Hello :) Today is Day 238!
A quick summary of today:
- chapter 10 of the Machine Learning Algorithms in Depth book
- rehearsal for the KB AI competition
Chapter 10: Fundamental DL algorithms
This chapter introduced just basic concepts
- Multilayer perceptron consists of multiple densely connected layers followed by a nonlinearity
- Cross-entropy loss is used in classification tasks, while mean squared error loss is used in regression tasks
- Neural networks are optimized via a backpropagation algorithm, based on the chain rule
- Increasing model capacity to avoid underfitting can be achieved by changing the model architecture (e.g., increasing the number of layers and hidden units per layer)
- Regularization to avoid overfitting can occur on multiple levels: weight decay, early stopping, and dropout
- Convolutional neural nets work exceptionally well for image data and use convolution and pooling layers instead of vectorized matrix multiplications
- Classic CNN architectures consist of a convolutional layer followed by ReLU nonlinear activation function and a max pooling layer
- Pretrained CNNs can be used to extract feature vectors for images and used in applications such as image search
- Recurrent neural nets are designed to process sequential data
- Application of RNNs include language generation, sequence classification, and sequence translation.
- Multi-input neural nets can be constructed by concatenating the feature vector representations from individual input branches.
- Neural network optimizers include stochastic gradient descent, momentum, Nesterov momentum, AdaGrad, RMSProp, and Adam.
KB AI competition rehearsal
Here is a picture of the room - Samsung Hall in Ehwa Women University in Seoul.
We arrived at the university at around 3.30pm and until the rehearsal we just walked around, and my project partner practiced his delivery.
The rehearsal started officilly at 5.30 and lasted until around 9pm. Each team did their presentation and my initial thoughts are that we are the most unique. While I have bias as a competitor, almost all the other projects revolve around using an LLM for some text assistance - from that point of view we are very unique and different in our topic, tools and overall project goal .
We will see tomorrow ^^
That is all for today!
See you tomorrow :)