I recently had the opportunity to discuss current IBM artificial intelligence developments with Dr. Lisa Amini, an IBM Distinguished Engineer and the Director of IBM Research Cambridge, home to the MIT-IBM Watson AI Lab. Dr. Amini was previously Director of Knowledge & Reasoning Research in the Cognitive Computing group at IBM’s TJ Watson Research Center in New York. Dr. Amini earned her Ph.D. degree in Computer Science from Columbia University. Dr. Amini and her team are part of IBM Research tasked with creating the next generation of Automated AI and data science.
I was interested in automation’s impact on the lifecycles of artificial intelligence and machine learning and centered our discussion around next-generation capabilities for AutoAI.
AutoAI automates the highly complex process of finding and optimizing the best ML model, features, and model hyperparameters for your data. AutoAI does what otherwise would need a team of specialized data scientists and other professional resources, and it does it much faster.
AI model building can be challenging
“How Much Automation Does a Data Scientist Want?”
Building AI and machine learning models is a multifaceted process that involves gathering requirements and formulating the problem. Before model training begins, data must be acquired, assessed, and preprocessed to identify and correct data quality issues.
Because the process is so complex, data scientists and ML engineers typically create ML pipelines to link those steps together for reuse each time data and models are refined. Pipelines handle data cleansing and manipulation operations for model training, testing and deployment, and inference. Constructing and tuning a pipeline is not only complex but also labor-intensive. It requires a team of trained resources who understand data science, plus subject-matter experts knowledgeable about the model’s purpose and outputs.
It is a lengthy process because there are many design choices to be made, plus a myriad of tuning adjustments for various data processing and modeling stages.
The pipeline’s high degree of complexity makes it a prime candidate for automation.
IBM AutoAI automates model building across the entire AI lifecycle
According to Dr. Amini, AutoAI does in minutes what would typically take hours to days for a whole team of data scientists. Automated functions include data preparation, model development, feature engineering, and hyperparameter optimization.
End-to-end automation of an entire model building process can result in significant resource savings. Here is a partial list of AutoAI features:
AutoAI provides a significant productivity boost. Even a person with basic data science skills can automatically select, train, and tune a high-performing ML model with customized data in just a few mouse clicks.
However, expert data scientists can rapidly iterate on potential models and pipelines, and experiment with the latest models, feature engineering techniques, and fairness algorithms. This can all be done without having to code pipelines from scratch.
Future AI automation projects
IBM Research is working on several next-generation AI automation projects, such as next-generation algorithms to handle new data types, bring new automated quality and fairness, and dramatically boost scale and performance.
Dr. Amini provided a deep dive into two especially interesting next-generation capabilities for scaling enterprise AI: AutoAI for Decisions and Semantic Data Science.
AutoAI for improved decision making
Time series forecasting is one of the most popular but one of the most difficult predictive analytics. It uses historical data to predict the timing of future results. Time series forecasting is commonly used for financial planning, inventory, and capacity planning. The time dimensions within a dataset make analysis difficult and require more advanced data handling.
IBM’s AutoAI product already supports Time Series forecasting. It automates the following steps of building predictive models:
Dr. Amini explained that after a time series forecast is created in many settings, the next step is to leverage that forecast for improved decision-making.
For example, a data scientist might build a time series forecasting model for product demand, but the model can also be used as input for inventory restocking decisions with the goal to maximize profit by reducing costly over-stocking of too much inventory or avoiding lost sales due to stock outages.
Simple heuristics are sometimes used for inventory restocking decisions, such as determining when inventory should be restocked and by how much. In other cases, a more systematic approach, called decision optimization, is leveraged to build a prescriptive model to complement the predictive time series forecasting model.
Prescriptive analytics (as opposed to predictive analytics) use sophisticated mathematical modeling techniques and data structures for decision optimization and leverage expertise in short supply. However, products for automated decision optimization pipeline generation created directly from data, like AutoAI for predictive models, do not exist today.
Dr. Amini explained that the best results are obtained by using both machine learning and decision optimization. To support that capability, IBM researchers are working on multi-model pipelines that could accommodate the needs of predictive and prescriptive models. Multi-models will allow business analysts and data scientists to use a common model to discuss aspects of the problem from each other’s perspectives. Such a product would also promote and improve collaboration between diverse but equally essential resources.
Automation for Deep Reinforcement Learning
The new capability to automate pipeline generation for decision models is now available through the Early Access program from IBM Research. It leverages deep reinforcement learning to learn an end-to-end model from data to decision policy. The technology, called AutoDO (Automated Decision Optimization), leverages reinforcement learning (RL) models and gives data scientists the capability to train machine learning models to perform sequential decision-making under uncertainty. Automation for reinforcement learning (RL) is critical because RL algorithms are highly sensitive to internal hyperparameters. Therefore, they require significant expertise and manual effort to tune them to specific problems and data sets.
Dr. Amini explained that the technology automatically selects the best reinforcement learning model to use according to the data and the problem. Using advanced search strategies, it also selects the best configuration of hyperparameters for the model.
The system can automatically search historical data sets or any gym-compatible environment to automatically generate, tune, and rank the best RL pipeline. The system supports various flavors of reinforcement learning, including online and offline learning and model-free and model-based algorithms.
