And she looks like the moon. So close and yet, so far.aim highmore quotes

art: beautiful

EMBO Practical Course: Bioinformatics and Genome Analysis, 5–17 June 2017.

art + science

Bloomberg Businessweek Design Conference — San Francisco, 2013

Design loves science and science loves design, but doesn't always know it. (Bloomberg Businessweek)

science ♥ design

Together with Alberto Cairo, I presented at the Bloomberg Businessweek Design Conference (highlights) on the topic of design and communication in the sciences.

Alberto, as the journalist, motivated why communication should include access to detail through an engaging narrative. He made the distinction between the specialist (heavy on detail) and the communicator (focus on narrative) and emphasized that the distinction is artificial, though often played out (watch video).

I, as the scientist, underscored the importance of clear communication between scientists. As the specialists, they are often very poor communicators. Pick up any science journal and you'll quickly discover that scientists either aren't good at telling stories or are discouraged to do so by the medium. The consequence is the same: papers read like a wall of text. TL;DR anyone? The quality of visual communication in general ranges from muddled to abysmal (watch video).

We need more leaders in the field, such as Nature Publishing Group, to reward and emphasize good visual communication (e.g. Nature Cancer Review 2013 Figure Calendar).

Our presentations concluded with a 15 minute moderated discussion with Sam Grobart, senior Businesssweek writer. Everyone got a little cheeky. Good fun.

presentation slides

This was a lightning 7 minute talk. I did more planning about what to say than I usually do, given that there was virtually no opportunity for any kind of backtracking, and include a running narrative below each slide.

1/32

My slides are available as PDF, keynote (zipped) or Quicktime. The format is 16:9 HD.

The reality of redesign is disruptive. How can we pursue new ideas and opportunities without leaving consumers confused or angry? Businessweek puts that question to some of the world's most accomplished designers. (Bloomberg Businessweek Design Issue)

On 28 Jan 2013, Bloomberg Businessweek Design Issue will capture the ideas from the conference and the personalities that generated them.

During the conference, each talk was captured in a series of sketches by Tom Wujec: my talk sketch and moderated discussion sketch.

Date completed: ongoing — an accurate assessment of the state of the visual communication field in science. (read article)
VIEW ALL

Essentials of Data Visualization—8-part video series

Mon 16-01-2017

In collaboration with the Phil Poronnik and Kim Bell-Anderson at the University of Sydney, I'm delighted to share with you our 8-part video series project about thinking about drawing data and communicating science.

Essentials of Data Visualization: Thinking about drawing data and communicating science.

We've created 8 videos, each focusing on a different essential idea in data visualization: encoding, shapes, color, uncertainty, design, drawing missing or unobserved data, labels and process.

The videos were designed as teaching materials. Each video comes with a slide deck and exercises.

P values and the search for significance

Mon 16-01-2017
Little P value
What are you trying to say
Of significance?
—Steve Ziliak

We've written about P values before and warned readers about common misconceptions about them, which are so rife that the American Statistical Association itself has a long statement about them.

This month is our first of a two-part article about P values. Here we look at 'P value hacking' and 'data dredging', which are questionable practices that invalidate the correct interpretation of P values.

Nature Methods Points of Significance column: P values and the search for significance. (read)

We also illustrate how P values can lead us astray by asking "What is the smallest P value we can expect if the null hypothesis is true but we have done many tests, either explicitly or implicitly?"

Incidentally, this is our first column in which the standfirst is a haiku.

Altman, N. & Krzywinski, M. (2017) Points of Significance: P values and the search for significance. Nature Methods 14:3–4.

Krzywinski, M. & Altman, N. (2013) Points of significance: Significance, P values and t–tests. Nature Methods 10:1041–1042.

Intuitive Design

Thu 03-11-2016

Appeal to intuition when designing with value judgments in mind.

Figure clarity and concision are improved when the selection of shapes and colors is grounded in the Gestalt principles, which describe how we visually perceive and organize information.

One of the Gestalt principles tells us that the magenta and green shapes will be perceived as as two groups, overriding the fact that the shapes within the group might be different. What the principle does not tell us is how the reader is likely to value each group. (read)

The Gestalt principles are value free. For example, they tell us how we group objects but do not speak to any meaning that we might intuitively infer from visual characteristics.

Nature Methods Points of View column: Intuitive Design. (read)

This month, we discuss how appealing to such intuitions—related to shapes, colors and spatial orientation— can help us add information to a figure as well as anticipate and encourage useful interpretations.

Krzywinski, M. (2016) Points of View: Intuitive Design. Nature Methods 13:895.

Regularization

Fri 04-11-2016

Constraining the magnitude of parameters of a model can control its complexity.

This month we continue our discussion about model selection and evaluation and address how to choose a model that avoids both overfitting and underfitting.

Ideally, we want to avoid having either an underfitted model, which is usually a poor fit to the training data, or an overfitted model, which is a good fit to the training data but not to new data.

Nature Methods Points of Significance column: Regularization (read)

Regularization is a process that penalizes the magnitude of model parameters. This is done by not only minimizing the SSE, $\mathrm{SSE} = \sum_i (y_i - \hat{y}_i)^2$, as is done normally in a fit, but adding to this minimized quantity the sum of the mode's squared parameters, $\mathrm{SSE} + \lambda \sum_i \hat{\beta}^2_i$.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Regularization. Nature Methods 13:803-804.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Model Selection and Overfitting. Nature Methods 13:703-704.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Classifier evaluation. Nature Methods 13:603-604.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Logistic regression. Nature Methods 13:541-542.

Model Selection and Overfitting

Fri 04-11-2016

With four parameters I can fit an elephant and with five I can make him wiggle his trunk. —John von Neumann.

By increasing the complexity of a model, it is easy to make it fit to data perfectly. Does this mean that the model is perfectly suitable? No.

When a model has a relatively large number of parameters, it is likely to be influenced by the noise in the data, which varies across observations, as much as any underlying trend, which remains the same. Such a model is overfitted—it matches training data well but does not generalize to new observations.

Nature Methods Points of Significance column: Model Selection and Overfitting (read)

We discuss the use of training, validation and testing data sets and how they can be used, with methods such as cross-validation, to avoid overfitting.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Model Selection and Overfitting. Nature Methods 13:703-704.