1. Activation Function
- A function applied to the output of each neuron to introduce non-linearity, enabling the network to learn complex patterns. Examples include ReLU, Sigmoid, and Tanh.
2. AdaGrad
- An optimizer that adapts the learning rate for each parameter based on the historical gradient information. It’s useful for sparse data.
3. Adam
- A popular optimizer that combines the benefits of AdaGrad and RMSprop, using adaptive learning rates and momentum.
4. Autoencoder
- A type of neural network designed to learn efficient representations (encodings) of data by training to reconstruct the input from a compressed form.
5. Backpropagation
- The algorithm used to calculate gradients for updating weights during training by propagating errors backward through the network.
6. Batch Normalization
- A technique to normalize inputs within a network layer to stabilize and speed up training.
7. Bias
- An additional parameter in a neuron that allows the model to fit the data better by shifting the activation function.
8. Bidirectional RNN
- An RNN architecture where the input sequence is processed in both forward and backward directions to capture context from both past and future states.
9. BLEU Score
- A metric for evaluating the quality of text generated by models, such as in machine translation or image captioning, by comparing it to reference outputs.
10. Bounding Box
- A rectangular box used to define the location of an object in an image, commonly used in object detection tasks.
11. Convolutional Neural Network (CNN)
- A type of neural network designed for processing structured grid data, like images, using convolutional layers to extract features.
12. Cost Function
- Another term for loss function, it quantifies the difference between the predicted output and the actual output.
13. Cross-Entropy Loss
- A loss function commonly used for classification tasks, measuring the difference between the predicted probability distribution and the actual distribution.
14. Data Augmentation
- Techniques used to increase the size and diversity of the training dataset by applying random transformations like rotation, flipping, or cropping.
15. Deep Learning
- A subset of machine learning that uses neural networks with many layers (hence “deep”) to learn hierarchical representations of data.
16. Dense Layer
- A fully connected layer where each neuron is connected to every neuron in the previous layer, often used in feedforward networks.
17. Dropout
- A regularization technique where randomly selected neurons are ignored during training to prevent overfitting.
18. Epoch
- A full pass through the entire training dataset. Multiple epochs are often required to train a model.
19. Exploding Gradient
- A problem where gradients grow exponentially large during backpropagation, causing the model to become unstable.
20. Feature Map
- The output of a convolutional layer, representing the activation of filters applied to the input data.
21. Filter
- A small matrix applied to the input data in convolutional layers to detect specific patterns like edges or textures.
22. Fine-Tuning
- Adjusting a pre-trained model on a new, related task by training it further with a small learning rate.
23. Fully Connected Layer
- A layer where each neuron is connected to every neuron in the previous layer, typically found at the end of CNNs.
24. GAN (Generative Adversarial Network)
- A type of neural network where two models (a generator and a discriminator) are trained together to produce realistic data and distinguish it from real data.
25. Global Average Pooling
- A pooling technique that reduces the spatial dimensions of feature maps to a single value per feature map, typically used at the end of CNNs.
26. Gradient Descent
- An optimization algorithm that adjusts the model’s parameters by moving in the direction of the steepest decrease in the loss function.
27. Gradient Vanishing
- A problem where gradients become too small during backpropagation, making it difficult for the network to learn.
28. Graph Neural Network (GNN)
- A type of neural network designed to operate on graph-structured data, such as social networks or molecules.
29. Hyperparameters
- Settings that define the model’s architecture or training process, such as learning rate, batch size, or the number of layers.
30. ImageNet
- A large dataset used for training and evaluating image recognition models, consisting of millions of labeled images across thousands of categories.
31. Instance Normalization
- A normalization technique often used in style transfer tasks, normalizing feature maps for each individual input.
32. Keras
- A high-level neural networks API, written in Python, and capable of running on top of TensorFlow, CNTK, or Theano.
33. Learning Rate
- A hyperparameter that controls the step size during gradient descent. A lower learning rate means smaller steps, leading to slower convergence.
34. Learning Rate Decay
- A technique where the learning rate is gradually reduced during training to allow finer adjustments as the model converges.
35. Leaky ReLU
- A variation of the ReLU activation function where a small negative slope is introduced for negative inputs to avoid dead neurons.
36. LSTM (Long Short-Term Memory)
- A type of RNN architecture designed to better capture long-term dependencies by incorporating memory cells that can maintain information over time.