Scaling AI with automation
Automation for reinforcement learning tackles two pressing problems for scaling AI in the enterprise.
First, it provides automation for sequential decision-making problems where uncertainty may weaken heuristic and even formal optimization models that don’t utilize historical data.
Secondly, it brings an automated, systematic approach to the challenging reinforcement learning model building domain.
Semantic Data Science
State-of-the-art automated ML products like AutoAI can efficiently analyze historical data to create and rank custom machine learning pipelines. It includes automated feature engineering, which expands and augments the feature space of data to optimize model performance. Automated methods currently rely on statistical techniques to explore the feature space.
However, if a data scientist understands the semantics of the data, it is possible to leverage domain knowledge to expand the feature space to increase model accuracy. This expansion can be done using complementary data from internal or external data sources. Feature space is the group of features used to characterize data. For example, if the data is about cars, the feature space could be (Ford, Tesla, BMW).
Complementary feature transformations may be found in existing python scripts or relationships described in the literature. Despite this, knowing which features and transformations are relevant, a user must have sufficient technical skills to decipher and translate from code and documents.
New semantic power for data scientists
Dr. Amini described another powerful new capability created by IBM Research called Semantic Data Science that automatically detects semantic concepts for a given dataset. Semantic concepts characterize concepts to help understand the words and sentences to provide a way for meanings to be represented. Once AutoAI has detected the proper semantic concepts, the program uses those concepts in a broad search for relevant features and feature engineering operations that may be present in existing code, data, and literature.
AutoAI can use these new, semantically-rich features to improve the accuracy of generated models and provide human-readable explanations with these generated features.
Even without having domain expertise to assess these semantic concepts or new features, a data scientist can still run AutoAI experiments. However, data scientists who want to understand and interact with the discovered semantic concepts can use the Semantic Feature Discovery visual explorer to explore discovered relationships.
Users can go directly from the visual explorer into the python code or document where the new feature originated simply by clicking the Sources hyperlink, as shown in the graphics below.
The Semantic Data Science capability is also available as an IBM Research Early Access offering. Some of the capabilities are even available for experimentation on IBM’s API Hub.
Dr. Amini concluded our conversation and summed up the vast research effort IBM is pouring into AutoAI with one single yet efficient sentence:
“We want AutoAI and Semantic Data Science to do what an expert data scientist would want to do but may not always have the time or domain knowledge to do by themselves.”
Wrap-up key points
For more information and comments about quantum computing and artificial intelligence, follow Paul Smith-Goodson on Twitter @Moor_Quantum
Moor Insights & Strategy, like all research and tech industry analyst firms, provides or has provided paid services to technology companies. These services include research, analysis, advising, consulting, benchmarking, acquisition matchmaking, or speaking sponsorships. The company has had or currently has paid business relationships with 8×8, A10 Networks, Advanced Micro Devices, Amazon, Ambient Scientific, Anuta Networks, Applied Micro, Apstra, Arm, Aruba Networks (now HPE), AT&T, AWS, A-10 Strategies, Bitfusion, Blaize, Box, Broadcom, Calix, Cisco Systems, Clear Software, Cloudera, Clumio, Cognitive Systems, CompuCom, CyberArk, Dell, Dell EMC, Dell Technologies, Diablo Technologies, Dialogue Group, Digital Optics, Dreamium Labs, Echelon, Ericsson, Extreme Networks, Flex, Foxconn, Frame (now VMware), Fujitsu, Gen Z Consortium, Glue Networks, GlobalFoundries, Revolve (now Google), Google Cloud, Graphcore, Groq, Hiregenics, HP Inc., Hewlett Packard Enterprise, Honeywell, Huawei Technologies, IBM, IonVR, Inseego, Infosys, Infiot, Intel, Interdigital, Jabil Circuit, Konica Minolta, Lattice Semiconductor, Lenovo, Linux Foundation, Luminar, MapBox, Marvell Technology, Mavenir, Marseille Inc, Mayfair Equity, Meraki (Cisco), Mesophere, Microsoft, Mojo Networks, National Instruments, NetApp, Nightwatch, NOKIA (Alcatel-Lucent), Nortek, Novumind, NVIDIA, Nutanix, Nuvia (now Qualcomm), ON Semiconductor, ONUG, OpenStack Foundation, Oracle, Panasas, Peraso, Pexip, Pixelworks, Plume Design, Poly (formerly Plantronics), Portworx, Pure Storage, Qualcomm, Rackspace, Rambus, Rayvolt E-Bikes, Red Hat, Residio, Samsung Electronics, SAP, SAS, Scale Computing, Schneider Electric, Silver Peak (now Aruba-HPE), SONY Optical Storage, Springpath (now Cisco), Spirent, Splunk, Sprint (now T-Mobile), Stratus Technologies, Symantec, Synaptics, Syniverse, Synopsys, Tanium, TE Connectivity, TensTorrent, Tobii Technology, T-Mobile, Twitter, Unity Technologies, UiPath, Verizon Communications, Vidyo, VMware, Wave Computing, Wellsmith, Xilinx, Zayo, Zebra, Zededa, Zoho, and Zscaler. Moor Insights & Strategy founder, CEO, and Chief Analyst Patrick Moorhead is a personal investor in technology companies dMY Technology Group Inc. VI and Dreamium Labs.