w3 Flashcards
What makes deep learning ‘deep’?
o a) The complexity of features learned
o b) The sophistication of the neural connections
o c) The number of layers in the network
o d) The size of the input data
c) The number of layers in the network
Which of the following is true about convolutional neural networks?
o a) They classify objects by matching pixel colors
o b) They apply filters to detect features in an image
o c) They use a single layer to capture complex features
o d) They don’t use activation maps
b) They apply filters to detect features in an image
What is the main function of the classification module in a ConvNet?
o a) To calculate the number of layers
o b) To output a set of activation maps
o c) To determine the class with the highest confidence score
o d) To increase network depth
c) To determine the class with the highest confidence score
What is an epoch in the context of training a neural network?
o a) A function for determining classification
o b) A measure of data accuracy
o c) One full pass through the entire training dataset
o d) The confidence score of the model
c) One full pass through the entire training dataset
What drives a ConvNet to converge?
o a) Random initialization of weights
o b) Lack of training data
o c) Weight adjustments stopping as the network stabilizes
o d) Reduction in the number of layers
c) Weight adjustments stopping as the network stabilizes
True or False
The activation maps in ConvNets are similar to activation patterns in the human brain’s visual system.
True
True or False
The number of layers in a neural network directly reflects the quality of learning, regardless of the network’s architecture.
False
True or False
In a ConvNet, the highest convolutional layer is used as the final output layer for classification.
False
True or False
Edge detection typically occurs in the first layer of a ConvNet.
True
Why is object recognition challenging for neural networks, despite it seeming easy for humans?
Object recognition is complex for neural networks because they must account for variations in shape, size, orientation, and lighting, whereas the human brain processes visual information with remarkable flexibility and adaptability.
Explain the role of convolutional layers in a ConvNet and how they process image data.
Convolutional layers apply filters to the input image, creating activation maps that highlight specific features such as edges, textures, or object parts.
Convolutional layers extract features from image data by applying filters (kernels) that slide over the input, These filters detect patterns such as edges, textures, or shapes by focusing on local regions of the image (local receptive fields) creating activation maps. By stacking multiple convolutional layers, the network learns hierarchical features, with early layers detecting simple patterns and deeper layers capturing complex structures, enabling the ConvNet to understand and classify images effectively.
What does the term ‘convergence’ mean in neural network training, and why is it important?
Convergence refers to the stabilization of the network’s weights as training progresses.
How does backpropagation help a ConvNet learn, and what role do epochs play in this process?
Backpropagation enables the ConvNet to adjust its weights and biases by propagating the error backward and using gradients to minimize the loss function.
Epochs determine how many times the model iterates over the dataset, allowing it to progressively refine its understanding of the data.
Together, backpropagation and multiple epochs allow a ConvNet to effectively learn from data and improve its performance over time.
Backpropagation calculates errors from the output layer and adjusts the weights in each layer to minimize these errors.
What are some ways ConvNets are similar to the human brain’s visual processing system?
ConvNets process images in a hierarchical manner, where each layer builds upon the previous one to identify increasingly complex features.
What was the purpose of the Amazon Mechanical Turk in building the ImageNet dataset?
a) To design deep learning algorithms
b) To crowdsource image labeling for training data
c) To create the 1,000 categories in ImageNet
d) To analyze model performance
b) To crowdsource image labeling for training data
What is the ‘top-5 accuracy’ metric in the ImageNet competition?
a) The model’s top five categories must be correct
b) The correct label must appear among the model’s top five predictions
c) Five categories are selected for each image
d) A model with five layers achieves higher accuracy
b) The correct label must appear among the model’s top five predictions
Why did ConvNets start outperforming other methods in image recognition?
a) Increased access to large datasets and faster parallel computing hardware
b) ConvNets required less data than other models
c) Better labeling techniques reduced training time
d) ConvNets used top-1 accuracy to improve classification
a) Increased access to large datasets and faster parallel computing hardware
Which model’s success in the ImageNet competition is considered a turning point for ConvNets?
c) AlexNet
What is one of the main challenges ConvNets face in comparison to human object recognition?
a) ConvNets consistently outperform humans
b) ConvNets struggle with localization tasks
c) ConvNets are unable to recognize any real-world images
d) ConvNets cannot achieve top-5 accuracy
b) ConvNets struggle with localization tasks
True or False
ConvNets are now considered fully human-equivalent in object recognition across all tasks.
False
True or False
The ImageNet competition used Amazon Mechanical Turk to improve GPU processing speeds.
False
True or False
ImageNet has 1,000 categories, and models are evaluated using top-5 accuracy.
True
True or False
Data snooping, seen during some ImageNet competitions, refers to a type of cheating where test data is improperly accessed.
True
What were the two key factors that allowed ConvNets to dominate the field of computer vision? Explain how each contributed.
