|(Image source: Pixabay)|
The decision to use deep learning in your research and development can be a good one. But only if you carefully consider the outcome. With so much hype surrounding AI, and deep learning in particular, it's hard not to jump on the train and think that AI will solve all of your data analysis problems. Speaking at the 2019 Pacific Design and Manufacturing Show, Jesse Livezey, a postdoctoral researcher in the Neural Systems and Engineering Lab at Lawrence Berkeley National Laboratory, said that this is not necessarily the case. Deep neural networks (DNNs) can offer some powerful computing benefits to researchers, but it is not something everyone necessarily needs.
In Livezey's own research, he and his team were able to apply deep learning to speech recognition for brain-computer interfaces (BCIs). The object is to create a BCI that can accurately interpret speech from brainwave patterns to assist the disabled. Livezey's research has found that when deep learning neural networks make errors in correlating consonant vowel sounds, it corresponds to what physical structure (the lips, tongue, etc.) are associated with making that sound. Since different sub-regions of the brain control structures like the lips, jaw, tongue, and larynx, it gives researchers more insight into where in the brain their BCI should be gathering speech signals from.
That's just one specific example. But Livezey was quick to caution the audience that while deep learning offers many benefits, it also has downsides that don't make it the one-size-fits-all solution that many are tempted to look at it as.
1. ) Do You Really Need Deep Learning?
Let's get the most obvious concern out of the way first. Is deep learning even the best solution for you?
“Have you done something simpler and more interpretable first?” Livezey asked. He said researchers need to figure out what type of result the end user is looking for. Just having really good performance is not necessarily what's needed.
He said the big advantage of DNNs is they offer “high accuracy, high targeting, and high precision.” In some use cases, that might be all that you care about it. But in other areas, such as as scientific or medical research, you might want to understand what the algorithm is doing (more on this in a bit) and you don't need to be as performative as possible. That's where you may want to look at other methods.
Moreover, there is the issue of understanding the right neural network or combination for your task. Are you using a CNN? An RNN? An MNN? Some combination of those? Something else entirely? You'll want to look at the best options for your task at hand, and the work of programming and training a neural network may not necessarily provide the best end result in the most optimal time frame.
2.) Is There Bias?
Algorithms are only as good as the data that's fed into them. Training a neural network on biased data can lead to all sorts of adverse effects. It can lead the AI to solve the wrong problem or draw an incorrect conclusion. At worst, you get AI that practices predatory lending and demonstrates racial bias.
“You want to know if biases exist in your dataset, and are they being used by your machine learning algorithm to make decisions?” Livezey said. “You have to make sure your algorithm is generalizing correctly by pulling pieces out of the dataset that you think are important.”
3. ) Do the Pros Outweigh the Cons?
DNNs excel at performance, however Livezey cautioned performance isn't necessarily the end-all-be-all.
Deep networks are very flexible and you can have many different types of input and output mappings. “Deep networks really do scale to larger datasets better than traditional methods like linear regression,” Livezey said.
The downside to this is it also means there are many more design choices for researchers and engineers to sort through. It can mean a lot more work on the front end with planning, and unless the DNN offers significant benefits, all of that work may not be worth it.
The biggest downside to neural networks, however, is the lack of transparency. Like the human neurons they mimic, deep neural networks are somewhat of a mystery in how they actually perform their functions in many respects. “We don't have great ways of interpreting what these deep networks are doing,” Livezey said. If you're used to using methods such as regression or decision trees, the inner workings are pretty clear and straightforward. Not so much with DNNs. “There's still lot of research into how deep networks are working and making the decisions they do,” he added.
Chris Wiltz is a Senior Editor at Design News covering emerging technologies including AI, VR/AR, blockchain, and robotics.
ESC BOSTON IS BACK!
The nation's largest embedded systems conference is back with a new education program tailored to the needs of today's embedded systems professionals, connecting you to hundreds of software developers, hardware engineers, start-up visionaries, and industry pros across the space. Be inspired through hands-on training and education across five conference tracks. Plus, take part in technical tutorials delivered by top embedded systems professionals. Click here to register today!