It’s a lazy Saturday afternoon and you’re scrolling through social media, procrastinating on that article you’ve due tomorrow when you inevitably come across that dime-a-dozen reel about AI and how it’s going to take your job. From these trivial interactions on social media to even national news, AI has been a hot topic recently. We’re all quick to talk about machine learning, deep learning and its other subfields, painting them as boons or banes to our liking.
But let’s pause for a moment and take a look at a neglected piece of technology that powers almost every single piece of AI tech we use today, Neural Networks. In this article, we’ll be discussing what they are, why they’re useful and where they’ve been applied.
Neural Networks: What are they?
We can think of Artificial Neural Networks(ANNs) as a recreation of the way the human mind works. Let’s say you pick up an apple. There are a load of sub-tasks to ‘picking up an apple’, for example, you recognize the apple as an apple, move your arm to pick it up, etc. Let’s take a look at recognizing an apple. You may consider a variety of factors like colour, shape, smell and size before thinking about it and making a decision.
In ANNs, the millions of neural interactions in our brain are replicated by using layers. Each of these layers is made up of parts called nodes. In essence, a node is a function, which takes in an input from the previous layer or the user and gives an output between 0 and 1. These nodes are distributed across 3 main layers. The input layer receives the input from the user, the hidden layers where most of the processing occurs, and the output layer where you can obtain your desired output.
Of course, certain things have more value in the decision. You’re more likely to care about the appearance than taste when guessing if a fruit’s an apple. That’s exactly why each output given by a node is multiplied by a factor called “weight’’. This ensures that factors that have a larger impact on the final decision also have a higher value during the entirety of the process which leads to better accuracy when making predictions.
There’s one final important term to be discussed here which is ‘bias’. Let’s head back to the analogy and make a small change. This time, you’re offered two distinct apples and can pick only one of them. How would you do so? Well, whichever one’s more vibrant in colour, larger in size and overall appears more desirable would be the obvious choice. So, there’s a certain threshold of appearance below which your choice would be a resounding ‘no’. That’s exactly what the bias serves as. We add it after multiplying the output of the node with its corresponding weights. The ‘bias’ here is usually a negative value, signifying the fact that it’s a barrier, which allows only values greater than it to get taken into account.
Neural Networks: The Family, The Genus, The Species
Now, the analogy provided for explaining neural networks may be very simple but don’t let that fool you. The term ‘neural network’ serves as an umbrella for numerous algorithms which follow a similar basic format but differ in terms of intended purposes and use cases. Here we’ll discuss a few of the more popular types. Those being:
- Feedforward Neural Networks (FNNs)
- Convolutional Neural Networks (CNNs)
- Recurrent Neural Networks (RNNs)
FNNs:
You’ve likely all seen the above diagram, as FNNs are the ones commonly portrayed as ‘Neural Networks’ despite being a subtype. In these networks, data flows in one direction, from the input layer to the output layer, hence the name. These networks are super simple to use and understand, but with that simplicity comes challenges as well. They’re not suited for data of unknown dimensions, i.e., you have to know almost everything about the data you’re feeding through it for it to give you desirable results. They’re also horribly inefficient on large scales. So if you’re trying to build a search engine, this is a type to steer clear of!
CNNs:
Let’s say you’ve got to process an image that’s 500x500 pixels. That’s 250,000 individual data points you’ve got to comb through to get a result. Using FNNs for this purpose is neither viable nor even possible since it would take an ungodly amount of time to get you a response. That’s where CNNs come in. They’re made up of a convolutional layer, which applies a filter to each data point and applies the ReLU function giving us a value that lies between 0 and 1 but isn’t an exact whole number (this is what allows for these models to provide “confidence scores”). Then, there are the pooling layers, which reduce the number of parameters being fed into the model by aggregating values from a region of the image together and averaging their values. Finally, the ‘fully-connected layer’ gives you the output.
RNNs:
Usually, neural networks are built to have a clear line between inputs and outputs. However, with RNNs, that line gets blurred ever so slightly. To put it simply, these models use info from both the previous layer’s inputs and outputs to influence the decision the next layer makes. Also, rather than having an individual weight for every node, they utilize a collective weight for each layer. This makes them more efficient than their feedforward counterparts.
Neural Networks: Applications
Let’s take a break from that apple tree and its apples we’re so fond of and take a small trip to the skies. Why, you ask? Well, to discuss one of the most prominent applications of neural networks of course!
Neural networks are a prominent contributor to AAVs, i.e., autonomous aerial vehicles). These are vehicles that have no pilot and absolutely no human interaction. However, without human intervention, it’s unthinkable for a drone to be able to pilot itself right? Well, that’s where neural networks come in.
Most of these AAVs have 2 things in common. 1) They need to collect a lot of information, and 2) They need to avoid being blown to smithereens while doing so. For these purposes, AAVs are usually chock-full of complicated sensors like LiDAR units, GPS systems and Inertial Measurement Units. We’re not going to dive into the specifics of these units here, but rather a consequence of having so many on board.
Think about having to choose between 2 things when you’re given only 3 lines about each of them. It’s quite a simple decision to make, right? Now imagine that you’re given 30 objects and a whole essay for each of them. Of course, your decision would be better informed than the prior case but would take hours if not days more. That’s why neural networks are used in a process called ‘sensor fusion’. They’re trained on vast amounts of data and can comprehend the inputs and provide an output that provides the most fruitful path for the AAV. This data can range from being based off of the averages of the readings from each sensor in High-Level Fusion to the raw data from each sensor in Low-Level Fusion.
Even DJI has caught on to the power that sensor fusion holds, with their Phantom 4 Pro being famous for its state-of-the-art system!
Let’s jump to point number 2 shall we? An AAV isn’t too useful if you can’t have it fly for an hour without having to send it back for repairs. Not only would this be a waste of time, but also be an effective way to burn all your hard-earned cash. Well, here come ANNs to the rescue once more! What we do is connect the AAV’s main camera to a CNN. This CNN processes the image and identifies the main obstacles in its path. It then relays this information to the AAV, which can then be used to avoid said obstacles. Over time, this results in an AAV that can learn from its surroundings and recognise a variety of roadblocks and hiccups, resulting in a smoother journey every time.
For example, Skydio’s Skydio 2 drone, incorporates 9 deep-learning models for 360-degree obstacle avoidance. Of course, they’re not alone. Plenty of other companies like Autel Robotics, DJI, and many others have also jumped onto this bandwagon.
Neural Networks: Everywhere At the End of Time
In conclusion, ‘Neural Network’ isn’t just a buzzword that’s being thrown around. It’s a revolutionary piece of tech with a sizable impact. With its numerous types like RNNs, CNNs and more, it’s impacted a variety of fields including autonomous vehicles, healthcare, finance and more. In a world driven by data, neural networks have emerged as the pioneering force that drives humanity towards a future where innovation knows no bounds.
References
https://www.ibm.com/topics/neural-networks
https://www.ibm.com/topics/convolutional-neural-networks
https://www.ibm.com/topics/recurrent-neural-networks
https://www.youtube.com/watch?v=aircAruvnKk
https://www.uavnavigation.com/company/blog/what-is-an-IMU
https://www.mdpi.com/1424-8220/21/6/2140
