Deep Learning and Neural Networks

 

Deep Learning and Neural Networks

 

“Deep learning is a subset of machine learning that utilizes multilayered neural networks, known as deep neural networks, to simulate the complex decision-making processes of the human brain. Unlike traditional machine learning models that use simple neural networks with one or two layers, deep learning models employ three or more layers, often reaching hundreds or thousands of layers”. They are composed of interconnected layers of “neurons” which perform mathematical operations. The strengths of the connections are adjusted using machine learning, providing more accurate outputs. In 2010 Yann Le Cun coined the term “self-supervised learning” to describe the way to train neural networks.

But how do neural networks work? They are inspired by the human brain, sending “signals” by multiplying the mathematical operations by weights, resulting in outputs. These weights can be adjusted through machine learning influencing the way the input is transformed into an output. If the result is not what is expected, the models learns by correcting the weight, changing the output. The neurons are connected to successive layers, the output of a layer becoming the input of another layer. The model learns in the hidden layers, which separates this process from a normal machine learning algorithm. Once the calculations are performed, these are allocated in the output layer, resulting in the prediction. The final output is compared to the true label using a loss function. The loss function measures how far the model’s prediction is from the actual value. Backpropagation calculates the gradient of the loss with respect to each weight in the network using the chain rule from calculus. This tells us how much each weight contributed to the error. The model then adjusts the weights and voilá! You have a new prediction. This cycle repeats until the prediction is accurate.

The number of layers, the nodes and the mathematical operations are defined beforehand. Consider the can be billion of nodes, making neural networks massive! The training of the model requires high computational capacity, something that was only achieved in the early 2010s. The advances in capacity are the main reason why artificial intelligence has only become mainstream in the last few years. However, the evolution of the neural networks took decades, the First simple neural network with a single layer dating back to the 1950s!

Some of the most common use cases of Deep Learning and Neural Networks involve:

-            Computer Vision: Image classification (Identifying objects in images (e.g. dogs vs cats); Object Detection (Self-driving cars detecting pedestrians, signs, etc); Facial Recognition (Used in security, phones, and social media); Medical Imaging  (Detecting diseases from X-rays, MRIs (e.g. cancer screening)).

-            Natural Language Processing (NLP): Language Translation (Google Translate); Text Generation (ChatGPT, email autocomplete); Speech Recognition (Converting spoken language to text (Siri, Alexa)).

-            Audio & Speech Processing: Voice Assistants (Real-time speech-to-text and intent understanding); Audio Generation – Deepfake voices, music generation (e.g. Jukebox); Speaker Identification – Verifying who is speaking (biometric security).

-            Finance: Fraud Detection (Identifying unusual patterns in transactions); Algorithmic Trading (Predicting market movements using time series data); Credit Scoring (Evaluating risk based on customer data).

-            Healthcare: Drug Discovery (Predicting how molecules interact); Predictive Diagnostics (Anticipating patient deterioration or readmission); Personalized Treatment Plans (Based on patient history and genomics).

As I wrote some time back, the future was predicted by writers. Someone first imagined it was possible to replicate the human brain and then took action. Neural networks represent a true black box, where is it not certain how the machine readjusts itself to provide more accurate predictions. What have we humans yet to discover? What will unfold once complex mathematical and statistical equations, maybe thousand-year-old theorems are finally solved? As it looks, this is just the tip of the iceberg. Mimicking the human brain is the first step to the creation of real intelligence. The question is, will the machines turn against us? Stay tuned for more!

 

What Is Deep Learning? | IBM

Purpose of different layers in a Deep Learning Model

Introduction to Deep Learning - GeeksforGeeks