ArcFace Loss One of tha main challenges in feature learning using Deep Convolutional Neural Networks for large scale face recognition is designing the optimal loss function that enhance discriminative power. Before this paper, there were mainly 2 lines to train…
Deep Metric Learning Let’s say we want to create a model that can do face recognition(Face Identification and face verification). We CAN use traditional deep learning since it can perform really well but it requires a lot of training data.…
SILU SiLU is proposed as an activation function for neural network function approximation in reinforcement learning, and DSiLU is the derivative function for SiLU. DSiLU is a steeper and “overshot” version of the sigmoid function and it is proposed as…
Rubik’s Cube Due to the annotation of 3D medical data being hard to acquire, the number of annotated 3D images for training is often not enough. Self-supervised learning deeply exploiting the information of raw data can be a solution to…
Data Imbalance Data-Leveling is used when your training data is imbalanced. Methods Upsample: This is where you add new instance for the minority class to balance the data. Downsample: This is where you delete data from the majority class to…
These are the podcasts that are helping me learn AI. 1. “AI Today” by Cognilytica 2. “Super Data Science” by John Krohn 3. “Data Skeptic” by Kyle Polich 4. “Practical AI” by Changelog Media
KL Divergence KL Divergence measures the distance between probability distributions. This is used in various places such as the cross-entropy loss or as a loss function in VAE where you want to constrain the latent distribution to a standard distribution.…
These are the online courses that helped me get a job as an AI engineer from a residential architect. Udemy Machine Learning A-Z™: Hands-On Python & R In Data Science Artificial Intelligence A-Z™: Learn How To Build An AI Deep…
PVTv2 The previous PVT had mainly 3 limitations. When processing high-resolution images, the computational cost is still relatively high. Loses local continuity of the image because it treats the image as a sequence of non-overlapping patches. Inflexible for arbitrary image…
DETR Modern object detectors predict a set of bounding boxes and category labels for each object of interest by defining surrogate regression and classification problems on a large set of proposals. This means that their performances heavily rely on post-processing…