Communicating risk using whole numbers
David Mackay's “Sustainable Energy – without the hot air” showed
the power of translating complex quantities into whole
numbers that can be added and easily compared. Similar
ideas have been tried in communicating risk, the units being
the Micromort for acute risks and the Microlife for chronic
risks. I shall show how these can be used to see how
recklessly dangerous, or how boringly safe, your life is.
The Global Calculator: Trying to apply 'Sustainable Energy – without the hot air' to the world
Could 9 billion people have a rich world standard and style
of living and yet avoid the worst effects of climate change,
if only we built enough nuclear power stations? or covered
enough desert with solar panels? or all used low energy
light bulbs? or efficient cars? Inspired by David
Mackay's “Sustainable Energy – without the hot air” a team
managed by Sophie Hartfield at the UK Department of Energy
and Climate Change tried to build an interactive calculator
to help people to figure out the answer. This talk will
cover their journey and the results of their attempt.
Modern electronic devices consist of a multitude of IC
components: the processor, the memory, the RF modem and the
baseband chip (in wireless devices), and the graphics
processor, are only some examples of components scattered
throughout a device. The increase of the volume of digital
data that needs to be accessed and processed by such devices
calls for ever faster communication between these IC's.
Faster communication, however, often translates to higher
susceptibility to various types of noise, and inevitably to
a higher power consumption in order to combat the noise.
This increase in power consumption is, for the most part,
far from linear. In this talk I will give a short overview
of problems encountered in chip-to-chip communication, and
will advocate the use of a novel class of codes, called
“Chordal Codes”, to solve those problems.
Codes for efficient data storage on DNA molecules
Over the past 6 months I have been working with David MacKay
and with Nick Goldman (EBI) on developing a coding system
for storing data efficiently on DNA molecules. In this talk,
I will focus on David’s initial approach to the problem,
which provides an information theoretic background for the
achievable storage density and reliability. The assumption
for this model is a pure packet loss model with noiseless
recovery of packets, which David called the “fountain
channel”. This supposes that lower level coding can fully
compensate for synthesis, sequencing and amplification
errors such as insertions, deletions and substitutions. An
overview of the literature on coding for channels with
insertions, deletions and substitutions reveals surprisingly
few publications, the majority of which co-authored by David
MacKay, and none of which providing unbounded reliability in
the information theoretic sense. Hence, I will discuss a
model for DNA storage that combines the packet loss with the
noisy nature of the recovered packets, and present coding
strategies for exploiting this channel that David, Nick and
I have been developing.
A traditional calculator evaluates symbolic mathematical expressions (such as √5)
and produces a decimal number (such as 2.2360679775).
An inverse calculator does the opposite: it starts from the number
and suggests how the number might have been calculated in the first place.
This talk explains how to build such an inverse calculator, and
why having one might be useful.
(This 10-minute surpise talk was not in the printed programme.)
Mind the Gap (between science and society) – the case studies of climate, nuclear and GMOs
Scientists often have a very different idea of risk and
benefit from policymakers and the general public. With
special attention to three case studies – climate change,
nuclear power and genetically modified crops – Mark Lynas
looks at the gaps between what scientists understand and
what everyone else thinks, and how these gaps can lead to
sub-optimal policy outcomes in the real world.
Why Medicine Needs Deep Learning
My research on deep inference and learning reaches back to
the wake-sleep algorithm, published in 1995, and the paper
that David MacKay and I wrote in 1996 showing that belief
propagation in graphs with cycles can be used for accurate
inference. Most of my time is now spent on genomic medicine.
Deep learning will transform medicine, but not in the way
that many advocates think. The amount of data times the
mutation frequency divided by the biological complexity and
the number of hidden variables is small, so downloading a
hundred thousand genomes and training a neural network won't
cut it. There is what I call a "genotype-phenotype gap" and
the value of closing this gap exceeds Google's $200B ad
market. I believe that the only way to bridge this gap is to
build machine learning systems that properly incorporate
biological knowledge. I'll describe this approach, which is
being pursued by dozens of young investigators, has improved
our ability to “read the genome”, and will, I believe, be an
indispensable component in the future of medicine.
New statistical approaches to disentangle single-cell diversity
Many key biological processes are driven by differences in
the regulatory landscape between single cells. Recent
technical developments have enabled the transcriptomes and
epigenomes of hundreds of cells to be assayed in an unbiased
manner, opening up the possibility to identify and study new
physiologically relevant, sub-populations of cells. In this
talk I will discuss statistical advances to elucidate the
factors that drive single-cell heterogeneity. Accurate and
scalable latent variable models allow dissecting single-cell
transcriptome and epigenome studies, thereby disentangling
biological variation from technical and confounding factors.
I will illustrate these approaches in applications to large
datasets consisting of tens of thousands of cells and discuss
their relationship to methods from population genetics.
