The Truth About Deep Learning: What You Need to Know in Plain English
We’ve been on a journey through the different regions of Machine Learning. We’ve explored Supervised Learning, Unsupervised Learning, and Reinforcement Learning. If you haven’t had a chance to read those posts, that’s fine; you don’t need to read them in any particular order to make sense of the topic. We’ll conclude the trek with Deep Learning, which mimics the structure of the human brain. Deep Learning uses artificial neural networks, which follow the architecture and function of our own brains. With the rise of ChatGPT, image generation, and self-driving cars, Deep Learning is getting a lot of attention, and it deserves ours.
What is Deep Learning?
Deep Learning is a subset of Machine Learning that uses neural networks with many layers of artificial neurons. That’s a term that we need to understand for this to make sense. Artificial neurons are simple mathematical functions that roughly mimic how biological neurons process information. This imitation is very basic, though, while our biological information processing is extremely complex. But those layers — it’s the multitude of the layers that makes it “deep.”
At the front of the stack is the input layer, which takes in raw data. The learning model doesn’t do anything with it here; it just receives it and sends it to the next layer. There could be any number of “hidden” layers, where the actual processing occurs. They’re called ‘hidden’ because their internal operations aren’t visible — we only see what goes in and what comes out. These layers extract features and patterns from the input data. After all the layers have had a turn with the data, the results are passed to the output layer.
I’m giving you a very abbreviated description of Deep Learning. It’s much more complicated in execution. To give a simple analogy to help with the textual explanation, think about how children learn from watching their parents do things. We don’t explain everything that we’re doing, but they pick up a lot (sometimes more than we’d like). Sometimes we do tell them, “Watch me, and do it just like I do it.”
Another analogous example is the way Thomas Edison approached his work on creating an incandescent bulb. He didn’t know what the right answer was; he discovered it after testing many possibilities. Edison didn’t program the outcome. He discovered it through trial, error, and persistence — much like how deep learning models find their way to solutions through repeated feedback.
How Deep Learning Differs from Other Types of Machine Learning
Supervised Learning vs. Deep Learning
- SupervisedML typically uses simpler algorithms (e.g., decision trees, linear regression) that rely on human-engineered features
- Deep Learning automates feature extraction (e.g., it figures out edges, shapes, and patterns itself in an image)
- Deep Learning is more data-hungry and resource-intensive, but also more powerful
Unsupervised Learning vs. Deep Learning
- Unsupervised ML (e.g., clustering, dimensionality reduction) finds hidden patterns in unlabeled data
- Deep Learning can be unsupervised (e.g., autoencoders, generative models), but is more often used in supervised settings
- DL can uncover complex patterns that traditional unsupervised methods might miss
Reinforcement Learning vs. Deep Learning
- Reinforcement Learning is goal-directed learning through trial and error
- Deep Learning often provides the architecture for RL agents (Deep Q-Networks, policy gradients)
- Deep RL = reinforcement learning powered by neural networks
Key Components of Deep Learning
- Neurons and Layers: Input, hidden, and output layers
- Weights and Biases: How the model “learns”
- Activation Functions: Introducing non-linearity (ReLU, Sigmoid, etc.)
- Backpropagation and Gradient Descent: How errors are corrected and the model improves
Real-World Applications
We see Deep Learning models applied to voice assistants like Siri and Alexa. These models are used in autonomous vehicles, as well as facial recognition. Deep Learning is the primary driver behind image generation tools like Midjourney and Stable Diffusion. Healthcare is another application, in diagnosis and drug discovery, and industries can use it in predictive maintenance.
But wait, didn’t we also read about autonomous vehicles and voice assistants using other models of Machine Learning? Why, yes – yes, we did. Autonomous vehicles – and nearly everything else that uses Machine Learning – often uses several different types of ML. Each type brings a different strength to the operation, and no single type can do it all.
Challenges and Ethical Concerns
Anytime we introduce new technology, we face new questions, and some of those questions deal with ethics. We talked about that in my series on AI and Ethics, but we need to bring it up again. With many of the new products, we’ll have the same questions, and sometimes – but not always – the answers will be the same. We still need to address them.
Data hunger and environmental cost
These models need data to work with, and that data has to come from somewhere. More data to work with will almost always mean better outcomes. After all, the point of developing AI and ML technologies is to achieve better outcomes. Sourcing the data and processing it all requires a lot of processing capacity. Processing capacity requires energy, and it generates a lot of heat, which requires more energy to reduce or to somehow dissipate. We need to recognize and understand – and articulate – our tolerance for the environmental consequences.
The “black box” nature of deep learning
We can see the inputs and the outputs, but we can’t easily understand how the model made its decision in between. This may seem inconsequential, but it really does matter. For example, why did the model predict cancer in the scan? Can a doctor trust the presentation? Was that hiring decision fair, or was there an influence of race, gender, or zip code? Can we clearly understand why an algorithm flagged someone as “high risk”? Without transparency, it can be hard to explain outcomes, detect and correct bias, and build trust in the system.
Bias in training data
We’ve discussed bias in this article, but it’s worth bringing up again. It’s no less relevant in Deep Learning than it is anywhere else. We will receive the results of what we input, and if we provide examples that indicate that tables can fly, the model will provide us with the information that tables can fly. That’s not exactly bias in the traditional sense, but it illustrates how the model will reflect whatever ‘reality’ we feed into it.
What’s Next for Deep Learning
Assuming we can overcome the challenges and adequately address the concerns, where might we go from where we are now? For one thing, one might argue for research into more efficient and explainable models. As an experimental model, it’s fine to say, inputs go here, stuff happens, outputs come out here. But when we are making life-and-death decisions on those outputs, we need to know what contributed to them.
We might also see Deep Learning running on smaller devices, given the advances in compute power independent of larger CPUs. That, in itself, could facilitate faster integration with robotics, healthcare, and creative tools. Given the history of Machine Learning in general, we should expect a convergence with other fields like neuroscience and quantum computing.
Your Turn
Deep Learning is powerful, but it’s complex. With so many moving parts, usable outputs depend on reliable performance at every step. Machine Learning in general is still in its infancy.
I have mixed feelings about Deep Learning, partly because of the lack of transparency in that “black box.” What do you think? When you look at Deep Learning, what are your concerns, and what are your hopes? Let me know in the comment section, it’s just below the Related Posts bar.
For some more sources on Deep Learning, check out these links:
Difference Between Machine Learning and Deep Learning | GeeksforGeeks
My photography shops are https://www.oakwoodfineartphotography.com/ and https://oakwoodfineart.etsy.com, my merch shops are https://www.zazzle.com/store/south_fried_shop and https://society6.com/southernfriedyanqui.
Check out my New and Featured page – the latest photos and merch I’ve added to my shops! https://oakwoodexperience.com/new-and-featured/
Curious about safeguarding your digital life without getting lost in the technical weeds? Check out ‘Your Data, Your Devices, and You’—a straightforward guide to understanding and protecting your online presence. Perfect for those who love tech but not the jargon. Available now on Amazon:
https://www.amazon.com/Your-Data-Devices-Easy-Follow-ebook/dp/B0D5287NR3
