Machine learning (ML) has been overhyped and has generally under-delivered. In fact, I would argue that the term itself is a misnomer. A more accurate name might be machine classification or machine filtering because what the technology actually does is label buckets of data based upon a training corpus. The problem is that the so-called ‘learning’ taking place is often limited and short-lived because as the complexity of the data increases, the probability for misclassification increases. This can quickly lead to machine ignorance or machine amnesia.
Achieving eighty percent accuracy with ML is typically fairly easy. It is closing the gap on the remaining twenty percent that is the real challenge mainly because, in most cases, the data required to optimize accuracy and efficiency isn’t available. Without the right training data, ML inevitably hits a point of diminishing returns. And while many try to resolve this issue by using a larger training corpus, this often results in the homogenization and polluting of the data which actually degrades efficiency. The following graphs shows how this curve plays out visually.
Topos Labs’ cognitive text mining platform Gracie represents the next generation of machine learning based on human-augmented intelligence. Gracie seamlessly blends the nuanced knowledge, expertise, and context of humans with the speed, scale and efficiency of artificial intelligence to drastically improve the accuracy and real-world utility of machine learning.
Via a simple, intuitive interface, users fine-tune Gracie’s understanding of the world, helping her to focus on prioritized data, achieve incomparable degrees of granularity, and bridge the stubborn gap between machine learning and deep learning. This offers a dramatic improvement over traditional ML, where feature vectors are all treated equally, requiring pre-training corpus data to be manipulated to boost accuracy. Human-augmented learning not only saves time and money, it improves results and eliminates much of the guesswork from the process.
When Gracie hits a point of confusion, her users provide clarity and guidance. For example, when Gracie interprets two different conversations errantly as meaning the same thing, a user can step in to differentiate them for her. With the benefit of this human guidance, Gracie re-analyzes the data and this process continues in a virtuous loop until the desired level of accuracy is achieved. This is all done without programming and the adjustments take effect in near real-time, unlike traditional AI approaches in which simple feature vector modifications could take weeks or months.
Deep learning (DL) is touted as the next evolution of ML. And while DL has achieved some very interesting results within closed environments, classification discovery is less impressive, lacks a focused approach, and relies heavily on random luck. This is why deep learning requires so much computing power. The vast majority of cycles are wasted on analytics that have no relevance, and in many cases (such as ML), the raw data lacks the granularity needed to accurately classify results. Adversarial machine learning is a DL technique that holds considerable promise and will be utilized and optimized by Gracie in the near future. The difference is that human-machine teaming with Gracie can prioritize scenarios minimizing dead ends and reducing the need for massive compute power and big data sets.
Gracie was engineered to directly address the limitations, flaws and expense of conventional ML environments. Gracie’s people-centric, human-machine teaming environment empowers domain experts, and reduces the need for technology specialists. Gracie is AI for the rest of us.
About Dr. Couturier
Dr. Couturier is a pioneer and one of the world’s foremost experts in machine learning for security analytics. He created the first applications that analyzed encrypted traffic. Over the last 15 years he has architected dozens of security solutions to detect breaches, infections, fraud, and phishing attacks.