RGPV Machine Learning Exam Q&A
Uploaded by
royvvk921RGPV Machine Learning Exam Q&A
Uploaded by
royvvk921Common questions
Powered by AIRNNs are preferable over traditional feedforward networks when dealing with sequential data, such as text, speech, or time series, where contextual understanding over time is crucial. RNNs' inherent ability to maintain memory and context allows them to model temporal dependencies effectively, an essential feature for tasks like machine translation and sentiment analysis .
LSTM networks use gates to control the flow of information, addressing RNN's challenge with vanishing gradients by providing separate pathways for short-term and long-term information retention. Their architecture allows selective memory updates, maintaining relevant information over long sequences and efficiently managing dependencies to solve complex tasks such as language modeling and time-series prediction .
Convolutional layers in CNNs apply filters (kernels) over input data to capture spatial hierarchies through localized feature extraction, such as edges and textures, enhancing the network's ability to generalize image data. These transformations reduce computational complexity by focusing on essential features, facilitating efficient processing and leading to better performance in image recognition tasks .
Dropout is an effective regularization technique because it prevents overfitting by randomly dropping units during training, forcing the network to learn redundant representations of data. This randomness encourages a form of ensemble learning within the network, improving generalization by ensuring neurons do not rely on each other too much and preventing co-adaptations that might limit model robustness .
Neural networks mimic the brain's interconnected neurons and learn by adjusting weights through layers of neurons, capable of learning complex patterns via activation functions that introduce non-linearity. In contrast, traditional models often rely on predefined features and simpler decision boundaries. This depth allows neural networks, particularly deep learning models, to model intricate data patterns .
Transfer learning optimizes training by reusing a pre-trained model on a new, related task, which saves time and resources. It is particularly effective with limited data, as it leverages previously learned features. However, its limitation lies in its dependence on the relevance between the original and new tasks; if they're not sufficiently similar, transferred features may not enhance, or could even degrade, model performance .
Pooling operations, like max pooling, reduce the spatial dimensions of feature maps, maintaining significant or most activated features while minimizing computational load and complexity. These operations effectively downsample data, helping to control overfitting by reducing the network's sensitivity to positional variations, thereby enhancing model robustness and accuracy in image processing tasks .
Data preprocessing enhances model performance by cleaning raw data, removing noise, handling missing values, and normalizing features. This results in more accurate and reliable models because the algorithm can better discern patterns without biases from inconsistent or irrelevant data .
Activation functions are crucial in neural networks because they introduce non-linearity, enabling the network to learn complex patterns and hierarchies within data. Without them, a neural network could only function as a linear model. Different types, such as ReLU, Sigmoid, and Tanh, offer various benefits; for example, ReLU introduces sparse representation, improving efficiency. Overall, activation functions greatly increase neural network complexity and expressiveness .
The vanishing gradient problem occurs when gradients become too small during backpropagation in deep networks, hindering the learning of earlier layers. This is particularly prevalent in standard RNNs when dealing with long sequences. LSTMs mitigate this by utilizing gate mechanisms that regulate the flow of information and gradients, allowing the retention of significant gradients over long periods and effectively learning long-term dependencies .