A data processing bottleneck occurs when the time taken to load, transform, and prepare complex data—such as high-resolution images or large text corpora—exceeds the time required for actual model training. To prevent data preprocessing from severely slowing down the overall machine learning pipeline, practitioners must implement efficient algorithms and utilize advanced tools.

Data Processing Bottleneck

Beyond standard categorical and numeric variables, deep learning practitioners frequently work with myriad unstructured data formats. These complex data types include text strings for natural language processing, images and point clouds for computer vision, and audio data. Handling these diverse formats requires specialized approaches and advanced tools compared to traditional tabular data.

Claude

University of Michigan - Ann Arbor

https://www.coursera.org/learn/neural-networks-deep-learning?specialization=deep-learning

Neural Networks and Deep Learning

A dataset is collections of data, which can be in the form of a table, network, text, image, video, sound, ...

Dataset

A common imputation heuristic for missing categorical fields is to treat the absence of data itself as a distinct category. For example, if a column contains both valid string categories and NaN entries, it can be transformed into multiple binary dummy variables—where one specific column serves to indicate whether the original value was missing by using $$1$$ for missing and $$0$$ otherwise.

Categorical Imputation via Dummy Variables

When dealing with missing numerical data, a common imputation heuristic is to replace the missing entries with the mean value of the corresponding column. This approach assumes that the average of the observed values provides a reasonable estimate for the unobserved ones.

Numerical Imputation via Mean

Dive into Deep Learning

(1) Standard neural networks: real estate, online advertising

(2) Convolutional neural networks: applied in image recognition including photo tagging

(3) Recurrent Neural Networks: applied in sequence data including speech recognition, machine translation

(4) custom/hybrid neural networks: complex applications such as autonomous driving

application of neural networks in supervised learning

http://datahacker.rs/neural-networks-representation/

deep-learning representation

Because we have access to large columns of data, more computational power, and advanced algorithms.

For big data, large neural networks perform better than traditional ML models (SVM, logistic regression...)

Why is deep learning taking off?

(1) initialize parameters
(2) Implement the forward propagation to get the prediction
(3) compute the loss
(4) Implement the backward propagation to compute the derivatives of the parameters
(5) update the parameters
(6) repeat steps (2) to (5) to reduce the cost.

The learning circle of the neural network

- Start off trying to have one neural network do multiple things (can actually lead to better performance)
- differences and similarities amongst tasks is utilized to output a better result

Multi-task Learning

Benchmarking is a way to analyze how well your neural network is performing on a given standard set.
-Ensembling
-Multicrop at test time
-using open source code


Benchmarking

• Single neuron == linear regression without applying activation(perceptron)

• Basically a single neuron will calculate weighted sum of input(W.T*X) and then we can set a threshold to predict output in a perceptron. If weighted sum of input cross the threshold, perceptron fires and if not then perceptron doesn’t predict.

What is a Neural Network(NN)?

The working memory is the system that enables the ability to explicitly hold and manipulate pieces of information that are used to achieve a goal. Unlike humans, neural networks do not have a working memory.

Working Memory

Deep Learning Data Types

- Number of data points (instances)
- Attributes (variables)
- Types of attributes (variables)
- Number of attributes (variables)
- Statistical measures

Characteristics of a dataset

Here are a list of free resources to find datasets. Most will be available in CSV format for easy integration into R or Python. 

Sample Datasets

It important to examine your dataset before starting to work with it for a machine learning task to:
- Understand how much missing data there is in the dataset.
- Get an idea of whether the features need further cleaning.
- Understand what techniques to use for analyzing the data.

The first step to analyze a dataset:

- 1) Equilibrium
- 2) Cyclic (periodic)
- 3) Random 
- 4) Complex

Wolfram's four classes of empirical data

Data can be distributed in several ways. Some of these include:
- Normal (Gaussian) distributions
- Uniform distributions
- Exponential distributions

Data Distributions

Supervised learning is a type of machine learning problem where the objective is to predict a designated unknown label based on known inputs. This approach requires a training dataset consisting of examples for which the ground truth labels are already known, allowing the model to learn the mapping from inputs to classifications.

Supervised Learning

A machine learning model is the computational machinery designed to ingest data of a specific type, apply mathematical transformations, and output predictions. In modern applications, the focus is on statistical models whose internal transformations can be estimated and refined using a dataset.

Machine Learning Model

Before data can be fed into any machine learning model, its overall quality must be evaluated and improved. Real-world datasets are often plagued by flaws such as extreme outliers, faulty sensor measurements, and human recording errors. Addressing these discrepancies is an essential part of the data preparation process to ensure the model learns from reliable information.

Data Quality

In real-world machine learning scenarios, datasets are often not provided in a single, simple CSV file. Instead, they may be stored in a relational database, where information is distributed across multiple tables. For example, an e-commerce application might store customer addresses in one table and their purchase data in another, requiring practitioners to extract and merge these separate files before modeling.

Relational Database

The success of a machine learning model heavily depends on having the right data, as encapsulated by the principle 'garbage in, garbage out.' If a dataset contains mistakes, features that do not predict the target, or reflects societal prejudices by under-representing certain groups, the learning process will fail or produce biased, harmful results.

Machine Learning Dataset Quality

A machine learning example, interchangeably referred to as a data point, sample, or instance, represents an individual record within a dataset. To be useful for learning algorithms, each example is characterized by a specific numerical representation, forming the basis for a model's predictions.

Machine Learning Example

Comma-separated values (CSV) files are a ubiquitous format for storing tabular, spreadsheet-like data. In a CSV file, each line represents a distinct data record, and the individual fields within that record are separated by commas.

CSV File

A data batch, commonly referred to as a minibatch, is a discrete, smaller subset of a larger dataset that is processed simultaneously during a single iteration of machine learning model training or evaluation. Training typically requires multiple passes over a dataset, grabbing one minibatch at a time. Each minibatch consists of a tuple of features and corresponding labels. Processing data in minibatches rather than the entire dataset balances computational efficiency with hardware memory constraints.

Data Batch

When loading datasets for machine learning, the order in which data is read depends on the current mode. For training, data is typically read in a random order (shuffled) to properly update the model. For validation, data is read in a pre-defined, sequential order, which is important for debugging purposes and consistent evaluation.

Training vs. Validation Data Reading Order

The `pandas.get_dummies()` function is used to convert categorical variables into binary indicator columns. To apply categorical imputation, the `dummy_na=True` argument can be passed, which instructs the function to treat `NaN` as a valid category and generate a dedicated dummy column to explicitly flag missing entries.

pandas.get_dummies

For missing numerical values, a standard imputation heuristic is to substitute `NaN` entries with the mean of all available values in that column. In `pandas`, this is achieved by chaining the `.fillna()` and `.mean()` methods, such as `inputs.fillna(inputs.mean())`, which replaces all missing occurrences with the calculated column average.

pandas.DataFrame.fillna

The `mean()` method in `pandas.DataFrame` calculates the mathematical average of the values along a specified axis. When applied to a dataset containing missing values (`NaN`), it computes the mean of the available numerical entries in each column by default. This method is frequently used to programmatically determine the substitution values required for numerical imputation.

```python
# Calculate the mean of each column
column_means = inputs.mean()
```

Learn Before

Related

Learn After