent Wolfram blog post, they show how to train an image classifier, throwing it on ONNX, and then converting it to Core ML so it can be used on iOS devices. Includes code!

Deploy a Neural Network to Your iOS Device Using the Wolfram Language-Wolfram Blog

January 7, 2021 – Jofre Espigule-Pons, Machine Learning Today's handheld devices are powerful enough to run neural…


Machine Learning Index w/ Code

A huge index with several hundred projects per index on all things machine learning, includes computer vision and NLP. You can find the Super Duper NLP Repo on it U+1F60E.


500 AI Machine learning Deep learning Computer vision NLP Projects with code …


Repo Cypher U+1F468‍U+1F4BB

A collection of recent released repos that caught our U+1F441


Ask2Transformers automatically annotates text data.. aka zero-shot. U+1F525


This repository contains the code for the work Ask2Transformers – Zero Shot Domain Labelling with Pretrained…



A parameter efficient Transformer-based model which combines the newly proposed Sandwich-style parameter sharing technique.


This repository contains the code for the Subformer. To help overcome this we propose the Subformer, allowing us to…



Open-domain QA evaluation library, it includes efficient reader comparison, reproducible research, and knowledge source for applications.


A Simple and Fair Evaluation Library for Open-domain Question Answering Oepn-domain QA Evaluation usually means days of…



Arabic BERT returns for a 2nd week in a row on the Cypher. This time its ARBERT and MARBERT. It also includes ArBench a benchmark for Arabic NLU based on 41 datasets across 5 different tasks.


This is the repository accompanying our paper ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic. In the…



CRSLab is an open-source toolkit for building Conversational Recommender System (CRS). Includes models and datasets.


CRSLab is an open-source toolkit for building Conversational Recommender System (CRS). It is developed based on Python…


Dataset of the Week: StrategyQA

What is it?

“StrategyQA is a question-answering benchmark focusing on open-domain questions where the required reasoning steps are implicit in the question and should be inferred using a strategy. StrategyQA includes 2,780 examples, each consisting of a strategy question, its decomposition, and evidence paragraphs.”


Example 1

“Is growing seedless cucumber good for a gardener with entomophobia?”

Answer: Yes
Explanation: Seedless cucumber fruit does not require pollination. Cucumber plants need insects to pollinate them. Entomophobia is a fear of insects.

Example 2

“Are chinchillas cold-blooded?”

Answer: No
Explanation: Chinchillas are rodents, which are mammals. All mammals are warm-blooded.

Example 3

“Would Janet Jackson avoid a dish with ham?”

Answer: Yes
Explanation: Janet Jackson follows an Islamic practice. Islamic culture avoids eating pork. Ham is made from pork.

Where is it?

StrategyQA Dataset – Allen Institute for AI

The StrategyQA dataset was created through a crowdsourcing pipeline for eliciting creative and diverse yes/no questions…


Every Sunday we do a weekly round-up of NLP news and code drops from researchers around the world.

For complete coverage, follow our Twitter: @Quantum_Stat

Quantum Stat

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI

Feedback ↓