text
stringlengths
0
820
that domain knowledge and model assumptions are of
prime importance and that models must be challenged
1) to respect the reality of the physical/biological/chemi -
cal processes governing the system under study and 2)
to be accountable—by the transparency of their internal
reasoning—for the decisions to which they lead. This is
important, especially when models are intended to be
used for actual decision making and can affect balanc -
es of power or society-changing decisions. Later in this
article, we present ideas that are aligned with these di -
rections (see Table 1 ) and centered on the injection of
domain knowledge, but for different purposes. First we
discuss physics-aware ML, which has the goal of using
domain knowledge to restrict the solution spaces of the
models so that the outcome is physically plausible (di -
rection 4 in Table 1 ). This will ensure that the physical
consistency of the solutions is maintained while avoid -
ing aberrant outcomes that break physics (e.g., mass and
energy conservation).
We then discuss how best to obtain human-under -
standable interpretations and explanations of the inner functioning of the models, to understand why and how
models make decisions (direction 5 in Table 1 ). This has the
advantage of making the model trustable and nonfalsifi -
able and of avoiding situations where the right conclusions
are reached for the wrong reasons.
Explainability also enhances the potential for testing nov -
el hypotheses and acquiring new scientific knowledge from
the analysis of the model’s functioning. Directions 4 and 5
can be combined: they use domain knowledge in various
ways with different goals. The transparency of the models’
weights is not absolutely necessary at this stage, as interpre -
tations can be achieved by the analysis of inputs (e.g., Local
Interpretable Model-agnostic Explanations, i.e., LIME [ 23])
and physics awareness realized by modified loss functions.
Yet, as mentioned previously, science is about under -
standing the world in which we live, not just approximat -
ing it. We argue that, without learning causal relationships
from observational data and assumptions, this ambitious
goal of understanding the Earth system will not be possible
(direction 6 in Table 1 ). In this case, learning of cause-and-
effect relationships is a mix of the previous ingredients, as
domain knowledge is needed to design the model in such
a way that it can reveal (maybe novel) cause-and-effect
relationships that can be then explained using domain
TABLE 1. A SUMMARY OF THE SIX RESEARCH DIRECTIONS PRESENTED IN THIS ARTICLE.
DIRECTION IN A NUTSHELL REFERENCES CURRENT ISSUES 10 YEARS FROM NOW
1 Go beyond recognition
toward induction, deduction,
spatial and temporal
reasoning, and structural
inference.[6], [7] Missing or very limited
benchmarks and novel tasks as
well as reasoning models; the
interpretability is unsolved.Intelligent systems linking meaningful
transformation of entities, e.g., over space
or time, and deriving knowledge as the
way people understand the visual world
and its processes.
2 Think beyond the raster and
consider all the possible
inputs and sources of supervi -
sion, in particular, geotagged
social media data. [8]–[10] The presence of data set biases
and of label noise; a spatiotem -
poral mismatch between data
sources and scalability, with an
increasing number of sources. Systems that use a wide variety of sources
to enable a fine-grained understanding of
the world, all with minimal human effort
required for data set building and system
design.
3 Query the world by asking
questions about images and
create descriptions. [11], [12] Simplistic language model,
limited choice of thematic
interactions, and a lack of
large-scale infrastructure. Visual search engines have an understand -
ing of questions about images and are able
to adapt to different types of requests and
are usable for everyone.
4 Make models learned
using deep neural networks
consistent with domain-
specific knowledge, like
equations from physics.[1], [13]–[16] Networks’ outputs are not
physically consistent; networks
are often used as emulators of
simulations but do not explore
beyond current simulators’
constraints: they cannot
discover new physical rules.Systems trainable with much fewer data
because they constrain output space
via physical knowledge; systems that
learn a new hypothesis for new science
generation.
5 Enhancing interpretability
and explainability to
understand processes in ML
models in a better way. [17], [18] A lack of human-understand -
able interpretation, with a