The availability of large datasets (e.g., ImageNet) provided ample training data, allowing models to learn complex features. Additionally, advancements in GPU technology enabled faster parallel computing.
Why is top-5 accuracy used as a benchmark in the ImageNet competition, and how does it differ from top-1 accuracy?
Top-5 accuracy is used because it allows the correct label to appear within the top five predictions.
Explain why object recognition is not considered ‘solved’ by AI, even though ConvNets perform well on tasks like ImageNet.
Object recognition is not considered “solved” by AI, despite ConvNets performing well on tasks like ImageNet, because they struggle with generalizing to unseen data, are vulnerable to adversarial attacks, and lack true understanding. While ConvNets can classify objects accurately, they do not have semantic grounding or contextual reasoning, meaning they may fail when faced with novel scenarios, variations, or real-world complexities. Additionally, real-world object recognition involves handling dynamic environments, contextual shifts, and integrating with broader cognitive tasks, areas where ConvNets alone fall short. This highlights the gap between benchmark performance and real-world robustness, making object recognition a still-unsolved challenge in AI.
How did the success of AlexNet in the ImageNet competition influence the development of deep learning and ConvNets?
AlexNet’s success showed that deep ConvNets could outperform traditional methods, leading to a surge in research.
What role did the ImageNet competition play in the evolution of computer vision benchmarks, and why is it considered a symbol of progress in AI?
The ImageNet competition provided a clear, quantitative measure of progress in computer vision.
Why might ConvNets focus on irrelevant parts of an image, like backgrounds, when classifying objects?
They learn patterns based solely on observed data.
What does the ‘long tail’ problem refer to in machine learning?
Poor performance on rare or unusual examples.
What is an adversarial example?
An input deliberately modified to mislead an AI model.
Which type of AI learning is most suitable for developing general-purpose AI?
Unsupervised learning.
true or false
ConvNets are naturally capable of learning on their own, without the need for labeled data.
False
true or false
Humans and ConvNets learn in fundamentally different ways, which affects AI’s trustworthiness and robustness.
True
true or false
The gorilla incident is an example of how biased data can lead to unfair or inaccurate AI predictions.
True
true or false
Explainable AI (XAI) is designed to make AI decisions easier to understand and interpret.
True
true or false
ConvNets perform well on non-linearly separable problems without the need for non-linear activation functions.
False
How does the way ConvNets learn differ from human learning, and why does this affect their robustness?
Unlike humans, who use context and reasoning, ConvNets learn only from labeled data.
Explain the ‘long tail’ problem in supervised learning and why it limits the development of general-purpose AI.
The long tail problem refers to a model’s difficulty with rare or unusual examples.
What are adversarial examples, and why do they pose a challenge for ConvNets?
Adversarial examples are inputs that are subtly modified to trick AI models into misclassifying them.
Discuss the importance of explainable AI (XAI) and how it might improve trust in AI systems.
Explainable AI aims to make AI decisions more transparent by allowing users to understand the reasons behind predictions.
Why might supervised learning result in biased AI systems, and what is a real-world example of this?
Supervised learning can lead to biased AI systems if the training data is unrepresentative of diverse groups.
activation map
An activation map is a visual representation of the features that a convolutional neural network (CNN) detects in an input image at a specific layer.
- It highlights the regions in the input that the network considers important for identifying patterns or making predictions. For example, in image classification, an activation map in an early layer might show edges or textures, while in deeper layers, it might highlight more complex structures like shapes or objects. Activation maps provide insight into what the network is “looking at” and help interpret how the model processes and understands the input data.
percepton learning rule
The perceptron learning rule is an algorithm used to adjust the weights of a perceptron during training. It works by comparing the predicted output to the actual target, then updating the weights to reduce the error. If the prediction is correct, no change is made; if incorrect, the weights are adjusted by adding or subtracting a small value proportional to the error and the input. This process is repeated for each training example, gradually improving the model’s ability to classify the data correctly.
Differences Between Perceptron and Deep Neural Networks (DNNs)
Architecture:
Perceptron: A single-layer neural network
DNNs: Multi-layer networks
Activation Function:
Perceptron: Uses a binary activation function
DNNs: Use continuous activation functions like sigmoid, tanh, or ReLU, which allow for more nuanced outputs and better handling of complex problems.
Learning Capabilities:
Perceptron: Can only solve linearly separable problems (e.g., AND, OR),
DNNs: Can solve both linearly separable and non-linearly separable problems (XOR)
Complexity:
Perceptron: Simple model with limited capacity, suitable for basic classification tasks.
DNNs: complex and powerful, capable of learning intricate patterns and representations, making them suitable for tasks like image recognition,
Training:
Perceptron: Uses the Perceptron learning rule for weight updates,
DNNs: Use backpropagation to adjust weights across multiple layers, (with gradient descent).
