DISTILL IS DEDICATED TO MAKING MACHINE LEARNING CLEAR AND A modern medium for presenting research. The web is a powerful medium to share new ways of thinking. Over the last few years we’ve seen many imaginative examples of such work.But traditional academic publishing remains focused on the PDF, which prevents this sort of

communication.

THREAD: CIRCUITS

GROWING NEURAL CELLULAR AUTOMATA Growing models were trained to generate patterns, but don't know how to persist them. Some patterns explode, some decay, but some happen to be almost stable or even regenerate parts! Persistent models are trained to make the pattern stay for a prolonged period of time. . Interstingly, they often develop some regenerative capabilities without being explicitly instructed to do MULTIMODAL NEURONS IN ARTIFICIAL NEURAL NETWORKS Nick Cammarata†: Drew the connection between multimodal neurons in neural networks and multimodal neurons in the brain, which became the overall framing of the article. Created the conditional probability plots (regional, Trump, mental health), labeling more than 1500 images, discovered that negative pre-ReLU activations are often interpretable, and discovered that neurons sometimes UNDERSTANDING RL VISION Understanding RL Vision. With diverse environments, we can analyze, diagnose and edit deep reinforcement learning models using attribution. Attribution from a hidden layer to the value function, showing what features of the observation (left) are used to predict success (middle) and failure (right). Applying dimensionality

reduction (NMF

SELF-CLASSIFYING MNIST DIGITS DISTILL PRIZE FOR CLARITY IN MACHINE LEARNING Distill prizes are expected to be $10,000 USD. The Distill Prize has a $125,000 USD initial endowment, funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project. Logistics for the prize are handled by the Open Philanthropy Project. DISTILL IS DEDICATED TO MAKING MACHINE LEARNING CLEAR AND A modern medium for presenting research. The web is a powerful medium to share new ways of thinking. Over the last few years we’ve seen many imaginative examples of such work.But traditional academic publishing remains focused on the PDF, which prevents this sort of

communication.

THREAD: CIRCUITS

GROWING NEURAL CELLULAR AUTOMATA Growing models were trained to generate patterns, but don't know how to persist them. Some patterns explode, some decay, but some happen to be almost stable or even regenerate parts! Persistent models are trained to make the pattern stay for a prolonged period of time. . Interstingly, they often develop some regenerative capabilities without being explicitly instructed to do MULTIMODAL NEURONS IN ARTIFICIAL NEURAL NETWORKS Nick Cammarata†: Drew the connection between multimodal neurons in neural networks and multimodal neurons in the brain, which became the overall framing of the article. Created the conditional probability plots (regional, Trump, mental health), labeling more than 1500 images, discovered that negative pre-ReLU activations are often interpretable, and discovered that neurons sometimes UNDERSTANDING RL VISION Understanding RL Vision. With diverse environments, we can analyze, diagnose and edit deep reinforcement learning models using attribution. Attribution from a hidden layer to the value function, showing what features of the observation (left) are used to predict success (middle) and failure (right). Applying dimensionality

reduction (NMF

SELF-CLASSIFYING MNIST DIGITS DISTILL PRIZE FOR CLARITY IN MACHINE LEARNING Distill prizes are expected to be $10,000 USD. The Distill Prize has a $125,000 USD initial endowment, funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project. Logistics for the prize are handled by the Open Philanthropy Project. PUBLISHING IN THE DISTILL RESEARCH JOURNAL Distill publishes articles explaining, synthesizing and reviewing existing research. This includes Reviews, Tutorials, Primers, and Perspective articles. The editorial team is especially interested in explorable explanations. Examples: Why Momentum Really Works,

Attention and

HOW TO CREATE A DISTILL ARTICLE It is the assumed layout of any direct descendents of the dt-article element. .l-body. For images you want to display a little larger, try these: .l-middle. .l-page. All of these have an outset variant if you want to poke out from the body text a little bit. For instance: .l-body-outset. .l-middle-outset. GROWING NEURAL CELLULAR AUTOMATA Growing models were trained to generate patterns, but don't know how to persist them. Some patterns explode, some decay, but some happen to be almost stable or even regenerate parts! Persistent models are trained to make the pattern stay for a prolonged period of time. . Interstingly, they often develop some regenerative capabilities without being explicitly instructed to do DISTILL PRIZE FOR CLARITY IN MACHINE LEARNING Distill prizes are expected to be $10,000 USD. The Distill Prize has a $125,000 USD initial endowment, funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project. Logistics for the prize are handled by the Open Philanthropy Project. COMPUTING RECEPTIVE FIELDS OF CONVOLUTIONAL NEURAL NETWORKS r l =2. So, we obtain the general recurrence equation (which is first-order, non-homogeneous, with variable coefficients ): r l − 1 = s l ⋅ r l + ( k l − s l) This equation can be used in a recursive algorithm to compute the receptive field size of the

network, r 0. .