37. Margin
- In SVMs and related models, the margin is the distance between the decision boundary and the nearest data points of any class.
38. Max Pooling
- A pooling operation that reduces the size of the feature maps by taking the maximum value from a group of neighboring pixels.
39. Mean Squared Error (MSE)
- A loss function commonly used in regression tasks, measuring the average squared difference between predicted and actual values.
40. Momentum
- An optimization technique that accelerates gradient descent by adding a fraction of the previous update to the current one, helping to overcome small local minima.
41. Neural Architecture Search (NAS)
- The process of automatically finding the best architecture for a neural network, often using techniques like reinforcement learning or evolutionary algorithms.
42. Normalization
- The process of scaling input data or intermediate activations so they have a mean of zero and a standard deviation of one, improving training stability.
43. One-Hot Encoding
- A representation of categorical variables as binary vectors, where only one element is “hot” (set to 1), and all others are “cold” (set to 0).
44. Overfitting
- A scenario where a model learns the training data too well, including noise and outliers, resulting in poor generalization to new data.
45. Parameter Sharing
- A concept in CNNs where the same filter (weights) is applied across different parts of the input, reducing the number of parameters.
46. Perceptron
- The simplest type of artificial neuron, consisting of a linear function followed by a threshold activation function.
47. Pooling Layer
- A layer in CNNs used to reduce the spatial dimensions of feature maps, making the network more efficient and less sensitive to small translations in the input.
48. Precision
- A metric used to evaluate classification models, defined as the number of true positives divided by the sum of true positives and false positives.
49. Recurrent Neural Network (RNN)
- A type of neural network designed to handle sequential data, where connections between nodes form a directed cycle, allowing information to persist.
50. ReLU (Rectified Linear Unit)
- A popular activation function that outputs the input directly if it’s positive, otherwise outputs zero. It helps to mitigate the vanishing gradient problem.
51. Residual Network (ResNet)
- A deep neural network architecture that uses skip connections (or residual connections) to allow the model to learn residual functions, mitigating the vanishing gradient problem.
52. Ridge Regression
- A type of regression that includes a penalty for large coefficients, helping to prevent overfitting by shrinking the coefficients toward zero.
53. RMSprop
- An optimizer that uses a moving average of squared gradients to normalize the gradient, helping to deal with the vanishing and exploding gradient problems.
54. ROC Curve
- A graphical representation of the performance of a binary classifier, plotting the true positive rate against the false positive rate at various threshold settings.
55. Semantic Segmentation
- A computer vision task where each pixel in an image is classified into a category, such as labeling all pixels belonging to a person, car, or tree.
56. Sensitivity (Recall)
- A metric that measures the proportion of actual positives correctly identified by the model, calculated as true positives divided by the sum of true positives and false negatives.
57. Sigmoid Function
- An activation function that squashes input values between 0 and 1, often used in binary classification tasks.
58. Softmax Function
- An activation function used in multi-class classification tasks that converts logits (raw scores) into probabilities, where the sum of all probabilities equals one.
59. Sparse Coding
- A representation method where the input data is expressed as a sparse combination of basis vectors, often used in feature learning.
60. Spectral Normalization
- A technique used to stabilize GAN training by normalizing the spectral norm (maximum singular value) of the weight matrices.
61. Stride
- The step size by which the convolutional filter or pooling window moves across the input image. A larger stride results in a smaller output size.
62. SVM (Support Vector Machine)
- A supervised learning model that finds the optimal hyperplane that separates classes in a high-dimensional space with maximum margin.
63. Transfer Learning
- A method where a model pre-trained on one task is adapted for a new, related task, often improving performance when data is limited.
64. True Positive Rate (TPR)
- Also known as recall or sensitivity, it’s the proportion of actual positives correctly identified by the model.
65. Underfitting
- A situation where a model is too simple to capture the underlying patterns in the data, resulting in poor performance on both training and testing data.
66. Upsampling
- The process of increasing the spatial dimensions of feature maps, typically used in tasks like image generation or semantic segmentation.
67. Vanishing Gradient
- A problem in deep networks where gradients become very small during backpropagation, making it difficult for the network to learn.
68. Weight Initialization
- The process of setting the initial values of a network’s weights before training begins, crucial for ensuring proper convergence.
69. Weight Sharing
- A concept in CNNs where the same filter (set of weights) is applied across different parts of the input image, reducing the number of parameters.
