I was taught that serious academics favored Support Vector Machines over Neural Networks, which industry only loved because they didn’t have proper education. oops…
also, Computer Vision was considered “AI-complete” and likely decades away. ImageNet dropped a couple years I graduated. though I guess it ended up being “AI-complete” in a way…
Before AlexNet, SVMs were the best algorithms around. LeNet was the only comparable success case for NNs back then, and it was largely seen as exclusively limited to MNIST digits because deep networks were too hard to train. People used HOG+SVM, SIFT, SURF, ORB, older Haar / Viola-Jones features, template matching, random forests, Hough Transforms, sliding windows, deformable parts models… so many techniques that were made obsolete once the first deep networks became viable.
The problem is your schooling was correct at the time, but the march of research progress eventually saw 1) the creation of large, million-scale supervised datasets (ImageNet) and 2) larger / faster GPUs with more on-card memory.
It was fact back in ~2010 that SVMs were superior to NNs in nearly every aspect.
HOG and Hough transforms bring me back. honestly glad that I don’t have to mess with them anymore though.
I always found SVMs a little shady because you had to pick a kernel. we spent time talking about the different kernels you could pick but they were all pretty small and/or contrived. I guess with NNs you pick the architecture/activation functions but there didn’t seem to be an analogue in SVM land for “stack more layers and fatten the embeddings.” though I was only an undergrad.
do you really think NNs won purely because of large datasets and GPU acceleration? I feel like those could have applied to SVMs too. I thought the real win was solving vanishing gradients with ReLU and expanding the number of layers, rather than throwing everything into a 3 or 5-layer MLP, preventing overfitting, making the gradient landscape less prone to local maxima and enabling hierarchical feature extraction to be learned organically.
No, you are correct. Hinton began researching ReLUs in 2010 and his students Alex Krizhevsky and Ilya Sutskever used it to train a much deeper network (AlexNet) to win the 2012 ILSVRC. The reason AlexNet was so groundbreaking was because it brought all of the gradient optimization improvements (SGD with momentum as popularized by Schmidhuber, and dropout), better activation functions (ReLU), a deeper network (8 layers), supervised training on very large datasets (necessary to learn good general-purpose convolutional kernels), and GPU acceleration into a single approach.
NNs, and specifically CNNs, won out because they were able to create more expressive and superior image feature representations over the hand-crafted features of competing algorithms. The proof was in the vastly better performance, it was a major jump when the performance on the ILSVRC was becoming saturated. Nobody was making nearly +10% improvements on that challenge back then, it blew everybody out of the water and made NNs and deep learning impossible to ignore.
Edit: to accentuate the point about datasets and GPUs, the original AlexNet developers really struggled to train their model on the GPUs available at the time. The model was too big and they had to split it across two GPUs to make it work. They were some of the first researchers to train large CNNs with GPUs. Without large datasets like the ILSVRC they would not have been able to train good deep hierarchical convolutions, and without better GPUs they wouldn’t have been able to make AlexNet sufficiently large or deep. Training AlexNet on CPU only for ILSVRC was out of the question, it would have taken months of full-tilt, nonstop compute for a single training run. It was more than these two things, as detailed above, but removing those two barriers really allowed CNNs and deep learning to take off. Much of the underlying NN and optimization theory had been around for decades.
I was taught that serious academics favored Support Vector Machines over Neural Networks, which industry only loved because they didn’t have proper education. oops…
also, Computer Vision was considered “AI-complete” and likely decades away. ImageNet dropped a couple years I graduated. though I guess it ended up being “AI-complete” in a way…
Before AlexNet, SVMs were the best algorithms around. LeNet was the only comparable success case for NNs back then, and it was largely seen as exclusively limited to MNIST digits because deep networks were too hard to train. People used HOG+SVM, SIFT, SURF, ORB, older Haar / Viola-Jones features, template matching, random forests, Hough Transforms, sliding windows, deformable parts models… so many techniques that were made obsolete once the first deep networks became viable.
The problem is your schooling was correct at the time, but the march of research progress eventually saw 1) the creation of large, million-scale supervised datasets (ImageNet) and 2) larger / faster GPUs with more on-card memory.
It was fact back in ~2010 that SVMs were superior to NNs in nearly every aspect.
Source: started a PhD on computer vision in 2012
HOG and Hough transforms bring me back. honestly glad that I don’t have to mess with them anymore though.
I always found SVMs a little shady because you had to pick a kernel. we spent time talking about the different kernels you could pick but they were all pretty small and/or contrived. I guess with NNs you pick the architecture/activation functions but there didn’t seem to be an analogue in SVM land for “stack more layers and fatten the embeddings.” though I was only an undergrad.
do you really think NNs won purely because of large datasets and GPU acceleration? I feel like those could have applied to SVMs too. I thought the real win was solving vanishing gradients with ReLU and expanding the number of layers, rather than throwing everything into a 3 or 5-layer MLP, preventing overfitting, making the gradient landscape less prone to local maxima and enabling hierarchical feature extraction to be learned organically.
No, you are correct. Hinton began researching ReLUs in 2010 and his students Alex Krizhevsky and Ilya Sutskever used it to train a much deeper network (AlexNet) to win the 2012 ILSVRC. The reason AlexNet was so groundbreaking was because it brought all of the gradient optimization improvements (SGD with momentum as popularized by Schmidhuber, and dropout), better activation functions (ReLU), a deeper network (8 layers), supervised training on very large datasets (necessary to learn good general-purpose convolutional kernels), and GPU acceleration into a single approach.
NNs, and specifically CNNs, won out because they were able to create more expressive and superior image feature representations over the hand-crafted features of competing algorithms. The proof was in the vastly better performance, it was a major jump when the performance on the ILSVRC was becoming saturated. Nobody was making nearly +10% improvements on that challenge back then, it blew everybody out of the water and made NNs and deep learning impossible to ignore.
Edit: to accentuate the point about datasets and GPUs, the original AlexNet developers really struggled to train their model on the GPUs available at the time. The model was too big and they had to split it across two GPUs to make it work. They were some of the first researchers to train large CNNs with GPUs. Without large datasets like the ILSVRC they would not have been able to train good deep hierarchical convolutions, and without better GPUs they wouldn’t have been able to make AlexNet sufficiently large or deep. Training AlexNet on CPU only for ILSVRC was out of the question, it would have taken months of full-tilt, nonstop compute for a single training run. It was more than these two things, as detailed above, but removing those two barriers really allowed CNNs and deep learning to take off. Much of the underlying NN and optimization theory had been around for decades.