SEQUENCE MODELING WITH CTC The function L (Y) L(Y) L (Y) computes the length of Y Y Y in terms of the language model tokens and acts as a word insertion bonus. With a word-based language model L (Y) L(Y) L (Y) counts the number of words in Y. Y. Y. If we use a character-based language model then L (Y) L(Y) L (Y) counts the number of characters in Y. Y. Y. The language model scores are only included when a prefix is AN OVERVIEW OF EARLY VISION IN INCEPTIONV1 An Overview of Early Vision in InceptionV1. This article is part of the Circuits thread, a collection of short articles and commentary by an open scientific collaboration delving into the inner workings of neural networks. The first few articles of the Circuits project will be focused on early vision in InceptionV1 — for our purposes, the THE PATHS PERSPECTIVE ON VALUE LEARNING The paths that remain are the paths that the agent will follow at test time; they are the only ones it needs to pay attention to. This sort of value learning often leads to faster convergence than on-policy methods Try using the Playground at the end of this article to compare between approaches. . Double Q-Learning. FEATURE VISUALIZATION This article focuses on feature visualization. While feature visualization is a powerful tool, actually getting it to work involves a number of details. In this article, we examine the major issues and explore common approaches to solving them. We find that remarkably simple methods can produce high-quality visualizations.

CURVE DETECTORS

Part one of a three part deep dive into the curve neuron family. DISTILL IS DEDICATED TO MAKING MACHINE LEARNING CLEAR ANDDISTILL BETADISTILL STOCK CHECKERDISTILL VIDEO DOWNLOADERDISTILL WEB MONITORDISTILL WEBSITE A modern medium for presenting research. The web is a powerful medium to share new ways of thinking. Over the last few years we’ve seen many imaginative examples of such work.But traditional academic publishing remains focused on the PDF, which prevents this sort of

communication.

THREAD: CIRCUITS

GROWING NEURAL CELLULAR AUTOMATA Growing models were trained to generate patterns, but don't know how to persist them. Some patterns explode, some decay, but some happen to be almost stable or even regenerate parts! Persistent models are trained to make the pattern stay for a prolonged period of time. . Interstingly, they often develop some regenerative capabilities without being explicitly instructed to do MULTIMODAL NEURONS IN ARTIFICIAL NEURAL NETWORKS Nick Cammarata†: Drew the connection between multimodal neurons in neural networks and multimodal neurons in the brain, which became the overall framing of the article. Created the conditional probability plots (regional, Trump, mental health), labeling more than 1500 images, discovered that negative pre-ReLU activations are often interpretable, and discovered that neurons sometimes UNDERSTANDING RL VISION Understanding RL Vision. With diverse environments, we can analyze, diagnose and edit deep reinforcement learning models using attribution. Attribution from a hidden layer to the value function, showing what features of the observation (left) are used to predict success (middle) and failure (right). Applying dimensionality

reduction (NMF

SELF-CLASSIFYING MNIST DIGITS DISTILL PRIZE FOR CLARITY IN MACHINE LEARNING Distill prizes are expected to be $10,000 USD. The Distill Prize has a $125,000 USD initial endowment, funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project. Logistics for the prize are handled by the Open Philanthropy Project. HOW TO USE T-SNE EFFECTIVELY A popular method for exploring high-dimensional data is something called t-SNE, introduced by van der Maaten and Hinton in 2008 .The technique has become widespread in the field of machine learning, since it has an almost magical ability to create compelling two-dimensonal “maps” from data with hundreds or even thousands of

dimensions.