New statistical methods for the analysis of genome variation data
The last two decades have seen an exponential growth in the
quantity of DNA sequencing data, at a rate more than twice
as fast as Moore’s law for growth in computing power.
Individual sequences at one position in the genome are
related by descent from a common ancestor in a tree, but
this genealogy is made much more complex across the whole
genome by recombination, which means that you inherit
different sections of your genome from different recent
ancestors. A clean mathematical model exists for the full
genealogy but it is too complex to use for inference with
data sets available today. I will discuss various
algorithmic and representational innovations introduced to
model and exploit genetic relatedness in analysing ever
larger amounts of genome data, illustrating them with our
work on the Haplotype Reference Consortium, which has
assembled over 65,000 whole genome sequences with data at
nearly 40 million variable sites.
Emergence, dynamics, and behaviour
How do the principles of psychology relate to the activities
of 1011 nerve cells in the human brain? Large physical
systems generally have emergent collective behaviours. Real
computers—the brain included—compute by following a change
of computer ‘state’ with time. When that change of state
can be described by collective variables, robust dynamics
emerge whose mathematical description can be entirely unlike
that of the underlying microscopic variables. We illustrate
the useful emergent computational dynamics of two elementary
neural networks. The first deals with the problem of
variable cadence when recognizing dynamical sensory patterns
spread over time. The second achieves goal-directed
behaviour when multiple candidate goals are simultaneously
Neural substrates of decision-making in rats
The most common behavioral observation in decision-making,
experienced both in our daily lives and in laboratory
settings, is that easy decisions (where we are likely to
choose the correct response) are done quickly, whereas
difficult decisions (less likely to choose correctly) are
much slower. An appealingly simple model was proposed in the
behavioral literature many decades ago to account for this
observation. This model, sometimes known as the “gradual
accumulation of evidence” model, has been used to explain
many behavioral data sets. Does the brain implement
something well approximated by this model? If so, how does
the brain’s network of neurons actually carry out the
implementation? We have been using using a combination of
computational and experimental approaches with rats to try
to answer these questions.
Decoding the population activity of grid cells for spatial localization and goal-directed navigation
Mammalian grid cells discharge when an animal crosses the
points of an imaginary hexagonal grid tessellating the
environment. I will show how animals can navigate by reading
out a population vector of such activity patterns across
multiple spatial scales. The theory explains key
experimental results about grid cells, makes testable
predictions for future physiological and behavioural
experiments, and provides a mathematical foundation for the
concept of a "neural metric" for space. For goal-directed
navigation, the proposed allocentric grid cell
representation can be readily transformed into the
egocentric goal coordinates needed for planning movements.
Joint work with Martin Stemmler and Alexander Mathis.
Next-generation text entry
Per Ola Kristensson
Text entry is a common everyday computing task. However,
despite its ubiquitousness it is difficult to devise an
efficient text entry method that users are willing to adopt.
In this talk I will explain the narrow design space of text
entry research and make the case that successful
next-generation text entry methods are likely to be based on
designs that merge behavioural solution principles with
information engineering techniques. I will exemplify this
principle with several new text entry methods we have
developed for a variety of use-cases.
How Dasher has touched lives
Dasher is an information-efficient text-entry interface, driven by
natural continuous pointing gestures. Dasher is a competitive
text-entry system wherever a full-size keyboard cannot be used.
(This 7-minute surprise talk was not in the printed programme.)
Democratising data science
There has never been a higher demand for data science and
data scientists. In science, in medicine, in business, in
sport, in every area of life, we now have the tools to
accumulate large quantities of data, but we lack the
corresponding tools to process and understand it. I will
talk about work that we have done in Microsoft Research to
try to build such tools, the challenges that we have
encountered and suggest how data science might be made much
more broadly accessible in the future.
Pseudo-Marginal Slice Sampling
This talk is about sampling from probability distributions
in a challenging setting: we can only compute noisy (but
unbiased) estimates of the probability density function at
individual settings of variables we choose. Previous work on
such "pseudo-marginal" methods have been used in inference
problems in genetics, continuous time stochastic processes,
hierarchical models, and "doubly-intractable" distributions.
I'll present algorithms that are easier to apply than
previous work and sometimes work a lot better.
Joint work with Matt Graham.
Can sensory cortex do backpropagation?
Stochastic gradient descent in multilayer networks of
neuron-like units has led to dramatic recent progress in a
variety of difficult AI problems. Now that we know how
effective backpropagation can be in large networks it is
worth reconsidering the widely held belief that the cortex
could not possibly be doing backpropagation. Drawing on
joint work with Timothy Lillicrap, I will go through the
main objections of neuroscientists and show that none of
them are hard to overcome if we commit to representing error
derivatives as temporal derivatives. This allows the same
axon to carry information about the presence of some feature
in the input and information about the derivative of the
cost function with respect to the input to that neuron. It
predicts spike-time dependent plasticity and it also
explains why we cannot code velocity by the rate of change
of features that represent position.