Exploring Enterprise ML
Since the early applications in the 1960’s, Machine Intelligence has had many names attached to it, including data science, adaptive models, and symbolic reasoning. But whatever the nomenclature du jour over the decades, this much is true nowadays: ML is one of the most effective tool sets that a company can use to build and maintain a competitive moat.
We’ve never been more excited for the state of this technology, which has come to be for several reasons:
Academic Research Wins
- Multi-agent Reinforcement Learning: Researchers from DeepMind, UMontreal, and Cambridge (amongst others!) are building successful applications using multi-agent reinforcement learning frameworks, which should provide opportunity to learn from more complex, real world or simulated (increasingly real world-like) environments
- Transfer Learning: We are beginning to see positive results in transfer learning in narrow cases which we believe is an incremental step toward pervasive enterprise intelligence, applicable in even highly regulated industries such as Healthcare and Insurance
- Multi-modal, Sparse Data Utility: Groups are reducing time-to-value and automating pre-processing of multi-modal training data including numerics, image and text into related or single models
Data And Infrastructure For Now
- More Than Moore’s Law: Moore’s Law, which states roughly that the number of transistors that can go on an integrated circuit doubles every couple of years, is more than half a century old. Amid concerns it may be nearing the end of its natural life. Many technical teams are pushing limits to improve hardware architecture from GPU’s to TPU’s and other ASIC chips built for AI-specific use cases. Relatedly, running federated models directly on chip / device is helpful and particularly important for low latency decision making and data security
- Data Generation Democracy: Training data, which is getting created at an exponential rate, can now be used in large scale simulations using tools such as Improbable, as well as generated using a gambit of open source or fit-for-purpose providers, such as Parallel Domain
- People Playing Nicely: Software tools such as Dataiku, Periscope, Alation and others are enabling better and faster software engineering and data science workflows (pre-processing data, evaluating and deploying models, monitoring, governance, etc.)
- The Data Company: C-level execs see success directly linked to data-driven experimentation. Every company isn’t just a technology company anymore — now they must be ‘data’ companies where optimized data flow and dataset normalization become major inputs to operational improvement
- HR Demand: The two most widely posted jobs on Linkedin are ML Engineer and Data Scientist, respectively. More importantly, both roles aren’t any longer siloed to specific departments, like marketing or pricing. Rather, they are now considered key to the success of nearly all departments and software deployments
- Automation Culture: Investments are being made in automation and RPA across the enterprise stack, and employed frequently by even non-technical users
Costanoa Ventures has historically had particular affinity toward computer vision applications. In the past five years, the technical capability has progressed significantly — particularly in image recognition, which now performs at super human-level in certain domains. If you feed enough labeled and formatted image data through well-designed deep net systems, these systems exhibit powerful image recognition capabilities. Perception, however, is far from being solved. Issues remain including the binding problem, applicability of transfer learning on new datasets, the understanding of contextual information, and iteratively reducing false positives, just to name a few. However, it is important to note that these advances are already creating value today in everything from inspection tasks (PreNav, Tractable), to retail automation (FocalSystems*) to continuous aquaculture monitoring (Aquabyte*). We’re particularly interested in connecting with talented technologists working on achieving contextual understanding from video, which should enable better search experiences for consumers and users in the enterprise.
Some argue that the ability to predict is the essence of intelligence. In the context of the enterprise, prediction is powerful in that it fundamentally enables better resource allocation — from people to pricing to product iterations. Rapid institutional learning drives more competitive enterprises, which is near and dear to every startup founder and investor’s heart. We’d love to help build companies in other subfields of applied machine intelligence as well as be participants in the infrastructure shifts that serve it.
Please join us for a tech talk to discuss Uber’s ML-as-a-service platform on May 17.
As perpetual students of technology, we at Costanoa leverage and learn from a deep network of technologists in academia as well as commercial organizations large and small, who are pioneering on both the infrastructure and application level. We’re excited to announce that we’ve invited Jeremy Hermann to our offices on May 17th to speak about Michelangelo: Uber’s Machine Learning Platform. Jeremy will discuss Uber’s ML-as-a-service platform and how they designed it to cover the end-to-end ML workflow: manage data, train, evaluate, and deploy models, make predictions, monitor predictions and support traditional ML models, time series forecasting, and deep learning. He will also talk about the challenges and solutions that drove their approach as well as what they continue to learn. Register: http://bit.ly/2qoQCdg.
*Note: Costanoa is an investor in Focal Systems, Alation and Aquabyte.