70. Word Embedding
- A representation of words as dense vectors in a continuous space, capturing semantic relationships between words, often used in NLP tasks.
71. Zero Padding
- A technique where extra zeros are added around the input image before applying a convolution, preserving the spatial dimensions of the output.
72. Attention Mechanism
- A technique that allows the model to focus on specific parts of the input data, enhancing the ability to capture relevant features, widely used in NLP and computer vision.
73. Bag of Words (BoW)
- A simple representation of text data where each document is represented by a vector indicating the presence or frequency of words, ignoring grammar and word order.
74. Bayesian Neural Network
- A neural network that incorporates uncertainty in its predictions by using Bayesian inference, typically resulting in probabilistic outputs.
75. BERT (Bidirectional Encoder Representations from Transformers)
- A pre-trained NLP model that captures context from both directions (left-to-right and right-to-left) in text sequences, achieving state-of-the-art results on many tasks.
76. Capsule Network
- A type of neural network that uses capsules (groups of neurons) to capture spatial relationships and improve the ability to recognize objects in different poses.
77. Catastrophic Forgetting
- A problem in neural networks where learning new information causes the model to forget previously learned information, particularly in sequential learning tasks.
78. Class Imbalance
- A situation where some classes are significantly underrepresented in the training data, leading to biased models that perform poorly on minority classes.
79. Class Weighting
- A technique used to handle class imbalance by assigning higher weights to underrepresented classes in the loss function, encouraging the model to pay more attention to them.
80. Clipping
- A technique used to prevent exploding gradients by capping the gradient values to a maximum limit during backpropagation.
81. Collaborative Filtering
- A technique used in recommendation systems where the model predicts user preferences by analyzing patterns of likes and dislikes across many users.
82. Compositionality
- The principle that complex concepts can be constructed by combining simpler ones, often used in models that need to understand relationships in data.
83. Contrastive Loss
- A loss function used in tasks like face recognition, where the goal is to bring similar data points closer together in the embedding space and push dissimilar points apart.
84. Data Preprocessing
- The process of transforming raw data into a format suitable for training a model, including tasks like normalization, scaling, and augmentation.
85. DropConnect
- A regularization technique similar to dropout, where individual connections between neurons are randomly dropped instead of entire neurons.
86. Dynamic Routing
- A process used in capsule networks to iteratively update the weights of connections between capsules based on their agreement, improving the capture of spatial hierarchies.
87. Early Stopping
- A regularization technique where training is stopped when the performance on the validation set starts to deteriorate, preventing overfitting.
88. Elastic Net
- A regularization technique that combines the penalties of both L1 (Lasso) and L2 (Ridge) regression, encouraging sparsity and reducing overfitting.
89. Encoder-Decoder Architecture
- A neural network design used in tasks like machine translation and image captioning, where the encoder processes the input and the decoder generates the output sequence.
90. Entropy
- A measure of uncertainty or randomness in a dataset, often used in loss functions like cross-entropy to quantify the difference between distributions.
91. Feature Extraction
- The process of automatically identifying and extracting relevant features from raw data, often performed by the layers of a neural network.
92. Generative Model
- A type of model that learns to generate new data samples similar to the training data, as opposed to discriminative models that classify or predict labels.
93. Gradient Clipping
- A technique used to prevent exploding gradients by capping the gradient values to a maximum limit during backpropagation.
94. Hinge Loss
- A loss function used primarily in SVMs, where the loss increases linearly if the margin is not large enough to correctly classify the data point.
95. Knowledge Distillation
- A technique where a smaller, simpler model (student) is trained to replicate the behavior of a larger, more complex model (teacher), often used for model compression.
96. Latent Space
- A lower-dimensional representation of data where similar data points are close to each other, often used in generative models like autoencoders and GANs.
97. Local Response Normalization (LRN)
- A normalization technique that normalizes over local input regions, typically used in early layers of CNNs to aid generalization.
98. Meta-Learning
- A type of learning where the model learns to learn, often by training on a variety of tasks and generalizing to new tasks with minimal data.
99. Nesterov Momentum
- An optimization technique that extends momentum by adding a lookahead step, making the updates more responsive to the current gradient.
100. Objective Function
- Another term for loss function, it represents the function that the model aims to minimize during training.
These 100 terms should provide a strong foundation for understanding deep learning concepts and help you navigate the field more effectively.