Mobile Phone Policy in Exams
Uploaded by
Dakshita BhatnagarMobile Phone Policy in Exams
Uploaded by
Dakshita BhatnagarCommon questions
Powered by AIConvolutional layers are preferred in image processing because they efficiently use spatial hierarchy by applying filters that preserve spatial relationships between pixels. This local connectivity and weight sharing drastically reduce the number of parameters compared to fully-connected layers, making models less prone to overfitting . Fully-connected layers, using large numbers of parameters indiscriminately, do not efficiently capture the spatial structure of images, thus are less effective in handling high-dimensional image data.
Self-attention allows models to weigh the importance of different words in a sequence dynamically, capturing contextual relationships better than models relying on fixed feature sizes . Multi-headed attention extends this by allowing the model to focus on different positions simultaneously, enhancing the model's ability to learn multiple aspects of meaning in parallel . This parallel processing and dynamic context understanding enable transformers to outperform traditional sequence models like RNNs and LSTMs, especially in tasks requiring long-range dependencies and complex features.
Regularization, including L1 and L2 techniques, generally reduces a model's variance by preventing it from capturing noise, thus improving generalization at the cost of a slight increase in bias . Dropout, another regularization method, decreases variance by randomly ignoring neurons during training, which forces the model to learn robust features, thus also potentially increasing bias slightly . Both methods aim to balance the trade-off between bias and variance, ultimately leading to improved generalization abilities.
Training a model from scratch on a small dataset like D2 may lead to overfitting due to statistically insignificant data and a lack of generalization. Models trained this way might only capture noise rather than relevant patterns . Transfer learning mitigates this issue by leveraging pre-trained models on large datasets like D1, which can significantly reduce training time and improve performance. It allows models to learn refined feature representations, which are then fine-tuned with smaller datasets, leading to better generalization . Thus, transfer learning addresses small dataset limitations by utilizing established knowledge from larger datasets.
Stochastic Gradient Descent (SGD) enhances model training by performing updates iteratively and randomly on subsets of the data (mini-batches), which leads to more frequent updates compared to using the entire dataset in Batch Gradient Descent. This randomness helps in escaping local minima by introducing fluctuations that guide the model towards a global minimum . The algorithm is particularly effective in large-scale learning, where processing full datasets is computationally expensive, thereby facilitating efficient error minimization through iterative learning rates and convergence to optimal solutions.
When applying PCA, it's crucial to consider the trade-off between dimensionality reduction and information loss . PCA is useful for identifying the directions (principal components) that maximize variance in the dataset, effectively capturing the most informative aspects with fewer dimensions . However, reducing dimensions too aggressively can lead to loss of significant information, impacting the model's accuracy. Therefore, selecting the number of principal components should balance dimension reduction with preserving important data features.
Increasing the number of hidden units in an artificial neural network typically enhances the model's capacity to learn complex features and relationships, thus decreasing bias by allowing for more detailed representations of the input data . However, it can also increase variance, as more parameters can lead to overfitting, especially if the training data is limited. Therefore, while additional hidden units can improve learning capacity, careful application of regularization and cross-validation is critical to ensure that improved complexity doesn't compromise generalization.
Datasets with label inaccuracies often hinder learning as models may learn incorrect predictions, leading to imprecise classification and lower generalization abilities . Such challenges can be addressed by employing robust validation techniques and regularization methods like L1 or L2 to mitigate the effects of noise on the model's performance . Additionally, manual correction of labels or usage of cleaner datasets through data preprocessing can significantly enhance model accuracy and reliability.
Convolutional Neural Networks (CNNs) are primarily used for image recognition and computer vision tasks due to their ability to capture spatial hierarchies in images. They apply convolutional filters to input data, detecting patterns such as edges or textures, which are crucial for image processing . In contrast, Recurrent Neural Networks (RNNs), especially Long Short-Term Memory networks (LSTMs), are tailored for temporal data due to their feedback loops, which allow information to persist across time steps. This makes them suitable for tasks like sequence prediction and natural language processing . Thus, while CNNs excel in capturing spatial features, RNNs are adept at handling sequential data.
GoogLeNet introduces the inception module, which processes information through parallel convolutional operations of different sizes, allowing the network to capture various feature patterns at multiple scales . In contrast, Residual Networks (ResNets) introduce shortcut connections that bypass one or more layers, addressing the vanishing gradient problem and allowing for deeper networks . These shortcuts help in training much deeper models without degradation, facilitating learning across more layers while overcoming gradient dispersion issues.