Video summary
The journey from traditional computer vision to deep learning began with biological inspiration in the 1950s when Hubel and Wiesel discovered how the brain processes visual information through hierarchical stages. This understanding led to the creation of the Neocognitron in 1980, an early artificial neural network that mimicked biological structures by using layers to detect simple features like edges before progressing to complex concepts. Although this model was biologically faithful and utilized Hebbian learning, it lacked practical efficiency for real-world applications. The field shifted focus in 1986 with the reinvention of backpropagation, which enabled neural networks to learn efficiently via gradient descent. This breakthrough gave rise to the LeNet family of architectures, the first modern convolutional neural networks designed specifically for images using convolution and pooling operations. While these early networks could handle simple tasks like digit recognition, they struggled with complex visual problems, leading to a period where handcrafted feature pipelines became the dominant approach again.
A significant resurgence occurred in the late 1990s and early 2000s driven by two critical advancements: hardware and data. In 1999, Nvidia launched the first GPU, providing the necessary parallel processing power, while the creation of ImageNet in 2009 introduced a structured dataset that fueled research momentum. These factors culminated in 2012 with the development of AlexNet, a convolutional neural network that leveraged GPUs, ReLU activations, and data augmentation to achieve state-of-the-art performance on image classification. This success marked the revival of deep learning by proving that networks could learn visual features directly from raw data, effectively replacing manual feature extraction methods like SIFT. Following this milestone, various architectures such as VGGNet, Inception, and ResNets further improved convolutional networks, establishing them as the standard for object detection, segmentation, and image generation throughout the 2010s.
As deep learning solidified its dominance in computer vision, a new era emerged with the introduction of transformer neural networks in 2017. Originally designed for language processing, transformers were adapted for computer vision tasks and demonstrated exceptional scalability when trained on large datasets. By the 2020s, these models became integral to the field, ushering in the age of foundation models that could handle a wide variety of visual problems with unprecedented efficiency. Ultimately, the transition from traditional computer vision to deep learning was defined by the moment neural networks became powerful enough to automatically learn complex visual features like edges and textures without human intervention. This shift allowed researchers to move away from decades of manually describing images using gradients and key points, marking the beginning of a new era where artificial intelligence could solve visual problems through direct learning rather than manual engineering.
Read the full video transcript
How did computer vision become deep
learning? To answer this question, we go
back to the 1950s with two researchers,
Hubel and Wiesel, who wanted to
understand how does the brain process
visual information. They conducted
experiments making a cat look at light
patterns and they monitored neural
activity. Based on their observations,
they proposed a hierarchical model of
visual processing and this suggests that
visual processing happens in a sequence
of stages. Based on this biological
model, Neocognitron was created in 1980.
And as such, this artificial neural
network had some properties that were
faithful to biology. Earlier layers had
neurons that detected simple features
like colors and edges. Later layers had
neurons that detected complex features
like abstract concepts.
The neuron also followed a layered
organization and this network learned
via Hebbian learning, which is a model
of how human beings also learn. Now,
shortly after in 1986, we had the final
reinvention of backpropagation.
Backpropagation is an efficient way for
neural networks to learn via gradient
descent. And this paved the way for the
LeNet family of architectures, which
were the first modern convolutional
neural networks, mostly because they
learned via backpropagation, which is
still used today. So, while biology
inspired Neocognitron,
we have practicality that inspired
LeNet. LeNet was structured and designed
explicitly to handle images with these
convolution and pooling operations and
they are hence very parameter efficient.
While this neural network could process
numbers and characters, it struggled
with fairly more complex computer vision
tasks.
So, this first convolution neural
network eventually died out. In the late
'90s to the early 2000s, we saw the rise
of handcraft feature pipelines, where we
extract features from an image manually
to perform downstream tasks.
Around this time, we had two major
changes happen. The first was hardware.
In 1999, Nvidia launched the industry's
first GPU.
The second was data. In 2009, with the
rise of unstructured data on the
internet, ImageNet created structure for
it in a knowledge ontology. This was
done to push forward research in
computer vision. And push forward
research it did. In 2012, we had the
creation of AlexNet. This was a
convolution neural network that could
perform well on image classification,
where images were more complex than
digits. And it did so with parallel
processing with GPUs, ReLU activations
for non-linearity, normalizations to
stabilize gradients, data augmentation,
and more. This network, trained with
backpropagation, was now the state of
the art for image classification,
beating out handcraft feature methods
like SIFT. And so started the revival of
convolution neural networks. Convolution
neural networks got better at image
classification. This was with VGGNet,
Inception, and ResNets.
Convolution networks became the state of
the art for object detection. This was
with architectures like R-CNN, faster
R-CNN, and YOLO.
Convolution networks became the state of
the art for segmentation.
And this was through fully convolution
networks, U-Nets, and Mask R-CNN.
Convolution networks also became the
soda for image generation. This was
through generative adversarial networks
and variational autoencoders, which were
using convolution neural networks as
their core components. And so in the
2010s, convolution neural networks were
used in most computer vision problems.
And then in 2017, we had the creation of
transformer neural networks. While
created for language, they incorporated
into computer vision, too. It was found
that these transformers scale well for
images when we have a large amounts of
training data. And so from the 2020s,
computer vision started incorporating
these transformers everywhere, and thus
ushering the age of foundation models,
too. So, if I were to summarize how did
computer vision become deep learning?
Well, computer vision became deep
learning when neural networks became
powerful enough to replace handcraft
features. For decades, the field tried
to describe images manually using edges,
corners, gradients, textures, and key
points. But once large data sets, GPUs,
and better training methods and CNN
architectures came together, deep
networks could learn those visual
features directly. And so began the age
of neural networks for solving computer
vision problems.
I hope all of this makes sense. Thank
you all so much for watching, and if you
like what you saw, do give this video a
like. Comment below what other kinds of
questions keep you curious in the field
of computer vision. Comment them below,
and I just might make a video on it.
Thank you all so much, and I'll see you
in the next one. Bye-bye.