Home | Blog | Wiki | About

• Transfer learning: 'storing knowledge gained while solving one problem and applying it to a different but related problem. For example, knowledge gained while learning to recognize cars could apply when trying to recognize trucks'
• Reinforcement learning:
  • 'focus is on finding a balance between exploration (of uncharted territory) and exploitation (of current knowledge).'
  • No need of labelled data or explicit correction. Sort of learns by itself based on what it sees.
  • 'concerned with how intelligent agents ought to take actions in an environment in order to maximize the notion of cumulative reward'
• ReLu: Rectified linear unit
  • A popular activation function in ANNs
  • f(x) = max(0, x)
• Meta-learning: 'Learning to learn'
  • https://lilianweng.github.io/posts/2018-11-30-meta-learning/
• Support vector machine (SVM)
  • For (linear) classification problem
  • Supervised training
• Perceptron (McCulloch-Pitts)
• Linear classification
  • Points of different classes drawn on a graph can be demarcated by straight lines
• Support vector machines (SVM): Can do linear classification

• Forecasting algorithms
  • ARIMA (Auto-regressive integrated moving average)
  • SARIMA (Seasonal auto-regressive integrated moving average)
• Autoregression
• Moving average

• Ablation study
  • Ablation => removing a component of model to see how it performance varies
  • Useful to gauge the necessity/contribution of a component/architecture/parameter.
  • Ablation study could be done to show that a newly inserted aspect does indeed make the model better.
    • The model with and without this new aspect can be made, to compare and show that the performance is better with the new aspect.
  • Eg: Changing activation function
• Continuous learning
  • Train a model, but the training never stops.
  • Allow model to learn new things by constantly retraining for it to be able to deal with new kinds of data
  • Without losing the knowledge gained from previous trainings
  • Allowing the model to adapt to changing need, without being retrained from scratch.
• Hyperparameter
  • Usually set explicity by the designer. Not 'learned' during training.
  • These are set before training
  • Eg: Batch size in training dataset, learning rate
  • Difference from 'parameter': Parameters are learnt during training process
    • The term parameter refers to model parameters.
    • Eg: Weights in a neural network, k-value in K-nearest-neighbour
    • https://machinelearningmastery.com/difference-between-a-parameter-and-a-hyperparameter/
• Choosing value of hyperparameters
  • Evaluating AI models
  • Grid search: Try all possible configurations => often impractical
  • Random search: Select random combinations as configurations
    • Results shown to be close enough to grid search
  • https://www.kaggle.com/code/willkoehrsen/intro-to-model-tuning-grid-and-random-search

• Sobel operator: Used for edge detection in images
  • aka Sobel filter
• Transformers
  • Dethroned CNNs' place as the choice for computer vision
  • Can handle both text and images
• GPT: Generative Pre-trained Transformers
• Diffusion models
• GANs: Generative Adversarial Neural-networks
  • Made popular by deepfakes
  • Made less relevant by diffusion models ??
• RNNs: Recurrent Neural Networks
  • LSTM: Long Short-Term Memory
    • No vanishing gradient problem
  • GRU: Gated Recurrent Unit
    • A simplified form of LSTMs

Reinforcement learning

• Value iteration

• Q-Learning

• Policy (π)

  • Are probabilities of 'transitions'
  • P(s'|s, a): probability of going to s' from s by taking action a
  • Expected utility of a π
  • Probabilistic => need not take the same path every time for the same π
  • π^*: optimal policy
    • Policy with maximum expected utility

• Discounted rewards

  • Discount factor = γ

• R(s): reward associated with state s

• Markov process

  • Next state is independent of history

• MDP: Markov decision process

—

Bellmann equation for utility

U(s) = R(s) + γ max [P(s'|s, a) * U(s')]
                a∈A

This essentially says this:

Utility of a state = Reward of that state + Utility of next state

Sketching

From here:

A sketch C(X) of some data set X with respect to some function f is a compression of X that allows us to compute, or approximately compute, f (X) given access only to C(X).

• 'compress data in a way that lets you answer queries' ¹¹
• Helps save bandwidth when streaming data over a network.

Fun facts

• Fun fact (2025): PyTorch is the most popular framework used in deep learning research
• Hugging Face also provide libraries. Like trainers.
  • https://huggingface.co/docs/transformers/en/main_classes/trainer

Libraries and frameworks

• PyTorch
• Tensorflow (dead??)
• Theano
• Caffe
• Keras

Acks

Much of the info here is result of online searches instigated by conversions with the following people: Vishnu, Likith, Eva