Artificial Intelligence Unfolded - Article 1: A Comprehensive Guide to ML, Neural Networks, and Deep Learning

In the ever-evolving landscape of technology, terms such as Artificial Intelligence (AI), Machine Learning (ML), Neural Networks (NN), Deep Learning (DL), and Generative AI are far more than mere buzzwords; they epitomise the cutting edge of innovative solutions across a multitude of sectors.

Currently, I'm attending an extensive course at the University of Oxford, which covers a broad range of topics. These include core Machine Learning and Deep Learning algorithms, the mathematics underpinning them, and power of Python. It starts getting more interesting though. The coverage extends to advanced topics such as LLMs, Generative AI, OpenAI, Prompt Engineering, Retrieval Augmented Generation (RAG), GraphRAG, Cloud platforms such as Azure OpenAI, AWS as well as exploring MLOps, Co-pilots, LlamaIndex, and notably, Autonomous AI Agents.

Course Director, Ajit Jaokar , a Visiting Fellow of the Department of Engineering Science at 英国牛津大学 and an influential industry figure himself, has lined up impressive leaders, Christoffer Noring, Jerry Liu, Alfredo Deza, Anthony Alcaraz, David Stevens, Wenqi Glantz, Dr Erika Tajra, Dr. Kakasaheb Nandiwale, Andy McMahon , Anjali Jain , Amita Kapoor to name a few, for discussion sessions on AI. The knowledge gained through these discussions with the group of notable and influential leaders in the field of artificial intelligence involves deep diving to understand the power of what we can achieve if AI is used ethically and carefully,?and it just blows the mind. Having said that though, it is also very easy to feel inundated with the plethora of jargons, technologies, models, tools, use cases, and the constant stream of evolving news encountered daily as everyone wants to onboard the journey of adopting AI.

So I thought, why not start consolidating the knowledge in a methodical order and put pen to paper to help me and possibly others who are embarking on the journey towards AI, implementing it carefully to make a difference in the business world. With this article, the first in a series to come, my goal is to lay down my thoughts and foundations of AI. This piece dives into the core concepts and applications of AI and ML, and distinguishes between these technologies, offering a comprehensive overview and insights into their key algorithms.

I'll post on other interesting topics in coming weeks, mainly those which have interested me the most and will try and keep it simple!

Artificial Intelligence: The Foundation of Future Technologies

AI represents the capability of software—or more broadly, systems—to carry out tasks that typically necessitate human intelligence. This encompasses a wide array of functionalities, from understanding natural language to recognising patterns in data.

AI applications are often driven by machine learning and deep learning algorithms, which allow these systems to learn from data and improve over time. These technologies are crucial in addressing stochastic problems, where outcomes are influenced by randomness, unlike deterministic problems that have predictable outcomes based on input parameters, conditions, and rules.

The Spectrum of AI: From Narrow AI to AGI

The realm of AI spans from Narrow AI, systems designed to perform specific tasks within a limited domain—like adjusting a thermostat based on environmental data or algorithmic trading in financial markets—to Artificial General Intelligence (AGI). AGI envisions highly autonomous systems capable of outperforming humans across most tasks of economic value, embodying or even exceeding human-level intelligence across a broad spectrum of intellectual activities. This includes:

• Natural Language Processing (NLP)
• Autonomous AI
• Visual Perception
• Intelligent Robotics
• Knowledge Representation
• Expert Systems
• Planning and Scheduling
• Speech Recognition
• Problem and Search Strategies

Machine Learning: The Engine of AI

Machine Learning, a critical subset of AI, empowers systems with the capability to autonomously learn from data and enhance their performance over time. This involves uncovering hidden patterns within data without being explicitly programmed to make specific predictions or decisions.

Supervised Learning

At the heart of Machine Learning lies Supervised Learning, a methodology where the model is trained on a labeled dataset, which means that each training example is paired with an output label. This approach facilitates models in predicting outcomes for unforeseen data, making it foundational for numerous applications such as spam detection, sentiment analysis, risk assessment, and price prediction.

