The answer depends for which type of work you want to deploy a neural network for.
For tasks involving images like classification, detection etc- you are to use convolutional neural network based models so that you will use transfer learning with models like VGG-Net, InceptionV3, XceptionNet etc.
On the other hand, for tasks involving texts or time series, the concept is different where you need to use Recurrent Neural Network or LSTM based models. So use transfer learning like models like BERT, XLNet,RoBERT etc.
Most researchers suggests CNN, see e.g., Article Deep Convolutional Neural Networks for Computer-Aided Detect...
and https://pdf.sciencedirectassets.com/271475/1-s2.0-S0950061817X0024X/1-s2.0-S0950061817319335/Kasthurirangan_Gopalakrishnan_Pavement_cracking_2017.pdf?X-Amz-Security-Token=IQoJb3JpZ2luX2VjEHUaCXVzLWVhc3QtMSJHMEUCICCU2lFUHNpsR8m98SbtUp3HgAxfjx0xypWvnJxjgNN7AiEA9W8EY3BjQ1RRZVophSYfS9Il9ARYHPHtPUFIk6UlmykqgwQIjv%2F%2F%2F%2F%2F%2F%2F%2F%2F%2FARAEGgwwNTkwMDM1NDY4NjUiDMqW0uEPhTqldOeAsyrXA0KxoJ6GfrfcOOJ3W%2FiPzY5n6t0XAJstHnX9kgZ8nkERA%2FyFoNWcrYDenOwSVRFmezttiUz3O%2BZniSZwsiU3uDlcqU6cWnrZ073aBHKbBevFUj3qCQzzSsGdI1CBvbztvxZnX%2FY71Jd6vREqJSF%2BSMLrwp1Rj8fxDuan%2BsAOSmONAhE7aYbrcBmUjnuJHecxY0KVVlUytxgmUCuyQTnojhqKPvZTAmMEbXYd9gEY%2BG888n5x8MtwANzPxzqmCzNAO3eeUY1UrTKNk0J3IP58TWSbXkQFl%2F4K92xBjkfdwgcYJ8PNiHDnQJuXEzdpSsfYAOBoRslOviESv1HRcGE%2BxZprxlwCS%2FQM07SwUygiALno%2Bn5v6C%2BXuCnQiyaAGpVFIcBpRO5ue46fYTR3TCo71fsvl9C2sDXfVBBB84gHJiAvNNxZicOn0aUgj4WrhltRe6y1%2F9SASJx73UFcrZm2kT9k3FRBAYCLXQ%2BVwF19Y1FbVQ8tEhuTKUuU9t80N4nNNenB5O%2FPscAooBTeIyGjdhfC9HxUA9PXEgE%2FpEnj%2FxejBqOw5rLvnMCLXt0rynmWelcndsxRjk0KmV2OTHv%2BsetvlhzUunRV3az%2F9GUsEVcccTrJnu76yDCAiYuPBjqlAZhuJVf2RTiA6axAGE2V9W4GGBU7%2BrwvpfqZjdSxW3btxY4BmdEDECfLY%2FvqvXVZxABGcAu5VzRhf4YL%2FN6UnQHyA1VjZXGm9rSeKteYlZY%2B6RJD4XkNJy9eAi1fNyBvjRZE6SWNY%2FMm3N2DFLCBkRBJdhhMdhzpg64kYWc8lUMau5bW1A9Rq1A3lemow6tuGCnXLtAYe5PHTGiij2%2BaKwjtSnInSA%3D%3D&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Date=20220115T133111Z&X-Amz-SignedHeaders=host&X-Amz-Expires=300&X-Amz-Credential=ASIAQ3PHCVTYY4KZASYN%2F20220115%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Signature=583142a8f797482bf38ce47d939b4d23b2f9c5efe9675221251b0688f88baec1&hash=ec1b1cab79302971b0cad2961c2366b26089d8dae910fcceba5f8d9d9e0cc969&host=68042c943591013ac2b2430a89b270f6af2c76d8dfd086a07176afe7c76c2c61&pii=S0950061817319335&tid=pdf-3e98eecb-95e1-4858-acc3-2f3f92577f17&sid=dc23956997a8d1489a89dc15077cd91986fegxrqa&type=client
Transfer learning models focus on storing knowledge gained while solving one problem and applying it to a different but related problem. Instead of training a neural network from scratch, many pre-trained models can serve as the starting point for training.
I am sending a paper, hoping it will be useful to you, it is a comparison between pre-trained architectures (VGG16, VGG19, ResNet101, ResNet152, ResNet50, MobileNet, MobileNetV2, Xception, NASNetMobile, NASNetLarge, DenseNet121, DenseNet169 y DenseNet201) vs training from scratch. The experiments were carried out on CT images.
Conference Paper A machine learning approach for the feature extraction of pu...
It total depends on your type of data you are using; because every models has its traits and benifits as well. Pretrained CNN models give upto 90 to 95 % accuracy.
Simpler Models :
AlexNet
VGG16
VGG19
GoogLeNet
Models that gives Higher Accuracy: image classification and object detection
ResNet50
Inception-v3
Densenet-201
Xception etc.
Models for Edge Deployment: low-compute, low-power devices
In one sentence, The models are large and have a nice mathematical understanding inside to extract the features after full training. Suppose, the resnet50 is already trained in ImageNet if you want to use a transfer learning approach that should give a better result as the network already has the understanding that how it can extract some meaningful features as it is already trained. That is why it's easy and the adaptability of the neural network models are also useful in various domains.
Transfer learning means taking the relevant parts of a pre-trained machine learning model and applying it to a new but similar problem. This will usually be the core information for the model to function, with new aspects added to the model to solve a specific task.
Chapter Transfer Learning with Convolution Neural Networks Models: A...
Article Transfer learning for neural network model in chlorophyll-a ...