A VISUAL EXPLORATION OF GAUSSIAN PROCESSESSEE MORE ON DISTILL.PUB COMPUTING RECEPTIVE FIELDS OF CONVOLUTIONAL NEURAL NETWORKSSEE MORE ON

DISTILL.PUB

DISTILL IS DEDICATED TO MAKING MACHINE LEARNING CLEAR ANDDISTILL BETADISTILL STOCK CHECKERDISTILL VIDEO DOWNLOADERDISTILL WEB MONITORDISTILL WEBSITE A modern medium for presenting research. The web is a powerful medium to share new ways of thinking. Over the last few years we’ve seen many imaginative examples of such work.But traditional academic publishing remains focused on the PDF, which prevents this sort of

communication.

THREAD: CIRCUITS

GROWING NEURAL CELLULAR AUTOMATA Growing models were trained to generate patterns, but don't know how to persist them. Some patterns explode, some decay, but some happen to be almost stable or even regenerate parts! Persistent models are trained to make the pattern stay for a prolonged period of time. . Interstingly, they often develop some regenerative capabilities without being explicitly instructed to do MULTIMODAL NEURONS IN ARTIFICIAL NEURAL NETWORKS Nick Cammarata†: Drew the connection between multimodal neurons in neural networks and multimodal neurons in the brain, which became the overall framing of the article. Created the conditional probability plots (regional, Trump, mental health), labeling more than 1500 images, discovered that negative pre-ReLU activations are often interpretable, and discovered that neurons sometimes UNDERSTANDING RL VISION Understanding RL Vision. With diverse environments, we can analyze, diagnose and edit deep reinforcement learning models using attribution. Attribution from a hidden layer to the value function, showing what features of the observation (left) are used to predict success (middle) and failure (right). Applying dimensionality

reduction (NMF

SELF-CLASSIFYING MNIST DIGITS DISTILL PRIZE FOR CLARITY IN MACHINE LEARNING Distill prizes are expected to be $10,000 USD. The Distill Prize has a $125,000 USD initial endowment, funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project. Logistics for the prize are handled by the Open Philanthropy Project. HOW TO USE T-SNE EFFECTIVELY A popular method for exploring high-dimensional data is something called t-SNE, introduced by van der Maaten and Hinton in 2008 .The technique has become widespread in the field of machine learning, since it has an almost magical ability to create compelling two-dimensonal “maps” from data with hundreds or even thousands of

dimensions.

A VISUAL EXPLORATION OF GAUSSIAN PROCESSESSEE MORE ON DISTILL.PUB COMPUTING RECEPTIVE FIELDS OF CONVOLUTIONAL NEURAL NETWORKSSEE MORE ON

DISTILL.PUB

PUBLISHING IN THE DISTILL RESEARCH JOURNAL Distill publishes articles explaining, synthesizing and reviewing existing research. This includes Reviews, Tutorials, Primers, and Perspective articles. The editorial team is especially interested in explorable explanations. Examples: Why Momentum Really Works,

Attention and

HOW TO CREATE A DISTILL ARTICLE It is the assumed layout of any direct descendents of the dt-article element. .l-body. For images you want to display a little larger, try these: .l-middle. .l-page. All of these have an outset variant if you want to poke out from the body text a little bit. For instance: .l-body-outset. .l-middle-outset. GROWING NEURAL CELLULAR AUTOMATA Growing models were trained to generate patterns, but don't know how to persist them. Some patterns explode, some decay, but some happen to be almost stable or even regenerate parts! Persistent models are trained to make the pattern stay for a prolonged period of time. . Interstingly, they often develop some regenerative capabilities without being explicitly instructed to do DISTILL PRIZE FOR CLARITY IN MACHINE LEARNING Distill prizes are expected to be $10,000 USD. The Distill Prize has a $125,000 USD initial endowment, funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project. Logistics for the prize are handled by the Open Philanthropy Project. COMPUTING RECEPTIVE FIELDS OF CONVOLUTIONAL NEURAL NETWORKS r l =2. So, we obtain the general recurrence equation (which is first-order, non-homogeneous, with variable coefficients ): r l − 1 = s l ⋅ r l + ( k l − s l) This equation can be used in a recursive algorithm to compute the receptive field size of the

network, r 0. .