Key Concepts in Supervised Learning

• Data Pre-processing: Data is cleaned to remove any errors or inconsistencies. This step may include handling missing values, removing noise etc.?
• Feature Engineering: The art of selecting, modifying, or creating new features (variables) to boost model efficacy.
• Training and Test Data: Data is bifurcated into a training set, on which the model learns, and a test set, used to evaluate its performance.
• Model Selection: Suitable algorithm is selected based on the problem and data characteristics.
• Training Model: This involves feeding the training data into the model so it can learn the relationships between features and target variables. The model adjusts its internal parameters to minimise errors during this process.
• Model Evaluation: Evaluate the model using test data to assess its performance. Common evaluation metrics include accuracy, precision, recall (also known as sensitivity or true positive rate).
• Hyperparameter Tuning: The optimisation of algorithm parameters to maximise model performance.
• Cross-Validation: Enhances model evaluation accuracy by training the model multiple times with different subsets of the data.
• Overfitting and Underfitting: Critical challenges that arise from models learning the noise and outliers in the training data too well (overfitting) or being too simplistic to capture the underlying data structure (underfitting).

Classification

Classification, a pivotal type of supervised learning, aims to categorise data points into predefined classes. This method is foundational to applications such as email spam detection, customer segmentation, and sentiment analysis of user reviews.

Classification Algorithms

• Logistic Regression: Utilised for binary classification tasks (e.g., determining if an email is spam or not).
• Decision Trees: Offer a straightforward, interpretable decision-making process but can overfit if not properly tuned.
• Random Forest: An ensemble approach that averages the predictions of multiple decision trees to mitigate overfitting.
• Support Vector Machines (SVM): Excel at finding the optimal hyperplane to separate different classes in the feature space, often outperforming other classifiers in high-dimensional spaces.

The objective of the algorithm is to find the finest line or decision boundary that can separate n-dimensional space into classes such that one can put the new data points in the right class in the future. This decision boundary is called a hyperplane. In most of the cases, SVMs have a cut above precision than Decision Trees, KNNs, Naive Bayes Classifiers, logistic regressions, etc. In addition to this SVMs have been well known to outmatch neural networks on a few occasions. SVMs are highly recommended due to their easier implementation, and higher accuracy with less computation.

• Margin – Margin is the gap between the hyperplane and the support vectors.
• Hyperplane – Hyperplanes are decision boundaries that aid in classifying the data points.
• Support Vectors – Support Vectors are the data points that are on or nearest to the hyperplane and influence the position of the hyperplane.
• Kernel function – These are the functions used to determine the shape of the hyperplane and decision boundary.

• Naive Bayes: Applies Bayes' theorem with “naive” assumption of independence among features, suitable for text classification and high-dimensional datasets. Bayes theorem in simple English describes the probability of an event, based on prior knowledge of conditions that might be related to the event ** https://en.wikipedia.org/wiki/Bayes%27_theorem
• K-Nearest Neighbors (K-NN): A simple yet effective algorithm that classifies each data point based on the majority class of its nearest neighbors.
• Neural Networks: Deep learning models, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have revolutionised complex classification tasks such as image and speech recognition.

The evolution from traditional algorithms like Logistic Regression and Decision Trees to more complex models such as Support Vector Machines and Neural Networks illustrates the rapid advancement and diversification of machine learning techniques. Each algorithm has its unique strengths and is suited to specific types of problems, highlighting the importance of understanding the underlying principles to effectively apply them to real-world challenges.

As we delve deeper into the realms of AI and ML, it becomes apparent that the field is not just about selecting the right algorithm but also about understanding data, crafting features, and fine-tuning parameters to coax the best performance from models. The journey through Supervised Learning and Classification offers a glimpse into the meticulous and nuanced process of developing intelligent systems capable of making sense of vast and complex datasets.

In the table below I list various Classification algorithms and their use cases.

Algorithm

Logistic Regression

Use Cases

- Binary classification

- Probabilistic outcomes

When to Use

- When the outcome is binary or dichotomous (e.g., spam or not spam)

- Useful for understanding the impact of independent variables on the outcome due to its interpretability

Algorithm

Decision Trees

Use Cases

- Classification and regression

- Feature importance

When to Use

- When data has a hierarchical structure

- Easy to interpret and explain to non-technical stakeholders

- Handles both numerical and categorical data

Algorithm

Random Forest

Use Cases

- Multiclass classification

- Feature importance

When to Use

- When dealing with overfitting in decision trees

- For improving accuracy through ensemble learning

- Handles large datasets with higher dimensionality well

Algorithm

Support Vector Machines (SVM)

Use Cases

- Binary classification

- Multiclass classification

When to Use

- When there is a clear margin of separation in high-dimensional space

- Effective in cases where the number of dimensions exceeds the number of samples

