At Rise of AI 2019 in Berlin, Florian Kuhlmann, CTO and founder of Leverton, argues that today's narrow AI needs to take 'baby steps' toward more general capabilities before it can handle tasks that require reasoning and context. He describes Leverton's growth from two people to 90 across four offices, extracting information from legal and corporate documents with deep learning models trained daily in production. Using a hypothetical example of identifying twin brothers from a photo, and a real Leverton case of a two-column document with broken indentation, he shows that current deep learning excels only at single, narrow tasks and cannot combine visual, textual and semantic information to resolve ambiguity without heavy manual heuristics. He points to multitask learning, multimodal learning and differentiable programming as research directions that might narrow the gap toward artificial general intelligence, while stressing that businesses cannot wait for that goal. In the Q&A he discusses how Leverton stores extracted knowledge and says full natural-language understanding by machines remains far off.
Key takeaways
- Kuhlmann says Leverton grew from two people to 90 across four offices extracting information from legal and corporate documents.
- AI today excels only at narrow, single tasks and cannot combine reasoning across them, shown with a twin-identification example.
- A real Leverton document example shows layout detection fails on a two-column format without feedback from a later semantic step.
- He points to multitask learning, multimodal learning and Yann LeCun's differentiable programming as possible next steps toward general AI.
- In Q&A he says Leverton stores extracted knowledge mainly in normal or graph databases, and doubts full natural-language understanding is near.
Detailed notes
Leverton's narrow-AI production experience
Kuhlmann says Leverton grew over seven years from two people to 90, across four offices with headquarters in Berlin, building a software-as-a-service application that extracts information from legal and corporate documents into structured databases such as ERP systems, semi-automatically with a review process for complex cases. The company has used deep learning for about six years, for OCR, document classification and information extraction, training thousands of models. He describes today's AI as excellent at narrow, single tasks such as classifying a document, finding a person in a picture or recognizing traffic signs, sometimes reaching superhuman performance, but unable to do anything beyond what it was specifically trained for.
The twins example: why narrow AI can't reason
To illustrate the gap toward artificial general intelligence, Kuhlmann uses the example of identifying two twin brothers in a photo. A face-recognition model alone cannot tell them apart; combining it with OCR to read a name on a t-shirt helps identify one twin, but concluding the other twin's identity requires reasoning that they are brothers, which current systems cannot do without extensive hand-built heuristics chaining several models together.
A real document that breaks layout detection
In an actual Leverton case, a lawyer-drafted document has broken indentation that, on inspection, turns out to be a two-column layout rather than a formatting error. Detecting this from visual features alone is not possible; it requires feeding back information from a later, semantic processing step to an earlier layout-detection step, which Kuhlmann says again demands complex heuristics for what is still just one edge case among many.
Possible next steps toward general AI
Kuhlmann outlines research directions Leverton is watching: multitask learning, where a network learns several tasks at once and each benefits from the others, though it remains narrow; multimodal learning, combining inputs such as image, text and audio, which needs well-aligned training data that is hard to produce; and differentiable programming, an approach he associates with Yann LeCun, in which the heuristics and code wrapped around neural networks become themselves learnable and differentiable. He notes tools like TensorFlow and PyTorch come close but do not yet fully support this. His closing thought is that turning hand-built heuristics into trainable neural-network components tends to improve performance, and doing this systematically might move AI closer to general intelligence.
Q&A highlights
Asked how extracted knowledge is stored, Kuhlmann says Leverton mostly uses normal databases, with some experience using ontology-based graph databases, but says reaching artificial general intelligence would likely require storing knowledge inside neural network embeddings rather than structured databases. Asked what is fully automatable today in document extraction, he says accuracy depends on how similar documents are to each other: near-total automation is possible for very similar documents, but complex, cross-referencing document series such as a lease with amendments and renewals still need careful review. Asked whether machines will achieve real natural-language understanding in the next few years, he says no, arguing that basic information extraction is possible but full understanding, including grammar and sentence dependencies, remains far away.
Quotes
AI of today is very good at very narrow tasks, very singular tasks dedicated to one specific workflow.
Florian Kuhlmann
We reach sometimes superhuman performance.
Florian Kuhlmann
Questions answered in this talk
Why can't today's AI reason about context, like identifying twins in a photo?
According to Florian Kuhlmann at Rise of AI 2019, narrow AI models each solve one isolated task, such as face recognition or OCR, and chaining them cannot supply the reasoning (for example, knowing two people are brothers) needed to resolve ambiguity without extensive manual heuristics.
What database does Leverton use to store information extracted from documents?
According to Florian Kuhlmann at Rise of AI 2019, Leverton mainly uses normal databases, with some experience using ontology-based graph databases, though he says a shift toward artificial general intelligence would likely require storing knowledge in neural network embeddings instead.
What are possible next steps toward artificial general intelligence?
According to Florian Kuhlmann at Rise of AI 2019, promising directions include multitask learning, multimodal learning combining image, text and audio, and differentiable programming, an approach associated with Yann LeCun, that makes the heuristics around neural networks themselves trainable.