SEQUENCE MODELING WITH CTC The function L (Y) L(Y) L (Y) computes the length of Y Y Y in terms of the language model tokens and acts as a word insertion bonus. With a word-based language model L (Y) L(Y) L (Y) counts the number of words in Y. Y. Y. If we use a character-based language model then L (Y) L(Y) L (Y) counts the number of characters in Y. Y. Y. The language model scores are only included when a prefix is AN OVERVIEW OF EARLY VISION IN INCEPTIONV1 An Overview of Early Vision in InceptionV1. This article is part of the Circuits thread, a collection of short articles and commentary by an open scientific collaboration delving into the inner workings of neural networks. The first few articles of the Circuits project will be focused on early vision in InceptionV1 — for our purposes, the THE PATHS PERSPECTIVE ON VALUE LEARNING The paths that remain are the paths that the agent will follow at test time; they are the only ones it needs to pay attention to. This sort of value learning often leads to faster convergence than on-policy methods Try using the Playground at the end of this article to compare between approaches. . Double Q-Learning. FEATURE VISUALIZATION This article focuses on feature visualization. While feature visualization is a powerful tool, actually getting it to work involves a number of details. In this article, we examine the major issues and explore common approaches to solving them. We find that remarkably simple methods can produce high-quality visualizations.

CURVE DETECTORS

Part one of a three part deep dive into the curve neuron family.

Distill

About Prize Submit

March 4, 2021

Peer-reviewed

MULTIMODAL NEURONS IN ARTIFICIAL NEURAL NETWORKS Gabriel Goh, Nick Cammarata †, Chelsea Voss †, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah We report the existence of multimodal neurons in artificial neural networks, similar to those found in the human brain.

Nov. 17, 2020

Peer-reviewed

UNDERSTANDING RL VISION Jacob Hilton, Nick Cammarata, Shan Carter, Gabriel Goh, and

Chris Olah

With diverse environments, we can analyze, diagnose and edit deep reinforcement learning models using attribution.

Sept. 11, 2020

Commentary

COMMUNICATING WITH INTERACTIVE ARTICLES Fred Hohman, Matthew Conlen, Jeffrey Heer, and Duen Horng (Polo)

Chau

Examining the design of interactive articles by synthesizing theory from disciplines such as education, journalism, and visualization.

Aug. 27, 2020

Thread

THREAD: DIFFERENTIABLE SELF-ORGANIZING SYSTEMS Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, Michael Levin, and Sam Greydanus A collection of articles and comments with the goal of understanding how to design robust and general purpose self-organizing systems.

May 5, 2020

Peer-reviewed

EXPLORING BAYESIAN OPTIMIZATION Apoorv Agnihotri and Nipun Batra How to tune hyperparameters for your machine learning model using Bayesian optimization.

March 16, 2020

Peer-reviewed

VISUALIZING NEURAL NETWORKS WITH THE GRAND TOUR Mingwei Li, Zhenge Zhao, and Carlos Scheidegger By focusing on linear dimensionality reduction, we show how to visualize many dynamic phenomena in neural networks.

March 10, 2020

Thread

THREAD: CIRCUITS

Nick Cammarata, Shan Carter, Gabriel Goh, Chris Olah, Michael Petrov, and Ludwig Schubert What can we learn if we invest heavily in reverse engineering a single

neural network?

Jan. 10, 2020

Peer-reviewed

VISUALIZING THE IMPACT OF FEATURE ATTRIBUTION BASELINES Pascal Sturmfels, Scott Lundberg, and Su-In Lee Exploring the baseline input hyperparameter, and how it impacts interpretations of neural network behavior.

Nov. 4, 2019

Peer-reviewed

COMPUTING RECEPTIVE FIELDS OF CONVOLUTIONAL NEURAL NETWORKS André Araujo, Wade Norris, and Jack Sim Detailed derivations and open-source code to analyze the receptive

fields of convnets.

Sept. 30, 2019

Peer-reviewed

THE PATHS PERSPECTIVE ON VALUE LEARNING Sam Greydanus and Chris Olah A closer look at how Temporal Difference Learning merges paths of experience for greater statistical efficiency

Aug. 6, 2019

Commentary

A DISCUSSION OF ‘ADVERSARIAL EXAMPLES ARE NOT BUGS, THEY ARE

FEATURES’

Logan Engstrom, Justin Gilmer, Gabriel Goh, Dan Hendrycks, Andrew Ilyas, Aleksander Madry, Reiichiro Nakano, Preetum Nakkiran, Shibani Santurkar, Brandon Tran, Dimitris Tsipras, and Eric Wallace Six comments from the community and responses from the original

authors

April 9, 2019

Commentary

OPEN QUESTIONS ABOUT GENERATIVE ADVERSARIAL NETWORKS

Augustus Odena

What we’d like to find out about GANs that we don’t know yet.