Algorithm

Naive Bayes

Use Cases

- Text classification

- Spam filtering

When to Use

- When assumptions of feature independence hold

- Efficient with large datasets

- Good baseline for text-related tasks

Algorithm

K-Nearest Neighbors (K-NN)

Use Cases

- Classification

- Regression

- Recommender systems

When to Use

- When data is labeled and the dataset is not too large (to avoid performance issues)

- Useful in applications like recommendation systems where similarity to neighbours is a strong indicator

Regression

Moving beyond classification, Regression represents another cornerstone of Supervised Learning. Unlike classification, which deals with discrete outcomes, regression focuses on predicting continuous variables. It's instrumental in establishing a relationship between a dependent (target) variable and one or more independent (predictor) variables. Through regression models, we can fit a line, curve, or surface that best represents the data, providing a quantitative assessment of relationships among variables.

Key Concepts in Regression Analysis

• Coefficient of Determination (R2): This statistic measures how well the regression predictions approximate the real data points. An R2 of 1 indicates that the regression predictions perfectly fit the data.
• P-value: Offers an assessment of the statistical significance of each feature in the model, helping determine which variables have a meaningful contribution to the prediction.
• Residuals: The differences between the observed values and the model's predicted values, serving as a diagnostic tool to evaluate the model's accuracy and assumptions.

Regression models excel in forecasting and predicting outcomes, making them indispensable in fields such as economics, finance, and the biological sciences. They enable us to understand and quantify the relationship between variables, paving the way for informed decision-making and predictive analytics.

Regression Algorithms

• Linear Regression: The most fundamental form of regression, linear regression, uses a linear approach to model the relationship between the dependent and independent variables. It is represented by the equation Y = β? + β?X + ε, where Y is the dependent variable, X is the independent variable, β? is the intercept, β? is the slope, and ε is the error term.

Other notable regression algorithms include:

• Decision Trees: While commonly associated with classification, decision trees can also be adapted for regression, predicting continuous outcomes based on decision rules inferred from the data.
• Random Forests: An ensemble method that uses multiple decision trees to improve prediction accuracy and control over-fitting, suitable for both classification and regression tasks.
• Support Vector Regression (SVR): Adapts the principles of Support Vector Machines (SVM) for regression, focusing on fitting the error within a certain threshold to predict continuous outcomes.

The exploration of regression, alongside classification, underscores the versatility and depth of Supervised Learning. By understanding both discrete and continuous prediction models, practitioners can apply these techniques across a broad spectrum of real-world problems, from predicting stock prices to estimating medical outcomes.

Unsupervised Learning

Diverging from the supervised learning models discussed previously, Unsupervised Learning involves analysing and clustering unlabeled datasets. This approach allows us to discover hidden patterns or data groupings without the need for prior training on labeled data. Clustering, a key unsupervised learning technique, exemplifies this by grouping data points into clusters based on similarity measures.

Clustering

Clustering algorithms aim to segregate sets of objects into groups, such that objects within the same cluster exhibit higher similarity to each other than those in different clusters. This technique is invaluable for exploratory data analysis, revealing natural groupings, anomaly detection, and customer segmentation among others. It provides insights into data structure without predefined labels, driven by the intrinsic characteristics of the data itself.

Types of Clustering

• Centroid-based Clustering: Characterised by the representation of each cluster by a single mean vector, with K-means clustering being the typical example.
• Hierarchy-based Clustering: Forms a tree of clusters, which can be built bottom-up or top-down.
• Density-based Clustering: Defines clusters based on areas of high density, distinguishing between core points, border points, and outliers. DBSCAN and OPTICS are prominent examples.
• Distribution-based Clustering: Assumes data is generated from a mixture of distributions, such as Gaussian distributions, with Gaussian Mixture Models (GMM) being a notable instance.
• Grid-based Clustering: Quantises the space into a finite number of cells that form a grid, upon which clustering is executed. STING is an example of this approach.

Clustering Algorithms

• K-means Clustering: Segments the data into K clusters, minimizing the variance within each cluster. It iteratively refines the position of centroids until achieving convergence.
• Hierarchical Clustering: Constructs clusters by either merging smaller clusters into larger ones or dividing larger clusters into smaller ones, visualised through a dendrogram (a tree diagram).
• DBSCAN: Identifies clusters based on the density of data points, capable of discovering clusters with arbitrary shapes and distinguishing outliers.
• Gaussian Mixture Models (GMM): Models the data as originating from multiple Gaussian distributions, utilising the expectation-maximisation algorithm to estimate parameters.

