[Diesen Artikel auf Deutsch lesen]
Artificial intelligence (AI) is a hotly debated topic in science, business and with the media. Never before have we had such spectacular headlines promising tech innovations such as self-driving cars, autonomous chatbots, machine translation, and self-learning algorithms for speech recognition and some systems that can even detect cancer early.
But why is AI just now exploding onto the scene, and how does machine learning actually work?
We brought in Olaf, an expert on ML and he told us all about it how computers learn human language. At our workshop, he explained the basics of machine learning and answered our questions.
What did we learn? Quite a lot. Here are the highlights:
Automatic speech processing
By now, we’re all well acquainted with the question/answer dialogue games we play with Siri, Alexa, or Google Assistant, but in addition to these little robot helpers, automatic language processing can be found in numerous other applications.
These can include:
- The classification of texts such as those required in email spam filters.
- Clustering of texts for things such as customer questions or applications
- Automatic translation
So, how does automatic speech processing work?
The classic approach to teaching language to computers is to parse texts using grammar. The problem with this is that natural language is a tough nut to crack, and language is jam-packed with nuance and ambiguity.
- Fruit flies like bananas
- How time flies
In cases such as these, the word “flies” can be either a noun or a verb, and it’s up to the machine to figure out the difference based on context and usage.
The management of such ambiguities in language requires a sort of knowledge that is typically reserved for humans; the ability to decipher the context and meaning of words. Despite the difficulty this poses to artificial intelligence, machine learning is more promising than ever before. The reason? Artificial neural networks modeled vaguely on human synapses.
Artificial neural networks and language learning
Unlike the human brain, neural networks work with numbers. “Word embedding” allows for text to be input in the form of numbers – which computers understand. This allows for the construction of word vectors, typically in 50-300 dimensions. These vectors are not created manually, but automatically by reading large amounts of text, such as those on Wikipedia.
- Q&A answer systems and chatbots
- email sorting, spam filtering
- machine translation – term management and storage
Why are these advances in machine learning happening right now?
Feeding or training neural networks with a larger amount of of higher quality data is crucial for increasing the capabilities of machine learning.
Currently, there are two primary reasons why these leaps and bounds in machine learning are occurring now.
- Big data: Huge amounts of data are being input thanks to digitization. This data allows for extensive training and the expansion of ML capacity.
- Advances in hardware with sufficient computing power enables computers to manage this huge amount of information.
Another big step forward has been the new algorithms that have led to a higher level of learning success. Deep learning in particular has become the buzzword in AI.
So, what’s so special about deep learning?
The term “deep learning” probably stems from “deep networks” that have more than three layers – sometimes considerably more. The complexity of deep learning models enables a high degree of abstraction. Input data can be abstracted so that solutions can be found for cases that were not included in the original training data.
One great thing about deep learning models is that the neural networks can generalize and be abstract.
But there’s a problem. A complex neural network is like a “black box.” This means surprises.
The results do not always correspond with what one expects or hopes for.
For example, among thousands of cat images, one may be classified as a dog. Troubleshooting is extremely complicated because the source of the error is difficult to locate. This learning is more of an approximation process.
Another decisive point is the quality of the training data. This is getting better and better as more and more data becomes available. In the meantime, this has even become a business field of its own: crowdsourcing of training data for feeding deep learning machines.
What will the future bring?
Forecasting the future is difficult. When it comes to new technology, we often overestimage the short-term effects and underestimate the long-term. Predictions about the next “killer application” are often wrong.
One such instance of this could be the hype surrounding virtual reality (VR). The predicted successes have not yet truly been in line with the reality. On the other hand, before SMS or the Internet, nobody really saw these taking off – so you never know.
Among the most promising real-world applications for AI has been those in the medical fields. Here, intelligent image recognition methods are successfully being used for the early detection of certain cancers.
And that’s not all. Advances in the field can also be felt in everyday life with systems such as voice and image search, which are revolutionizing Internet usage.
The extent to which this trend will continue is unknowable, but if progress has been any indicator, we may be in for some big surprises.
Will AI change the world?
The ultimate question is extremely exciting: which areas will be changed by adaptive machines?
What do you think? When you think about artificial intelligence, which applications pop into mind first? What are your experiences with them? Have you done things such as book flights using voice search? Does your company use automated chatbots? What about machine translation tools like Google Translate?
Let us know in the comments and let’s talk about it!
[This article has been adapted from its original German edition by Adrian Anton]