Pretrained Models in CNN: ImageNet, AlexNet, and the Rise of Transfer Learning

What Are Pretrained Models ?

Pretrained models are neural network models that have already been trained on a large and diverse dataset, typically using significant computational resources. These models learn general and reusable feature representations, especially useful in image-based tasks. Instead of training a CNN from scratch, we can reuse these learned weights and adapt the model to our own problem.

Why Use Pretrained Models ?

When we work with CNNs, we usually deal with image-based problems. Deep learning models are highly data-hungry—to perform well, they require a large amount of data. In image classification tasks, this means a large number of images, and most importantly, labeled data.

For example, if we are building a cat vs dog classifier, we need many images of cats and dogs. While these images can be collected from sources like Google Images, they do not come labeled by default. We must manually label each image to indicate whether it is a cat or a dog. This manual labeling process is time-consuming, tedious, and costly, especially at scale. This is one of the primary reasons for using pretrained models.

The second major reason is training time. Training a CNN from scratch on large image datasets takes a significant amount of time and computational power, and the training process can be very slow. By using pretrained models, we can avoid this heavy training phase and instead build on top of an already trained model, making development faster and more efficient.

In short, pretrained models help us save data, time, and cost, while still achieving strong performance on image-based tasks.

ImageNet Dataset

ImageNet is a large-scale visual database of images created to support research in computer vision and deep learning. Around 2006, computer vision researcher Fei-Fei Li observed that most researchers were heavily focused on developing models and algorithms, while high-quality, large-scale datasets were still missing. She believed that, in the future, strong models would require equally strong datasets containing well-structured and meaningful visual data.

With this vision, she initiated the creation of the ImageNet database. During this process, she collaborated with a professor involved in building WordNet, which helped organize visual concepts in a hierarchical manner. Together, they built a dataset containing around 14 million images spanning approximately 20,000 categories. These categories represent everyday objects such as cats, vehicles, tables, chairs, and many others.

A key strength of ImageNet is its well-organized labeling structure. Images are annotated with clear visual descriptions and arranged hierarchically based on object relationships. In addition to classification labels, around one million images include bounding box annotations, where a box is drawn around the object of interest. This type of labeling is particularly useful for tasks such as object localization, as it provides information about both what the object is and where it appears in the image.

To achieve labeling at this scale, the ImageNet team relied on crowdsourcing. They collected annotations by asking large numbers of people to identify objects in images and mark their locations. For this purpose, they used a service provided by Amazon called Amazon Mechanical Turk.

The ImageNet dataset played a pivotal role in advancing deep learning and computer vision. By providing a large, diverse, and well-labeled dataset, it enabled the training of powerful deep learning models and ultimately changed the direction and future of deep learning research.

ILSVRC (ImageNet Large Scale Visual Recognition Challenge)

After the ImageNet database was built, researchers decided to use it as the foundation for a large-scale competition to benchmark image recognition models. This competition became known as ILSVRC, also commonly called the ImageNet Challenge. It started in 2010, with the primary goal of identifying and comparing the best-performing image classification models.

The dataset used in ILSVRC is a subset of the original ImageNet dataset. While ImageNet contains tens of millions of images across many categories, the challenge dataset consists of around one million images limited to 1,000 classes, which reduces complexity and makes fair comparison possible. Many research teams from around the world participated in this competition.

In the initial years (2010–2011), most winning approaches were based on traditional machine learning techniques. These methods relied heavily on manual feature extraction, such as SIFT and HOG features, followed by classifiers and ensemble techniques. In 2010, the winning model achieved an error rate of around 28%, meaning it misclassified 28 out of every 100 images. In 2011, this error rate improved slightly to about 25%.

A major breakthrough came in 2012, which is widely considered a landmark year for deep learning. In this year, a deep learning–based CNN model called AlexNet was introduced by researchers led by Geoffrey Hinton. This was the first time a deep learning model dominated the competition. AlexNet was trained using GPUs instead of CPUs, which significantly accelerated training, and it also popularized the use of ReLU activation functions in deep neural networks.

AlexNet reduced the top-5 classification error to around 16%, an improvement of more than 10 percentage points over the previous year. This dramatic performance gain brought a revolution in the field. The global research community began to focus heavily on CNNs and deep learning, recognizing their superiority for image recognition tasks.

In the years that followed, increasingly advanced CNN architectures continued to dominate the ILSVRC leaderboard, firmly establishing deep learning as the standard approach for computer vision problems.

AlexNet

In the ImageNet dataset, input images are 227 × 227 × 3 RGB images. AlexNet processes these images through a deep convolutional neural network designed for large-scale image classification.

In the first convolutional layer, 96 filters of size 11 × 11 are applied with a stride of 4. This is followed by a max-pooling layer with a 3 × 3 window and stride 2, which reduces spatial dimensions while retaining important features. As the network goes deeper, convolution and pooling layers progressively extract more complex and abstract features from the images.

At the end of the network, AlexNet uses three fully connected layers. The first has 9,216 neurons, followed by two layers with 4,096 neurons each. Finally, a softmax output layer with 1,000 units is used, corresponding to the 1,000 ImageNet classes, making this a multi-class classification problem.

AlexNet was the winning model of the 2012 ImageNet challenge and marked the beginning of the modern deep learning revolution. Its strong performance demonstrated the power of deep CNNs and shifted the global research focus toward deep learning for computer vision tasks.

Famous Architectures

After AlexNet, CNN architectures improved rapidly year by year. In 2013, ZFNet reduced the error rate to around 11.7%. In 2014, VGGNet further improved performance, achieving roughly 7.3% error. This was followed by GoogLeNet, which brought the error down to about 6.7%.

A major milestone came in 2016 with ResNet, which achieved an error rate of around 3.5%. This was remarkable because human-level performance on the same dataset is estimated to be around 5% error, meaning ResNet performed better than humans. Over time, as networks became deeper and more sophisticated, model complexity increased and error rates consistently decreased, leading to repeated wins in the ImageNet challenge.

Idea of Pretrained Models

Training deep learning models from scratch requires large amounts of data, which is expensive to collect and label. Even when sufficient data is available, training deep networks takes significant time and computational resources. Researchers realized that if highly effective models such as VGG and ResNet are already trained on massive datasets like ImageNet, it is often unnecessary to train a new model from the ground up.

Instead, we can reuse these pretrained models directly. This provides two major advantages. First, we do not need to collect and label large datasets from scratch. Second, we save a substantial amount of training time because the model has already learned rich and general visual features. Modern deep learning frameworks such as Keras provide many such pretrained models, making it easy to apply them to real-world problems.