The exploration of clustering within Unsupervised Learning underscores the methodology's ability to provide deep insights into the structure of datasets without relying on pre-labeled outcomes. This aspect of machine learning opens up possibilities for discovering new patterns and relationships in data, showcasing the versatility and depth of machine learning techniques.

Evaluation of Clustering

The evaluation of clustering algorithms presents unique challenges, distinct from those encountered in supervised learning. Without ground truth labels for comparison, traditional metrics like accuracy or precision are not applicable. Nevertheless, various metrics have been developed to assess the quality of clustering, providing insights into how well an algorithm has performed in grouping similar items together.

Silhouette Score

One of the most insightful metrics for evaluating clustering performance is the Silhouette Score. This measure calculates how similar an object is to its own cluster compared to other clusters. The score ranges from -1 to 1, where:

• A high score (close to 1) indicates that the object is well integrated into its own cluster and distinctly separated from other clusters.
• A score around 0 suggests that the object is on or very close to the boundary between two clusters.
• A low score (close to -1) signifies that the object is poorly matched to its own cluster and perhaps belongs in a different cluster.

The Silhouette Score provides a concise, yet powerful indication of the effectiveness of the clustering. High average scores across all data points suggest that the clustering configuration is appropriate and distinct, while low scores may indicate overlapping clusters or inappropriate cluster definitions.

The evaluation of clustering, through metrics such as the Silhouette Score, is crucial for validating the results of unsupervised learning algorithms. It guides data scientists in refining their models and selecting the most appropriate clustering techniques for their specific data and objectives.

In the table below I list various Clustering algorithms and their use cases.

Algorithm

K-means Clustering

Use Cases

- Market segmentation

- Document clustering

- Image segmentation

When to Use

- For large datasets due to its efficiency

- When instances can be clearly separated into non-overlapping clusters

Algorithm

Hierarchical Clustering

Use Cases

- Taxonomy generation

- Organisational chart creation

When to Use

- When the number of clusters is not known in advance

- For smaller datasets due to higher computational cost

- When a hierarchy of clusters is more informative than flat clusters

Algorithm

DBSCAN (Density-Based Spatial Clustering of Applications with Noise)

Use Cases

- Anomaly detection

- Spatial data clustering

- Identifying clusters of arbitrary shapes

When to Use

- When there is noise in the data and outliers are present

- For datasets where cluster density varies

- When the number of clusters is unknown and clusters have arbitrary shapes

Algorithm

Gaussian Mixture Models (GMM)

Use Cases

- Image segmentation

- Speech recognition

- Customer profiling

When to Use

- When clusters are assumed to have different sizes and covariance structures

- For datasets where clusters can overlap

- When a probabilistic cluster assignment is preferred

Neural Networks (NN)

Neural Networks stand at the core of deep learning, drawing inspiration from the human brain to mimic how biological neurons communicate. These networks comprise input and output layers, along with one or more hidden layers, where the actual processing occurs. The interconnected nodes, or neurons, within these layers apply activation functions to process inputs and generate outputs, enabling the network to learn from data patterns.

Types of Neural Networks

• Feedforward Neural Networks: The simplest form, where information moves in one direction from input to output layers without any loops.
• Recurrent Neural Networks (RNNs): These networks include loops, allowing them to retain information previously learned, which is crucial for tasks requiring context from earlier data points, like language modeling or time series prediction.
• Convolutional Neural Networks (CNNs): CNNs are optimised for processing data in multiple arrays (e.g., images), making them exceptionally good at recognising visual patterns.

Components of Neural Networks

• Artificial Neurons: The fundamental unit of computation, each neuron processes inputs using weights (a measure of its importance) and an activation function to produce an output.?

• Layers: Neural networks are structured into layers of these neurons. The simplest neural network consists of an input layer, to receive the input, and an output layer, to produce the final output. Most neural networks also have one or more hidden layers between the input and output layers, which allow them to learn complex patterns.
• Weights and Biases: Each connection between neurons has an associated weight, which is adjusted during the learning process. Each neuron can also have a bias, allowing it to shift the activation function to the left or right, which is also adjusted during learning.
• Activation Function: This function is applied to the weighted sum of the inputs to a neuron, producing the neuron's output. The purpose of the activation function is to introduce non-linearity into the network, allowing it to learn complex patterns. Common activation functions include the sigmoid, tanh, and ReLU (Rectified Linear Unit) functions.