April 2, 2019

Peer-reviewed

A VISUAL EXPLORATION OF GAUSSIAN PROCESSES Jochen Görtler, Rebecca Kehlbeck, and Oliver Deussen How to turn a collection of small building blocks into a versatile tool for solving regression problems.

March 25, 2019

Peer-reviewed

VISUALIZING MEMORIZATION IN RNNS

Andreas Madsen

Inspecting gradient magnitudes in context can be a powerful tool to see when recurrent units use short-term or long-term contextual

understanding.

March 6, 2019

Peer-reviewed

ACTIVATION ATLAS

Shan Carter, Zan Armstrong, Ludwig Schubert, Ian Johnson, and

Chris Olah

By using feature inversion to visualize millions of activations from an image classification network, we create an explorable activation atlas of features the network has learned and what concepts it typically represents.

Feb. 19, 2019

Commentary

AI SAFETY NEEDS SOCIAL SCIENTISTS Geoffrey Irving and Amanda Askell If we want to train AI to do what humans want, we need to study

humans.

Aug. 14, 2018

Editorial

DISTILL UPDATE 2018

Distill Editors

An Update from the Editorial Team

July 25, 2018

Peer-reviewed

DIFFERENTIABLE IMAGE PARAMETERIZATIONS Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and

Chris Olah

A powerful, under-explored tool for neural network visualizations and

art.

July 9, 2018

Peer-reviewed

FEATURE-WISE TRANSFORMATIONS Vincent Dumoulin, Ethan Perez, Nathan Schucher, Florian Strub, Harm de Vries, Aaron Courville, and Yoshua Bengio A simple and surprisingly effective family of conditioning mechanisms.

March 6, 2018

Peer-reviewed

THE BUILDING BLOCKS OF INTERPRETABILITY Chris Olah, Arvind Satyanarayan, Ian Johnson, Shan Carter, Ludwig Schubert, Katherine Ye, and Alexander Mordvintsev Interpretability techniques are normally studied in isolation. We explore the powerful interfaces that arise when you combine them — and the rich structure of this combinatorial space.

Dec. 4, 2017

Commentary

USING ARTIFICIAL INTELLIGENCE TO AUGMENT HUMAN INTELLIGENCE Shan Carter and Michael Nielsen By creating user interfaces which let us work with the representations inside machine learning models, we can give people new tools for

reasoning.

Nov. 27, 2017

Peer-reviewed

SEQUENCE MODELING WITH CTC

Awni Hannun

A visual guide to Connectionist Temporal Classification, an algorithm used to train deep neural networks in speech recognition, handwriting recognition and other sequence problems.

Nov. 7, 2017

Peer-reviewed

FEATURE VISUALIZATION Chris Olah, Alexander Mordvintsev, and Ludwig Schubert How neural networks build up their understanding of images

April 4, 2017

Peer-reviewed

WHY MOMENTUM REALLY WORKS

Gabriel Goh

We often think of optimization with momentum as a ball rolling down a hill. This isn’t wrong, but there is much more to the story.

March 22, 2017

Commentary

RESEARCH DEBT

Chris Olah and Shan Carter Science is a human activity. When we fail to distill and explain research, we accumulate a kind of debt...

Dec 6, 2016

EXPERIMENTS IN HANDWRITING WITH A NEURAL NETWORK Shan Carter, David Ha, Ian Johnson, and Chris Olah Several interactive visualizations of a generative model of handwriting. Some are fun, some are serious.

Oct 17, 2016

DECONVOLUTION AND CHECKERBOARD ARTIFACTS Augustus Odena, Vincent Dumoulin, and Chris Olah When we look very closely at images generated by neural networks, we often see a strange checkerboard pattern of artifacts.

Oct 13, 2016

HOW TO USE T-SNE EFFECTIVELY Martin Wattenberg, Fernanda Viégas, and Ian Johnson Although extremely useful for visualizing high-dimensional data, t-SNE plots can sometimes be mysterious or misleading.

Sept 8, 2016

ATTENTION AND AUGMENTED RECURRENT NEURAL NETWORKS Chris Olah and Shan Carter A visual overview of neural attention, and the powerful extensions of neural networks being built on top of it. Distill is dedicated to clear explanations of machine learning

About Submit

Prize

Archive RSS

GitHub Twitter

ISSN 2476-0757