I think the best
metaphor one can use and that I use quite a
lot, is thinking of those three terms being
put together as if they were in the
shape of an onion where AI is one of
the outer layers. Machine learning is kind of
the middle layer and then deep learning as the
most internal layer. Machine learning here
with a focus on learning relates to the machine or the
computer or the algorithm. Having to learn and
understand the data, the distribution of the
data and the characteristics of the dat
a. Deep learning is,
in fact, a family of algorithms within machine
learning that is characterized by the use of neural networks
as the main data structure that is used in order to train and
later on infer on the data. We need to have domain experts
that understand the data, understand the problem
domain, look at the data, inspect the data, understand
its distribution, and most importantly, decide what are
the most important features or characteristics that lie
within the data that are relevant
to the problem domain. Then we would need, once
those are identified, to extract those features
from the data, quantized them, meaning
digitize them somehow and then input those into
a vector of features. That vector of features
is a representation of those most important
features and characteristics of each sample of the data. Once we've extracted all the
features from all data samples that we have, we can take
the collection of vectors that we've created. And then feed those
vectors into the m
achine learning model of our choice. The fact that deep learning
allows us to train directly on the raw data
lies at the heart of the differentiation between
deep learning and machine learning. The most important
characteristics that lie within the data
and generalize on its own, on the most relevant
and important features for the solution that we
seek, or the decision that we want to have the
model perform for us. In deep learning, we simply
need to take the raw data and feed that into our deep
learning model in order for it to infer on it and make
a decision in machine learning. On the other hand,
just like we do in training before
we do inference, we need to extract and compute
features, organize those into a vector and that have
that Fed into our machine learning algorithm. That extraction and computation
and factorization of features is something that takes
up more time and resources in terms of computational
complexity, CPQ, RAM usage, et cetera having to do
manual feature extrac
tion is extremely counter
intuitive when we're talking about a
problem domain, which is innately adversarial. Sometimes, even if we have the
best cybersecurity experts, researchers, the best malware
analyst and reverse engineers, we don't even know
and understand all the different features
and characteristics that are found within the
data regarding known attack techniques, mutation
techniques and evasion methods, let alone in an area
like cybersecurity, where there are always
zero day exploitat
ion and vulnerabilities that
are found out there. If we're not aware of those,
we can possibly, as humans, extrapolate and
understand what are the relevant features
for those that may be found within the data. Deep learning algorithms allow
us to train on all of the data that's available to us rather
than just a fraction of it. It allows us to truly and
fully harness the power of the machine in order
to generalize on features and the correlations between
those features in a way that we as humans
simply cannot
do or do not understand. It allows us to create
models that consume less time and do it with less
computation resources, and it allows us to react
faster and better to a quickly, rapidly changing
threat landscape. By using deep learning,
we can deliver models that have much higher quality. It means we can deliver models
with a much higher detection rate and a considerably
lower false positive rates. These are models that can
be used for pre execution prevention across our entire
p
roduct offering and installed base.
Comments