The Learning Process in Neural Networks

The learning process involves several key steps:

The weights for each input to an artificial neuron in a neural network are determined through a learning process. Initially, weights are usually set to small random values. Then, as the network is trained on a dataset, these weights are incrementally adjusted based on the error of the network's output compared to the expected result. The goal of this training process is to minimise the error across all outputs, thereby allowing the network to learn the underlying patterns in the data. The process of adjusting the weights is achieved through several key steps:

1. Initialisation

• Initially, weights are often set close to zero but not exactly zero to break symmetry. Random initialisation helps to ensure that each neuron learns something different, making the training process more efficient.
• Techniques like Xavier/Glorot or He initialisation are commonly used to set the initial weights in a way that seeks to ensure the gradients are neither too small nor too large, which can help prevent the vanishing or exploding gradients problem, respectively.

2. Forward Propagation

• Data is fed into the network, passing through each layer from the input to the output.

• At each neuron, an input is received, a weighted sum is computed (including a bias), and an activation function is applied to this sum to produce the neuron's output.
• The process continues until the final output is produced.

3. Loss Calculation

• Once the network produces an output, the difference between this output and the true value (the label or target) is calculated using a loss function. Common loss functions include Mean Squared Error (MSE) for regression tasks and Cross-Entropy Loss for classification tasks.

4. Backpropagation

• The gradient of the loss function with respect to each weight in the network is calculated. This process is called backpropagation because it starts at the output layer and works backward through the network, using the chain rule of calculus to compute gradients.
• This gradient tells us how to change each weight to minimise the loss; specifically, it indicates the direction to adjust the weights to reduce the error.

5. Weight Update

• Once the gradients are computed, the weights are updated using an optimisation algorithm. The most basic form of this is Gradient Descent, though in practice, more sophisticated optimisers like Stochastic Gradient Descent (SGD), Adam, or RMSprop are used.
• The optimiser adjusts each weight by a small step in the direction that will reduce the loss, based on the gradient. This step size is controlled by a parameter called the learning rate.

6. Iterative Optimisation

• The process of forward propagation, loss calculation, backpropagation, and weight update is repeated for many iterations over the dataset, with the network's weights being incrementally adjusted each time.
• The network is usually trained in batches (mini-batch gradient descent), meaning that the weight update step is performed after computing the loss on a small subset of the training data, rather than the entire dataset at once. This approach helps to speed up the training process and can improve the convergence of the optimisation algorithm.

Through this iterative process of adjusting weights based on the back propagated error gradients, the network learns to map inputs to the correct outputs, effectively "deciding" the importance (weight) of each input in making predictions.

Through the intricacies of neural networks, from their structure inspired by the human brain to the sophisticated learning mechanisms they employ, we delve deeper into the essence of what makes deep learning so powerful. Neural Networks, with their diverse architectures and components, underscore the capacity of machines to not just process data, but to learn and interpret the world in ways that mimic human cognition.

Deep Learning

Deep Learning, a subfield of machine learning, leverages neural networks with multiple layers—hence "deep"—to learn from vast amounts of unstructured or unlabeled data. This approach is inspired by the structure and function of the human brain, specifically the interconnectedness and layered nature of neurons.

Deep learning models, through their depth, are adept at learning hierarchies of information, enabling them to tackle complex tasks that are beyond the reach of more traditional algorithms.

Deep Learning Models

Multi-Layer Perceptrons (MLPs)

Multi-Layer Perceptrons (MLPs) are a class of feedforward artificial neural networks, which consist of at least three layers of nodes: an input layer, one or more hidden layers, and an output layer. Each node, or neuron, in one layer connects with a certain weight to every node in the following layer, making the network fully connected. MLPs are a foundational model in deep learning, used for solving both regression and classification problems by learning complex patterns in data.

Training MLPs

Training MLP involves adjusting its weights and biases to minimise the loss function. This is typically done using gradient descent or variations thereof (e.g., stochastic gradient descent). The training process involves repeatedly:

• Performing forward propagation to compute the output for a batch of inputs.
• Computing the loss by comparing the predicted output to the actual target values.
• Performing backpropagation to compute the gradient of the loss function with respect to each weight and bias.
• Adjusting the weights and biases in the direction that minimally reduces the loss, using the computed gradients.

Convolutional Neural Networks (CNNs)

CNN is a type of forwarded neural network that excel in processing visual data, drawing inspiration from the biological visual cortex. They are particularly effective in image and video recognition, object detection, and other tasks requiring the analysis of visual content.

Key Components of CNNs

• Convolutional Layers: The core building blocks of a CNN. These layers perform a convolution operation that filters the input image to extract features such as edges, textures, or more complex patterns in higher layers. Each convolutional layer applies numerous filters to the input and generates a feature map for each filter. The filters are learned during the training process.
• ReLU (Rectified Linear Unit) Layer: Follows the convolutional layer; introduces non-linearity into the model, allowing it to learn more complex patterns. The ReLU function is applied to each pixel of the feature map, replacing negative pixel values with zero, which speeds up training without affecting the network's ability to converge.
• Pooling (Subsampling or Down-sampling) Layer: Reduces the dimensionality of each feature map but retains the most important information. Pooling can be of different types, such as max pooling (which takes the maximum value in each window of pixels) or average pooling (which takes the average value). This step reduces the computational complexity and the number of parameters.
• Fully Connected (Dense) Layer: After several convolutional and pooling layers, the high-level reasoning in the neural network is done through fully connected layers. Neurons in a fully connected layer have full connections to all activations in the previous layer, as seen in regular neural networks. The output from the convolutional/pooling layers is flattened (converted into a 1D vector) before being fed into the fully connected layer.
• Output Layer: The final layer, typically a softmax activation function is applied for classification tasks, which gives the probability of the input image belonging to a particular class.

How CNNs Work

• Input: The input to a CNN is typically an image, or a batch of images, represented as multi-dimensional array of pixels.
• Feature Learning: Through convolutional and pooling layers, the network learns to identify various features in the images. Early layers may identify simple features like edges and curves, while deeper layers can identify more complex features like parts of objects or even entire objects.
• Classification: After feature extraction, the network uses fully connected layers to classify the images based on the features extracted by the convolutional and pooling layers.

Training CNNs

CNNs are trained using a large set of labeled images. The training process involves:

• Feeding the network a batch of images.
• Performing forward propagation to compute the loss (difference between the predicted outputs and the actual labels).
• Using backpropagation to compute the gradients of the loss with respect to the weights.
• Updating the weights using optimisation algorithms like stochastic gradient descent, etc.

Recurrent Neural Networks (RNNs)

Designed for sequential data, RNNs can remember information from earlier inputs using loops within the network, making them ideal for time series analysis, natural language processing, and other domains where context matters

Long Short-Term Memory (LSTM)

LSTMs are an advanced form of RNNs capable of learning long-term dependencies, addressing the challenge of remembering information over extended sequences. This makes them particularly powerful for tasks in natural language processing and complex time series analysis.

How Deep Learning Works

Deep Learning models undergo a training process where they learn to identify patterns and features in data.

Feature Learning: Through layers of processing, the model learns to identify important features, starting from simple ones in early layers to more complex features in deeper layers.

Classification and Prediction: Utilising the features learned, the model makes predictions or classifies data, often through fully connected layers at the end of the network.

These models are trained on large datasets, using backpropagation and optimisation algorithms to adjust weights and minimise loss, allowing them to improve over time and handle tasks of increasing complexity.

And finally, in the below table I have listed the Deep Learning models and their use cases.

Algorithm

Multi-Layer Perceptron (MLP)

Use Cases

- Classification tasks

- Regression tasks

- For datasets with a high number of features

When to Use

- When the relationship between inputs and outputs is complex but does not involve temporal or sequential data

Algorithm

Convolutional Neural Networks (CNN)

Use Cases

- Image recognition

- Video analysis

- Image classification

- For tasks involving image or video data where spatial hierarchies in the data are relevant

When to Use

- When performance and accuracy in visual tasks are critical

Algorithm

Recurrent Neural Networks (RNN)

Use Cases

- Language modeling

- Speech recognition

- For sequential data such as text or time series

When to Use

- When context or the sequence of data points is important for prediction

Algorithm

Long Short-Term Memory (LSTM)

Use Cases

- Sequence prediction

- Time series forecasting

- Natural language processing

- For tasks that require learning long-term dependencies in sequential data

When to Use

- When RNN performance is limited by vanishing or exploding gradient problems